Empire Teas CRM and AI Initiative — Project Proposal
Status: issued to client (v0.5, 5 August 2026) — commercially accepted 16 August 2026 (email; USD 15,500 one-time + USD 360/month support). Content below is the issued text and is not edited retroactively. Post-issue notes for the SoW (recorded 18 August 2026 — not edits to the issued text): placeholders remain beyond §17 — §3 (value-case figures, never populated with client numbers), §5 (screenshots), §6 (the review window — five business days, finalised in the SoW) and §A.9 ([N] enhancement hours) — all finalised in the Statement of Work. §A.2.2’s development sub-total does not sum from its rows (~$10 gap); the SoW cost annexe re-derives it before any figure is quoted onward. §10’s “development, staging, and production” is superseded by the two-environment basis actually costed in §A.1/§A.2.1 — the SoW records the confirmation. Post-issue note added 21 August 2026 (later than the notes above; not an edit to the issued text): §A.2’s estate shape is superseded, and §A.1’s region assumption is now closed. Two decisions taken after issue change the shape §A.2.2 prices: both frontends are static hosting on S3 and CloudFront rather than the costed Fargate frontend task (each a React single-page application built with Vite, not the Next.js shown in the §5 diagram — AD #21), and UAT is a temporary environment stood up per acceptance round and destroyed after it, rather than the development environment doubling up as §A.2.1 assumed. Separately, §A.1 left the region “to be confirmed” and warned that a Singapore or European region runs 10% to 25% above the mid-cost basis used — the region chosen is US East, so that uplift does not apply. Taken together with the arithmetic gap already noted above, the re-derived estate sits below the §A.2.2 and §A.8 figures, so no change request arises. The re-derivation is in the companion SoW cost annexe (
sow-cost-annexe.md), which is the figure to quote onward; §A.2.2 and §A.8 stand as issued. Post-issue note added 31 August 2026 (not an edit to the issued text): the §A open-source tools table’s dependency-scanning row is superseded in implementation. §A lists “OWASP Dependency-Check” for dependency vulnerability scanning in the build; at Foundation slice 4 that tool proved unable to scan the pnpm-based frontends, and the gate was implemented on Trivy instead (AD #27, FS-4 spec Rev 2) — also open source and free, same commitment (a build-failing dependency-vulnerability scan across every repo), so no commercial impact and no change request arises. Post-issue note added 7 September 2026 (not an edit to the issued text): §3.2’s access model is widened, inside V1, at Empire Teas’ request. §3.2 states access is controlled by “role plus granular permission levels (read-only, edit, AI-use, and execute)” on the six named roles, and the PRD had deferred any structural org hierarchy. On 7 September Empire Teas asked for a fully configurable role-management system: Administrators author roles from a permission catalog, the six named roles become the built-in starting templates, the four levels survive as the category of each permission, and a reporting hierarchy (each user may name a manager) adds a team-level visibility scope set per permission. Recorded internally as AD #30 / PRD Rev 21. Treated by the delivery team as a V1 requirement — no change request arises; the added scope is carried inside the committed window (R-ARCH-001). The client-facing PRD, question list and role-access document are restated to the configurable model alongside this note; the six roles, the Managing Director approval rule and the human-in-the-loop AI rule in §3.2 stand as issued. Post-issue note added 11 September 2026 (not an edit to the issued text): one pricing detail in §13’s “Confirmed in the 3 August 2026 requirements meeting” list is superseded, and three items are recorded as outside V1. Destination-market pricing — Empire Teas stated on 9 September 2026 that a standard’s price does not vary between customers or destination markets, and the weekly price workbook they shared afterwards carries one price per standard per week; the delivery team records one price per standard (PRD Rev 23), and destination market becomes a customer attribute and a quote field rather than a price driver. The bulk-price-structure, no-volume-tiers, approval and payment-terms bullets stand as issued. Separately, the delivery team has excluded bulk or group email, newsletters and WhatsApp integration from V1 (team decision, 11 September 2026, stated to the client): the issued text never promised any of them — §4 lists two-way sync, templates and tracking, and WhatsApp appears only as where information lives today — so nothing sold is narrowed; it is recorded here so this proposal and the client-facing pack read the same way. No change to scope, price, timeline or providers; no change request arises. Post-issue note added 11 September 2026 (second note of the day; not an edit to the issued text): the 7 September note’s description of the access model is refined. That note said the six named roles become built-in starting templates and that the four permission levels survive as each permission’s category. On 11 September the delivery team simplified the model (team decision, stated to the client; no change request): no role templates — the system ships with one built-in Administrator role and Empire Teas’ Administrator creates every other role from the permission catalog, with Altrium alongside at go-live; one role per user; and AI is not a grantable level — permissions are read / edit / execute, AI the system runs on its own is not a permission, and AI help a person asks for comes with the permission it assists. §3.2’s six named roles, the Managing Director approval rule and the human-in-the-loop AI rule stand as issued. Recorded internally as AD #32 / PRD Rev 24. Post-issue note added 13 September 2026 (not an edit to the issued text): AWS account ownership and infrastructure billing are settled by a separate SOW (ET-002), which supersedes the proposal on these points. §10/§11/§12/§17 and Appendix A treat the infrastructure account as “to be confirmed” and speak of Empire Teas receiving running environments and the models running “inside Empire Teas’ own AWS account”. Empire asked Altrium to own and run the AWS environment for this project (Empire is non-technical; its cloud billing is normally handled by Dialog). This is now governed by the Managed Cloud Infrastructure & Hosting SOW (ET-002,aws-sow.md): Altrium owns and administers the AWS organisation and accounts and bills Empire a fixed USD 650/month (all-inclusive of AWS + AI usage + managed operations; separate from and additional to the USD 360/month application support in this proposal), with the accounts not transferred to Empire on exit — at close Empire receives a data + artifact export, not the running AWS accounts. Per ET-002 §14 the SOW expressly supersedes any inconsistent proposal provision on AWS account creation/ownership, infrastructure operations, infrastructure pricing, and account transfer on exit. The data-governance posture is unchanged in substance (in-account processing, no separate AI vendor), now with Altrium acting as Empire’s processor (ET-002 §10). The USD 360/month application support figure and all other proposal terms stand as issued. Recorded internally as AD-O-10 (Resolved) and PRD Rev 25. Post-issue note added 15 September 2026 (not an edit to the issued text): the “products catalog with bulk pricing” / “products catalog and pricing” named in the scope summary, the data-model deliverable, the §7 phase plan and the Administrator-training list is delivered as the Buying Standards book — one list, not two. Every item Empire Teas sells, bulk or private label, is a client-specific buying standard with its own never-reused running number, so the CRM keeps no separate product catalogue: the standard carries the grade, origin, brand and held certifications a catalog entry would have, and is created, edited and archived by the Tea Room. The same capability the proposal describes, delivered as one list; scope, price and timeline are unaffected (PRD Rev 27, team decision stated to the client — recorded here rather than absorbed, per L-001). Post-issue note added 16 September 2026 (not an edit to the issued text): the bulk-email exclusion recorded in the 11 September note is lifted to a stretch candidate; the WhatsApp exclusion in that note stands. That note recorded that the delivery team had excluded bulk or group email, newsletters and WhatsApp integration from V1. Empire Teas has since confirmed (16 September 2026) that sending in bulk is an important requirement and a feature they would genuinely use, naming two cases: sending the week’s prices to every customer at once, and sending a newsletter. Bulk email is therefore no longer declined — it is recorded as F37, a V1 stretch candidate (PRD Rev 28): low priority at Empire’s own framing, taken into V1 only if the schedule allows and otherwise the first post-V1 item, with the mid-November checkpoint as the decision point. Nothing sold is narrowed and nothing is added to what is owed: the issued §4 quotes two-way sync, templates and tracking, never a bulk send, so the committed V1 scope, price and timeline are unchanged and the CRM as quoted still sends one email at a time — the weekly prices included, per customer from the account page. If the checkpoint takes F37 into V1 that is added scope, agreed with the client at the time as a change request or against a compensating descope under R-ARCH-001; it is not absorbed silently (L-001). The newsletter half additionally depends on client answers on consent, unsubscribe, suppression and sending identity (OQ-62). WhatsApp integration remains outside V1, unchanged. Post-issue note added 16 September 2026 (second note of the day; not an edit to the issued text): lead triage researches each lead before ranking it, which supersedes the §A.3 cost line describing it and widens the §A.6 statement of what leaves the system for that task. §A.3.2 prices “triage and score one lead” as classification against fixed criteria on the cheaper model, and §A.6.1 states that for triage the minimum that leaves the system is the lead’s own fields. Both describe triage more narrowly than the requirement Empire Teas set: a lead’s priority is to be based on research into the company, not only on the words of the inquiry — which is the only way a priority over the 200-plus unreviewed backlog can mean anything, since the inquiry text alone cannot say whether a company imports tea, at what scale, into which market, or against which certifications. What changes in substance: before a lead is ranked, the company and the person named on it are looked up in publicly available sources, and the priority, its confidence and its stated reason are formed from what is found, with the source of each finding recorded and shown. So for this one task, the company and contact name from the lead leave the system to a search service, in addition to the lead’s own fields; nothing else about §A.6 changes — no buyer correspondence is sent for research, findings that cannot be traced to a source are discarded rather than guessed, and a person still approves every action. This is research, not a purchased dataset: the paid data-enrichment subscription listed among the deferred Pipedrive add-ons stays outside V1, unchanged. Commercially: the cost effect is immaterial — under two US dollars a month at the expected volumes, inside the fixed monthly infrastructure fee in SOW ET-002 — and the capability was a requirement rather than an addition, so no change request arises; the added build effort is carried inside the committed window (R-ARCH-001). §A.3’s other per-action figures, the §A.6 posture as a whole, and the human-in-the-loop rule stand as issued. Recorded internally as PRD Rev 30. Post-issue note added 17 September 2026 (not an edit to the issued text): the two acceptance releases are re-grouped — the same scope in a different order. §6/§7 group the build as Phase 0 (core CRM: accounts, contacts, pipelines and Kanban boards, activities, dashboard, global search, email, migration) then 1a (the tea-export workflow), with 1b (the AI layer) and 1c (parity) after, andproject-timeline.md§2 mirrored that. At Gate 3 (2026-09-17) the delivery team, on the lead’s instruction and with the client’s stated priorities, re-grouped the two releases: the first release (Phase 1 UAT, early November) covers signing in with Microsoft, users and roles with the reporting hierarchy, accounts and leads migrated from Pipedrive with AI research, ranking, summaries and drafted follow-ups, follow-up tasks on leads and accounts, email and templates inside the CRM, record sharing, and the weekly price publishing and per-customer send; the second release (Phase 2 UAT, late November) carries pipelines, deals and the Kanban boards, activities on deals, samples, quotes and the tea-export workflow, dashboards, the capture application, calendar and contact sync, list views, attachments, custom fields, import/export and the ERP handoff. Consequences stated plainly: the Phase 1 acceptance round does not include deals, and Pipedrive remains the record for deals until the second release; the migration brings accounts, contacts, leads and their notes (deals are a fresh start, per the 13 September election). Nothing sold is added or removed — every §4 capability is in one of the two releases — so scope, price and providers are unchanged and no change request arises; the §6/§17 milestone payments attach to the two releases as re-grouped. Recorded internally as Gate 3 Run 1 (engagement-gate-runs.md) and PRD Rev 32; the client is to be informed by the lead (draft line inengagement-gate-runs.md § Gate 3 Run 1, not yet sent as of 27 September 2026). Post-issue note added 24 September 2026 (not an edit to the issued text): the access model described in the 7 and 11 September notes is simplified again, and the Managing Director approval rule is delivered as configuration. §2 (the access passage those notes cite as §3.2) states that access is controlled by “role plus granular permission levels (read-only, edit, AI-use, and execute)”; AI-use was already removed as a level on 11 September. The delivery team has now simplified the rest (lead decision, 24 September 2026; to be stated to the client — draft line inengagement-log.md, not yet sent as of 27 September 2026): each area of the system has two permissions, View and Manage — Manage includes View and every action in that area (create, edit, archive, share, change the owner, ask the AI for a draft, and that area’s own steps such as submitting a quote) — plus a few named permissions that stand alone: publishing the weekly prices, approving a below-floor quote, user and role administration, system settings, and using one’s own mailbox. So the read-only level is View and the edit level is Manage; the execute level stays separately grantable as the named permissions — sending email, publishing a price, approving a below-floor quote — while an area’s other committing steps (submitting a quote, generating the ERP handoff) come with Manage on that area rather than as a level of their own. The system no longer ships with the built-in Administrator role the 11 September note described: it starts with one ordinary, editable role, “IT Admin”, holding user and role administration, and one named account holding it — a test mailbox on the test and development systems, and an address Empire Teas’ IT confirms on the live system, which becomes a new client dependency. Empire Teas’ administrator then creates every other role; §2’s six named roles are the recommended starting set. The Managing Director’s approval of below-floor quotes (§2, §5, §13) is delivered as the recommended configuration — only the Managing Director’s role holds the approval permission — rather than as a rule the software enforces against an administrator: whoever holds user and role administration can grant any permission to any role, their own included. Every change made in the system, and every sign-in and sign-out, is recorded in an audit trail (who, what, when), so any such grant can be attributed afterwards; there is no audit screen in V1. The client-facing PRD, question list, dependency list and role-access document are restated alongside this note. Recorded internally as AD #34, AD #35 and AD #36. Post-issue note added 27 September 2026 (not an edit to the issued text; lead decision, 27 September 2026): the AI work the system runs on its own runs in the background, which supersedes one sentence of §A.6.1. §A.6.1 states that the minimum a task needs leaves the system “only when a person triggers that task”, and that “nothing is sent automatically in the background” in V1. That describes AI a person asks for — a drafted reply, the quote or brief assistant — which still runs only when that person triggers it. It does not describe the AI the system runs on its own, which the requirement needs and the build already does: each lead is researched, triaged, ranked and summarised as it arrives, account and morning briefings are prepared, and duplicates and spam are flagged, as background jobs (recorded internally as PRD Rev 30 and AD #32; the research inputs are those in the 16 September second note). What does not change: each task sends only the minimum it needs; there is never a bulk upload of the database; the never-sent list in §A.6.1 stands; and no AI output is sent, saved as a decision or acted on until a person approves it. Commercially nothing changes — the cost is inside the §A.3 per-action figures as corrected by the 16 September note. The client-facing PRD (Section 3.3) states this; the client is to be informed by the lead. Post-issue note added 27 September 2026 (not an edit to the issued text; lead decision, 27 September 2026): the per-permission visibility scope named in the 7 September note is replaced by one fixed rule. The 7 September note describes the reporting hierarchy as adding “a team-level visibility scope set per permission”. There is no longer a scope to set: every permission on a record type reaches the records the person owns, the records shared with them and the records of everyone below them in the reporting line, and one extra permission per record type (“all records”) reaches every record. The reporting hierarchy and the configurable roles are unchanged, and nothing in §2 or §3.2 is narrowed. Recorded internally as AD #37 / PRD Rev 43; the client role document (Who Can Do What, revised 27 September) states it, and the client is to be informed by the lead with the access-model line already owed. Post-issue note added 28 September 2026 (not an edit to the issued text; lead decision, 28 September 2026): the Pipedrive migration covers leads, organisations and people only. §4 lists the one-time migration as “leads, contacts, organisations, deals, notes, activities, and business-critical custom fields”, and the 17 September note says the migration brings “accounts, contacts, leads and their notes”. Deals were already a fresh start at Empire Teas’ election (13 September). Pipedrive notes, activities and custom fields are not migrated either: Empire Teas had not asked for them, and they stay readable in Pipedrive, which remains available through the parallel run and until Empire Teas decides to switch it off. The migration still runs as §8 describes — a first load, then a final run after the data freeze — and the migration report still lets Empire Teas check that every lead, organisation and person came across. Every imported lead and account is owned at first by one Empire Teas user, who hands them out; the Pipedrive owner’s name is kept on each record for that. Recorded internally as PRD Rev 44; the client edition of the PRD (revised 28 September) states it. Post-issue note added 29 September 2026 (not an edit to the issued text): the §7 schedule and the AWS arrangement are replaced by the revised Statement of Work (SOW-ET-001 v1.1). §7’s indicative 3.5-to-4-month estimate, and the dates derived from it (sign-off target Tue 15 December 2026), are replaced by: Foundation, including the cloud environments and the deployment pipeline, Fri 16 October 2026; Phase 1 Fri 27 November 2026; Phase 1 UAT Thu 10 December 2026; Phase 2 (with the capture app) Fri 8 January 2027; Phase 2 UAT and stabilisation, and sign-off, Fri 29 January 2027. The dates assume Empire Teas provides the AWS environment, with the required access, by Fri 2 October 2026, and move day for day with it. §6’s review window is five business days for Phase 1 and interim deliveries and ten business days for Phase 2 and final acceptance. The AWS account is Empire Teas’ own, paid by Empire Teas’ card and billed to Empire Teas directly by AWS, which returns to this proposal’s original “billed direct to the account holder” position and supersedes SOW ET-002 (the Altrium-run account at USD 650/month) and the 13 September note above that recorded it; Altrium builds the infrastructure inside that account, with administrator access Empire Teas granted; Altrium’s monthly charge is the USD 360/month support fee, and no change to the USD 15,500 one-time fee has been recorded. The current schedule isproject-timeline.md; the delivery records areengagement-log.md(29 September 2026). Post-issue note added 30 September 2026 (not an edit to the issued text; lead decision, 30 September 2026): the synthesis model is Claude Sonnet 5.5, where Bedrock processes AI requests is stated, and the CRM’s domain is delegated to the project’s AWS account. §A.3 names Claude Haiku 4.5 and Claude Sonnet 5 (priced there at $3 / $15, since superseded by Sonnet 5’s permanent $2 / $10). Claude Sonnet 5.5 reached Amazon Bedrock on 28 September 2026 at the same $2 / $10, and replaces Sonnet 5; Haiku 4.5 stays until Claude Haiku 5.5 reaches Bedrock, which then replaces it. No commercial change arises. Where requests are processed: §A.3.1 and §A.6.5 say the models are served “inside Empire Teas’ own AWS account and region”. Every AI call is made from inside Empire Teas’ own AWS account, but, as the delivery team calls the models today, neither is processed in one Region only: Haiku 4.5 is processed in one of three US Regions, and Sonnet 5.5 — offered for now only through Bedrock’s global routing — may be processed in any commercial AWS Region worldwide. AWS already offers a US-only route for Sonnet 5 and Haiku 4.5, at about 10% more. Whether buyer data may be processed outside the US is Empire Teas’ decision, part of the data-privacy question the proposal left open (§A.6.5), and is asked in the client question list (Q18). Unless Empire Teas asks otherwise, the delivery team uses the worldwide route; moving to a US-only route, once AWS offers one for Sonnet 5.5, is a configuration change. Domain: the CRM’s addresses sit undercrm.empire.lk, which Empire IT delegates to the project’s AWS account with four NS records, so no further DNS changes are needed from Empire IT unless CRM email is sent from an@empire.lkaddress. The records areai-provider-strategy.md§4b andenvironments-and-identity.md§1.1. Post-issue note added 5 October 2026 (not an edit to the issued text; lead decision, 2 October 2026): the dependency-vulnerability control stops releases, not every build. §A.4 states “a build pipeline that fails on any dependency vulnerability at CVSS 7.0 or above, with exceptions requiring written justification and an expiry date”. The scan runs on every ready pull request and on the backend’s nightly run, and, once the development deploy is built, on every image it builds; it reports findings at CVSS 7.0 or above without failing them. A UAT or production release stops before deploying on any such open finding unless it has a written exception with an unexpired expiry date. Everything Empire Teas receives still meets the stated control; the change only stops a newly published advisory against an untouched dependency from blocking development. The record is AD #38 in.claude/CLAUDE.md. Post-issue note added 8 October 2026 (not an edit to the issued text; lead decision, 8 October 2026): AI drafts for leads are made when a person asks for them, not in advance. §3 (The value case) says V1 “triages that entire backlog on day one and puts a drafted, human-approved re-engagement in front of the team”. The backlog is still triaged on day one — every migrated lead researched, ranked and summarised, with a filter for the leads that have gone quiet — and every lead, new or migrated, has a button that drafts a reply or re-engagement in Empire Teas’ voice and shows it for review before anything is sent. What changes is only that the system does not draft every lead ahead of time. A person still approves every draft. No price or date change. The Phase 1 acceptance criteria (E5, E8) state it. Post-issue note added 9 October 2026 (not an edit to the issued text; lead decision, 9 October 2026): bulk email, a stretch candidate since the 16 September note, is committed to Phase 2. The issued text quotes email with two-way sync, templates and tracking (§4), places email campaigns among the later-phase Pipedrive add-ons (§ “Later phases”) and excludes a full marketing suite. On 9 October 2026 Empire Teas asked for newsletters, product promotions, trade-fair invitations, contact grouping and open and click figures, and the lead told them that Phase 2 includes bulk email — the weekly prices to every customer at once, and one message to a saved list of contacts for newsletters, promotions and invitations — with opens and clicks tracked for each send. This is added scope. The lead also told Empire that Phase 2’s features will be prioritised with them in a meeting at or before the Phase 1 hand-over, deciding which features they do not need; the work that moves out of Phase 2 there is the compensating descope, so the price and the dates are unchanged. The campaign product beyond the send — a segmentation builder, reporting across campaigns, an email designer, subscriber management — stays a later-phase item. Newsletters need Empire Teas’ answers on consent, unsubscribing and the sending address first (PRD Rev 57, OQ-62). Post-issue note added 9 October 2026, second note of the day (not an edit to the issued text; Altrium management’s decision, 9 October 2026): GuardDuty is not enabled for now. §A.2.2 prices “Threat detection (GuardDuty)” at $5 a month per environment, and §A.4 lists GuardDuty among the included security controls; the file-attachment scanning built on it (AD-O-4) was promised in the client-facing PRD (“uploads are validated and virus-scanned”). Altrium decided to go without GuardDuty — both account-wide threat detection and malware scanning of uploads — and to add it later if it is needed or Empire Teas asks for it. Uploads are still checked for type and size, and each is recorded as not scanned (never as clean). Turning scanning and threat detection on later needs no change to the software, only the AWS setup and two settings. While it is off, Empire Teas’ AWS bill is about $5 a month lower per environment thansow-cost-annexe.mdshows. Empire Teas is to be told before production (.forge/owed.md). Internal record:architecture.mdRev 52, PRD Rev 58, F38 spec Rev 7. Post-issue note added 11 October 2026 (not an edit to the issued text; lead decision, 11 October 2026): global search is not built. The scope summary, §2’s core CRM list, the §6/§7 phase plan and the parity line under “What success looks like” name global search, one search box across every kind of record. The delivery team dropped it (PRD Rev 60): each list screen — accounts, contacts, leads, users and every later list — carries its own search over the records the user reaches, and a user looking for a record goes to that list. The lead decided this is not raised with Empire Teas separately. No commercial change is recorded for it. Post-issue note added 11 October 2026, second note of the day (not an edit to the issued text; lead decision, 11 October 2026): the product-brief assistant is not built. §2’s AI layer lists a “product-brief assistant”, the AI that drafts a private-label product brief from a rep’s notes. The delivery team dropped it (PRD Rev 61): a rep fills the brief by hand, and the brief’s open-questions check is unchanged. Product briefs themselves (§2, tea-domain workflows) are still built. The lead decided this is not raised with Empire Teas separately. No commercial change is recorded for it. Post-issue note added 11 October 2026, third note of the day (not an edit to the issued text): corrections to the two notes above and to the 9 October GuardDuty note. (1) The passages the first two notes refer to are §3 Success criteria (the parity line naming global search), §4 Scope (the core CRM list naming global search, and the AI layer naming the product-brief assistant) and the §6/§7 phase plan; the notes’ “scope summary”, “§2” and “What success looks like” should be read as these. (2) The 17 September note’s “nothing sold is added or removed” no longer holds after 11 October: global search and the product-brief assistant are removed, as the notes above record. (3) The 9 October note says Empire Teas is to be told about GuardDuty before production; on 11 October the lead decided it is not raised separately — the client-facing PRD, revised 9 October, already states it. The client-facing PRD and the project timeline were revised on 11 October for the removed items; when they are sent is the lead’s decision. Post-issue note added 11 October 2026, fourth note of the day (not an edit to the issued text): the Pipedrive paid add-on equivalents are not planned for a later phase. §4’s Deferred (post-V1) list ends with “the Pipedrive paid add-on equivalents (website chatbot and live chat, email campaigns, data enrichment, LinkedIn integration, sales goals) pending the Pipedrive audit”. The delivery team dropped all five (PRD Rev 66): none is built in V1 and none is planned after it. They were never part of this proposal’s timeline or price, so no commercial change is recorded. Two things stay as agreed: the AI lead research from public sources (16 September note), and the bulk email committed for Phase 2 (PRD Rev 57), which was never the campaign product. The first 9 October note’s (bulk email) statement that the campaign product beyond the send “stays a later-phase item” no longer holds. If Empire Teas relies on one of the five today, it is raised through the change process. This is not raised with Empire Teas separately (decided 11 October 2026). Post-issue note added 13 October 2026 (not an edit to the issued text): the Pipedrive migration’s runs answer in the request, not on the background queue. §5’s architecture diagram shows the Pipedrive migration inside the background worker, and its prose says the migration “runs on a queue instead of making a user wait”. The import’s report (the dry run) and its writes (the first load and the final run after the data freeze) each answer in the request instead, every write run in one database transaction, while they fit the load balancer’s idle timeout; a write run that no longer fits becomes a background job (PRD Rev 70, F4-d spec FR-R5). The migration is run twice by the delivery team, so no user waits on it in daily work, and one transaction means a failed run leaves nothing half-written. Nothing changes in scope, price, timeline or providers, and no change request arises. Tharinda (lead) is asked to confirm on the F4-d spec PR. Post-issue note added 14 October 2026 (not an edit to the issued text; lead decision, 14 October 2026): the development environment is not scheduled off outside working hours while development is under way. §A.2.1 says “Development is scheduled off outside working hours”, and §A.2.2’s development figures assume it. The delivery team runs work against development overnight, so development stays on around the clock until development is complete; then it is scaled down, scheduled off, or decommissioned. Measured on 14 October 2026, development costs about USD 62 a month on Empire Teas’ AWS bill, against the cost annexe’s estimate of about USD 34; the annexe’s development row is to be re-derived (.forge/owed.md). Production is unaffected. Post-issue note added 14 October 2026, second note of the day (not an edit to the issued text; lead decision, 14 October 2026): two controls are optional in the development environment. §A.5 and §A.6.3 retain database backups “3 days in development”; §A.4 and §A.2.3 promise no long-lived keys for people and single sign-on with MFA for administrative access. In development, backups are optional, and the team’s shared command-line identity for development (svc-harness) uses an access key, with MFA on its console password optional; both are left to the DevOps team. Production is unchanged: 14-day backups with point-in-time recovery, and personal logins with MFA and no long-lived keys for people. Post-issue note added 15 October 2026 (not an edit to the issued text; lead decision, 14–15 October 2026): more of the account baseline is left out of the development environment. §A.2.1 and §A.4 describe the security baseline for each account, including an immutable audit log of infrastructure actions. In development there is no CloudTrail trail of its own — AWS’s 90-day CloudTrail event history, which every account has at no cost, is the record — and no root-user protection, account password policy, alternate contacts or account-wide S3 public-access block (each development bucket blocks public access on its own). Development’s Terraform state stays in the management account, and its AWS Budgets are arranged by the DevOps team with Empire Teas. Production is unchanged: it gets the full baseline, including the immutable audit log, and budgets with alerts. Post-issue note added 16 October 2026 (not an edit to the issued text; lead decision, 16 October 2026): the database password is not rotated on a schedule. §A.2.3 includes “secrets held in AWS Secrets Manager with rotation”, and §A.4’s Secrets Manager row lists “database credentials, with rotation”. The database password is still generated and held by AWS (RDS manages it in Secrets Manager, so no person sees it and it is never in source control), but in every environment it is not changed automatically on a schedule. It is changed by hand, with the application redeployed, if a leak is suspected or someone who could read it leaves. A scheduled change that the running application does not pick up cuts it off from its data, and the database is reachable only from inside the AWS network, so the scheduled change protected little. This is planned, and applied by Altrium’s DevOps team. The other secrets (the Microsoft sign-in certificate keys) are unchanged. Post-issue note added 17 October 2026 (not an edit to the issued text; lead decision, 17 October 2026): the development database is not encrypted at rest, and production’s is undecided. §A.4 promises encryption at rest, with customer-managed KMS keys “for database, storage, and backups”. In development the database stays unencrypted, as it was created, and development has no database or backups key; the file storage and the mailbox tokens are still encrypted with their own customer-managed keys. Development holds test data only. Whether the production database is encrypted at rest is not yet decided. Production holds Empire Teas’ buyer data, so a decision not to encrypt it would depart from §A.4 and would be recorded here and agreed with Empire Teas. Encryption costs nothing to switch on when the production database is first created, but adding it later needs a rebuild with downtime, so the decision is due before production is set up. Post-issue note added 21 October 2026 (not an edit to the issued text; lead decision, 21 October 2026): account briefings are prepared when a person asks for one, not in the background. The 27 September note lists “account and morning briefings are prepared” among the AI work the system runs on its own. The account briefing is instead prepared when someone whose role may edit the account presses a button on it, as the Phase 1 acceptance criterion E6 states (“a user asks for a briefing”); a role that may only view accounts does not start it. This sends less data, less often, than the 27 September note allows: §A.6.1’s original sentence (“only when a person triggers that task”) holds for the account briefing. The morning briefing is unchanged. Because no order data reaches the CRM in V1 (§A.6.1 is unaffected; past orders are an open question with Empire Teas), the briefing does not cover past orders, and on an account with no recent contact it suggests a re-engagement in place of a reorder prompt. No price or date change; the per-briefing cost stays inside §A.3.2’s figure. Post-issue note added 26 October 2026 (not an edit to the issued text; lead decision, 26 October 2026): the AI features run on Claude Haiku 4.5 until AWS allows the larger model on Empire Teas’ account. §A.3.2 assigns Claude Sonnet (Sonnet 5.5 since the 30 September note) to the writing and synthesis actions: email and sequence drafts, the quote assistant, the account briefing, the morning briefing and trade-show lead research. AWS has not made Sonnet 5.5 available on Empire Teas’ account. On 26 October 2026 it confirmed that a new account keeps lower Bedrock limits until it has a few months of usage and spend, and that it lifts them on its own review, with no date. Until then every AI feature runs on Claude Haiku 4.5, and the CRM will pace its AI requests to stay within the account’s limit. The design is unchanged: each feature moves back to Sonnet 5.5 by a configuration change once AWS allows it, after a check of the feature’s output on that model. Drafts and briefings may read less polished on Haiku 4.5 meanwhile; a person still reviews and approves every AI output. No price or date change; Haiku 4.5 costs less per request than Sonnet 5.5, so AI spend falls rather than rises against §A.3.2’s figures. Post-issue note added 27 October 2026 (not an edit to the issued text; lead decision, 27 October 2026): how two infrastructure passages are read in production. (1) §A.6.2 says AI calls are “server-to-server from the private network”. To avoid the cost of a NAT gateway (§A.2.1 prices none), the CRM’s backend runs in a public subnet of the private network, with a public address and a firewall rule that admits only the load balancer, so its calls to the AI service go server to server over TLS from that subnet, not from a private one. This departs from §A.6.2’s wording. The rest of the sentence holds: the browser never talks to the AI service, and no key reaches a user’s device. The database stays in private subnets and is not reachable from the internet. (2) §A.4 lists CloudTrail as an “immutable audit log of every AWS API call, to a central archive”. Each account keeps its own archive: production’s trail writes to a bucket in the production account, locked for one year so that records cannot be changed or deleted (only one emergency role can override the lock). No price or date change. Post-issue note added 28 October 2026 (not an edit to the issued text; lead decision, 28 October 2026): the production database is encrypted at rest. This settles the open point of the 17 October note. The production database, its automated backups and its snapshots are encrypted at rest through AWS KMS with AWS’s own managed key for the database service, rather than a key of the project’s own: the same encryption at no charge, the difference being that a key of the project’s own would also let its policy limit who may decrypt; file storage and the stored mailbox tokens keep their own customer-managed keys, as in development. §A.4’s encryption at rest and its KMS row are met: the issued row names KMS encryption keys for the database, storage and backups, not keys of the project’s own; the 17 October note’s description of §A.4 as promising “customer-managed” keys went further than the issued text. One further consequence: a snapshot encrypted with AWS’s managed key cannot be shared with another AWS account; restores within the account and copies to another region still work. The development and UAT databases stay unencrypted; they hold test data only. No price or date change. Post-issue note added 4 November 2026 (not an edit to the issued text; lead decisions, 2 and 4 November 2026): an addendum to the 30 September note — the background AI work uses Haiku 4.5’s worldwide route, and processing in any commercial AWS Region is accepted. The 30 September note says Haiku 4.5 is processed in one of three US Regions. From F43, the AI work the system runs on its own for each lead — its summary and spam check, and its research — calls Haiku 4.5 through Bedrock’s global routing, which may process a request in any commercial AWS Region worldwide, as Sonnet 5.5’s route already does; the AI calls a person waits for (inquiry extraction, drafts, the account briefing) stay on the US-only route. The reason is capacity: AWS gives each route its own request quota, and the account is held at 10 requests a minute on each until AWS lifts its new-account limits (26 October note), so background work on its own route no longer competes with people’s calls. The global route is billed at list price, about 10% below the US-only route, so no commercial change arises. Where requests are processed is also settled, by lead decision of 4 November 2026: processing in any commercial AWS Region is accepted. This replaces, for the processing location, §A.3.1’s and §A.6.5’s “inside Empire Teas’ own AWS account and region”. Every call is still made from Empire Teas’ own account, and on every route data between Regions stays on the AWS network, encrypted in transit. The model provider has no access to prompts or answers, and neither AWS nor the provider uses them for training (AWS Bedrock documentation, read 4 November 2026). Client question Q18, which asked this on 30 September and was not answered, is answered with this decision and closed. A US-only route remains a configuration change if Empire Teas ever needs one. The records areai-provider-strategy.md§4b and §5 row 9, andcloud-gated-validation.md§D. Post-issue note added 12 November 2026 (not an edit to the issued text; agreed with Empire Teas at the progress meeting of 10 November 2026, lead decision recorded 11 November 2026): the trade-show capture application moves from the second release to the first. The 17 September note placed the capture application in the second release (Phase 2). Empire Teas asked for it in time for its January 2027 trade fairs, and it is now delivered in Phase 1, to Phase 1’s dates, with its scope unchanged (the full application of §4). The capture-application review, which the revised Statement of Work holds during Phase 2, is proposed for Fri 20 November 2026, to be confirmed by both parties. The live phase assignment is inclient/project-timeline.mdandclient/phase-1-acceptance-criteria.md(Section L). Post-issue note added 12 November 2026 (not an edit to the issued text; from Empire Teas at the progress meeting of 10 November 2026, lead decision 12 November 2026): Empire Teas is replacing IFS. §ERP handoff and §A name IFS as Empire Teas’ ERP and the automated link as a priced follow-on once IFS API availability is confirmed. Empire Teas has said IFS is being replaced by an ERP not yet chosen. V1 is unchanged: it produces the handoff document. Any automated link remains a separately priced follow-on, and waits until Empire Teas names the new ERP and gives its details. Post-issue note added 12 November 2026 (not an edit to the issued text; lead decision, 12 November 2026): the capture application announces finished leads inside the app, not by email. §4 says each finished lead “is announced over WhatsApp and email rather than mobile push”, and §A.7 carries the WhatsApp Business Platform and Amazon SES for those notifications. WhatsApp was already left out (the 18 August decision recorded in the cost annexe); email is now left out too. The application shows each finished lead, and each lead a colleague converts or rejects, in a notifications list and an on-screen notice inside the app, which reaches the representatives who use it at the show; it sends no email and no push notification. The cost annexe’s lead-notification line falls to nothing. No change to price or timeline; no change request arises. Not raised with Empire Teas separately: this note and the client-facing PRD are the record (client/client-facing-prd.mdSection 3.5). Post-issue note added 12 November 2026 (not an edit to the issued text; decided by the delivery team on 12 November 2026 and accepted by the lead the same day): a captured lead is reviewed in the capture application before it reaches the CRM. §4 says “the captured leads feed straight into the CRM’s lead pipeline”. Each captured lead is now read and researched as §4 describes, then waits in the application until a person reviews it: converting it creates the lead in the CRM, with its trade-show source, owned by the person who converted it; rejecting it keeps it in the application only. This is in keeping with §4’s rule that “a person decides who is worth contacting”. Cards and badges are captured one at a time, and using the application needs its own permission, which an administrator gives to the roles of the people who work trade fairs. No change to price or timeline; no change request arises. Not raised with Empire Teas separately: this note and the client-facing PRD are the record (client/client-facing-prd.mdSection 3.5; PRD Rev 95). Post-issue note added 12 November 2026 (not an edit to the issued text; lead decision, 12 November 2026, after Empire Teas saw the plan screens at the progress meeting of 10 November): the capture application also plans each show before it starts. §4 describes capturing and researching contacts met at the stand, and does not name planning before a show. The application now also holds a plan for each show: its exhibitors, loaded from the show’s list or added one by one, each researched and ranked for a person to review, with visit notes and a visited mark for each stand. It is delivered in Phase 1 with the rest of the application (client/client-facing-prd.mdSection 3.5;client/phase-1-acceptance-criteria.mdL8 and L9; PRD Rev 96). Researching every exhibitor adds AI and web-search use that the running-cost annexe does not yet show as its own line; whether this addition changes the price, and the annexe line for it, are not yet decided. Version: 0.5 Date: 5 August 2026 Prepared by: Altrium Prepared for: Empire Teas (Pvt) Ltd.
1. Executive Summary
Empire Teas runs its front-office sales process (buyer inquiries, private-label product development, sample rounds, quoting, follow-up, and export coordination) on Pipedrive plus a spread of spreadsheets, email, and WhatsApp. That toolset was not built for the tea export business, and the gaps cost real revenue: more than 200 leads sit unreviewed, follow-ups are missed, sample and quote status is not visible end to end, and dormant reorder-ready accounts go quiet.
Altrium proposes to design and build a purpose-built B2B sales CRM that replaces Pipedrive without losing anything the team relies on, models the tea export and private-label workflow as a first-class process, and embeds a human-in-the-loop AI layer into the repetitive text and document work. A person reviews and approves every AI action, and a person sets all pricing. The AI never acts on its own.
Version 1 is delivered as a fixed-scope engagement. This proposal sets out what is built, how it is delivered and accepted, how the switch off Pipedrive is managed, what Empire Teas owns at the end, the assumptions and dependencies the plan rests on, and the commercial terms.
2. Understanding of Your Needs
Empire Teas is among Sri Lanka’s top five tea exporters, shipping to more than 70 countries across two commercial lines: bulk tea (sold by weight) and private label (manufactured under the buyer’s own brand). The sales process is relationship-heavy, sample-driven, and touches several internal functions along the way: sales, tea tasting and R&D, quality, export documentation, and finance.
Three things shape this project:
- Nothing can be lost in the move off Pipedrive. The team lives in Pipedrive today, including sending and receiving email from within it. The new system must replicate the day-to-day capabilities the team depends on, and the switch itself must be safe.
- The tea export workflow must be modelled directly. Multi-round sample development, certification checks, export-document sets by destination, and quote readiness gating are core to the business and are not served by any off-the-shelf CRM.
- AI should remove repetitive work, under human control. Reading and extracting inquiries, triaging leads, drafting replies in Empire Teas’ voice, briefing accounts, and checking quote readiness are all high-value assists, provided a person approves every output.
The system is used day to day by these roles in V1: Sales Rep, Sales Manager, R&D and tea taster, a dedicated Tea Room pricing team, and Administrator, with an Executive read-only view for management. Access is controlled by role plus granular permission levels (read-only, edit, AI-use, and execute). Quote pricing follows a required gross-profit margin: representatives quote freely at or above it, and any quote below it is approved by the Managing Director. Quality, export documentation, and finance are part of the wider business process the CRM supports, but are not separate system roles in V1; their views are covered by the Sales Manager and Administrator roles, and the finance and export paperwork itself remains in the ERP and with the export team.
A separate ERP programme is underway to serve as the transaction backbone (inventory, production, procurement, finance, and confirmed orders). That ERP is out of scope for this project. This CRM is the front-office productivity layer that hands a won deal over to the ERP.
3. Objectives, Success Criteria and Value
Success criteria
The engagement is successful when:
- The team can run a full working day without Pipedrive, including sending, receiving, and replying to email inside the new system, and Pipedrive can be safely decommissioned.
- Both sales pipelines (Private Label and Bulk) operate with drag-and-drop stage changes, weighted forecasting, and per-stage stale-deal alerts.
- The tea-domain workflows work end to end: multi-round sample tracking, quote readiness gating with human pricing approval, product briefs, certification-mismatch flagging, and destination-based export-document checklists.
- The AI features run under human review: lead triage and scoring, inquiry extraction, account briefings, quote assistance, dashboard briefings, and email and sequence drafting. The AI never sets price.
- Pipedrive data, including the 200-plus lead backlog, is migrated cleanly, deduplicated, and triaged, and a won deal produces a valid handoff to the ERP.
- The everyday parity features are in place: global search, filterable list views with bulk actions and export, file attachments on records, and admin-configurable custom fields.
The value case
The strongest single number already in Empire Teas’ hands is the 200-plus unreviewed leads sitting in Pipedrive today. These are inbound buyers who were never worked. V1 triages that entire backlog on day one and puts a drafted, human-approved re-engagement in front of the team.
To be completed with Empire Teas’ figures: if even [conversion rate] of the 200-plus backlog converts at an average deal value of [average deal value], that is [recovered revenue] in the first months of use. Against a one-time build cost of [V1 price] and the [annual Pipedrive licence cost] saved every year thereafter, the system pays for itself on the backlog alone, before counting faster response times, fewer missed follow-ups, and recovered dormant accounts.
The other outcomes to baseline and track with Empire Teas are: average inquiry response time, overdue follow-ups, samples stuck without feedback, quotes aging past threshold, and dormant high-value accounts re-engaged.
4. Scope
In scope (V1)
Thirty features across the following delivery themes, plus a companion mobile application:
- Core CRM (system of record): accounts and contacts; two independent pipelines with Kanban boards; activities and reminders; authentication, users, roles, and ownership; the operational dashboard and KPIs; global search.
- Migration and import: one-time migration of Pipedrive data (leads, contacts, organisations, deals, notes, activities, and business-critical custom fields); an ongoing spreadsheet import and export tool for Excel-held data.
- Tea-domain workflows: multi-round sample tracking; quotes with an eight-item readiness checklist and pricing-approval gate; private-label product briefs; certification-mismatch checking; destination-based export-document checklists; stale and dormant-account alerts.
- Human-in-the-loop AI layer: lead triage with scoring and drafted replies; inquiry extraction from email; account briefings; quote assistant; dashboard morning briefing and “what is stuck” read; product-brief assistant.
- Communication and parity: email in the CRM (two-way sync, templates, tracking); activity calendar and calendar sync; contact sync; configurable pipelines (stages, probability, rotting); outreach sequences; products catalog with bulk pricing; list views, saved filters, and bulk actions; file attachments; admin custom fields.
- Handoff: CRM-to-ERP handoff document produced on a won deal (see the ERP note below).
- Trade-show lead capture: a mobile lead-capture application (an installable progressive web app) for exhibitions and trade shows, feeding scanned, researched, and ranked contacts into the CRM’s lead pipeline (see the note below).
Migration boundary
The structured migration covers Pipedrive. Data held in Excel is brought in through the spreadsheet import tool. Information currently living in WhatsApp threads and personal email is not retro-migrated in bulk; going forward it is captured through the CRM’s email sync and normal data entry. This boundary is confirmed against the real data during the Pipedrive audit.
ERP handoff
V1 produces a structured handoff document when a deal is won, on a dedicated integration seam, and the team enters that into the ERP. Automated CRM-to-ERP integration is not in V1. The current ERP is IFS; once IFS API availability is confirmed, automated integration against IFS is offered as a defined, separately-priced follow-on that plugs into the same seam. The new ERP (targeted for approximately 2027) is deferred.
Trade-show lead capture (mobile app)
V1 includes a companion mobile lead-capture application for exhibitions and trade shows, delivered as an installable progressive web app so it reaches every representative’s phone through a link, with no app-store account, review cycle, or device-management overhead, and updates the same day it changes. At the booth, a representative photographs a business card or scans a badge QR during the conversation itself, optionally adding a short typed or dictated note of what was said (“wants organic samples before August, decides in Q4, price-sensitive”), and carries on; capture takes seconds and requires no other data entry. In the background the system reads the card, researches the company against public commercial sources (firmographics, role, recent news, and import-export signals), scores the contact against Empire Teas’ offering, and returns a ranked profile with a plain-language reason the lead matters and the specific hook to open with, typically within minutes and often before the representative has left the aisle. Each finished lead is announced over WhatsApp and email rather than mobile push, which is deliberate: that reaches every phone reliably, needs no home-screen install, and lands where leadership already works. The captured leads feed straight into the CRM’s lead pipeline, where they are triaged, scored, and drafted against by the same human-in-the-loop AI layer as any other inbound inquiry. This closes the gap the discovery identified directly, where exhibition contacts receive one conversation and are then lost to the stack of cards. The application captures and researches only; a person decides who is worth contacting and approves every message that goes out.
Out of scope
Explicitly not built in this project: ERP functionality (inventory, production, procurement, finance, confirmed orders); automated CRM-to-ERP integration (a priced follow-on, above); estate, plantation, and manufacturing operations; auction bidding and tea-grading AI; native mobile apps distributed through the Apple App Store or Google Play (the CRM is a responsive web application, and the trade-show capture app described above is an installable progressive web app; a store-distributed version can be added later on the same codebase); a general no-code automation builder; a third-party integration marketplace; a full marketing suite; direct submission to SLTB, ASYCUDA, or customs; and any autonomous AI decision-making.
Deferred (post-V1)
Intended for a later phase and not part of this proposal’s timeline or price: automated CRM-to-ERP integration (as above), an automated tea-auction price feed (a public API is confirmed available, to assist the weekly pricing workflow without ever auto-setting a price), a retrieval-based knowledge and compliance assistant, automated inbound-email triggering, multilingual AI extraction, store-distributed native mobile apps, per-office data partitioning, and the Pipedrive paid add-on equivalents (website chatbot and live chat, email campaigns, data enrichment, LinkedIn integration, sales goals) pending the Pipedrive audit.
5. Proposed Solution
A custom, single-tenant web application built specifically for Empire Teas, comprising:
- A structured data model for the tea export domain (accounts, contacts, leads, deals, pipelines and stages, product briefs, samples, quotes, a products catalog with effective-dated bulk pricing keyed to client-specific buying standards, activities, emails, and the reference data behind them).
- Two purpose-built pipelines seeded as editable defaults (Private Label and Bulk), with admin-configurable stages, per-stage win probability, and per-stage idle-deal alerting.
- Email, calendar, and contact integration with Microsoft 365, so mail auto-links to the right contact, deal, and lead, and the team works inside the CRM.
- A human-in-the-loop AI layer that reads, extracts, summarises, drafts, and checks, always presenting a reviewable draft that a person approves before anything is saved, sent, or acted on.
- A clean boundary to the ERP: the CRM owns everything up to the confirmed order and hands a won deal over.
It is a single-instance system serving all offices (Sri Lanka, Kenya, Netherlands, Shanghai, Spain), with access controlled by role and record ownership and a region tag for filtering and dashboards.
How it fits together
flowchart LR
reps(["Sales team<br/>five offices, in a browser"])
booth(["Representative<br/>at an exhibition"])
subgraph cloud["Empire Teas cloud account"]
direction TB
fe["<b>Frontend application</b><br/>React / Next.js<br/>screens, boards, dashboards"]
subgraph be["<b>Backend application</b> — one codebase"]
direction TB
api["API and business rules<br/>roles, permissions, approval gates,<br/>AI orchestration"]
worker["Background worker<br/>mail sync, AI batch jobs,<br/>alerts, Pipedrive migration"]
end
db[("CRM database<br/>PostgreSQL")]
files[("File attachments")]
queue[["Job queue"]]
end
pwa["Trade-show capture app<br/>installable on a phone"]
ai["AI service"]
m365["Microsoft 365<br/>mail, calendar, contacts"]
erp["ERP<br/>out of scope"]
reps --> fe --> api
booth --> pwa --> api
api --> db
api --> files
api --> queue --> worker
worker --> db
worker <--> m365
api --> ai
worker --> ai
api -. "handoff document<br/>on a won deal" .-> erp
Reading the diagram: everything inside the outer box runs in Empire Teas’ own cloud account.
Two applications are built, not three. The frontend renders every screen the team uses. The backend holds all business rules, role permissions, and the approval gates, and it is the only component that touches the database. The background worker is part of that same backend application, not a separate system: it is the same codebase, deployed alongside the API so it can be scaled independently, and it exists so that slow or scheduled work (mailbox synchronisation across five offices, AI batch jobs, the Pipedrive migration, stale-deal alerts) runs on a queue instead of making a user wait. Keeping it in one codebase is a deliberate choice at this team size: a separate service would add deployment, versioning, and operational overhead for no benefit at 10 to 20 users.
The three things outside the box are the AI service, Microsoft 365, and the ERP. The ERP boundary is a document handoff, not a live connection, in V1. Depending on the data-governance decision in Appendix A.6, the AI service can also run inside the same cloud account rather than outside it.
AI governance. Every AI-generated action is a draft that requires human approval before it sends or saves. Quote pricing follows a required gross-profit margin, and any quote below that margin is approved by the Managing Director. Compliance, export, and quality decisions remain human-owned. The AI runs on secure, reputable cloud infrastructure, and the specific data-handling arrangements are presented for Empire Teas’ formal approval at the contract stage.
You have already seen it
A working prototype of the designed system has been built and approved, and this proposal is grounded in it. It demonstrates the core screens: the operational dashboard with the AI morning briefing, the two pipeline boards, the account 360 view, the multi-round sample tracker, the quote readiness screen, and the bulk-pricing catalog with price history. This is a custom build you have already seen working, not a concept on paper.
Insert three to four prototype screenshots here (dashboard, a pipeline board, the sample tracker, the bulk-pricing screen).
6. Delivery Approach and Milestone Acceptance
Delivery is phased so value lands early and each phase builds on a stable base:
- Foundation sets up the repositories, build and deployment pipeline, authentication, the core data model, and the environments.
- Phase 0 delivers the core CRM (accounts, contacts, pipelines, activities, dashboard), email in the CRM, global search, and the Pipedrive migration, so the team can begin working in the system.
- Phase 1a adds the tea-domain workflows (samples, quotes, product briefs, certification checks, export-document checklists), the products catalog and bulk pricing, pipeline configuration, and the stale and dormant-account alerts.
- Phase 1b adds the human-in-the-loop AI layer and outreach sequences, built on the now-clean core data.
- Phase 1c closes the remaining parity conveniences: calendar and contact sync, list views, file attachments, admin custom fields, spreadsheet import and export, and the ERP handoff document.
- UAT and stabilisation runs regression, a client user-acceptance cycle, fixes, and sign-off.
Milestone acceptance. At the end of each phase, Altrium demonstrates the delivered work against that phase’s agreed acceptance criteria to a named Empire Teas signatory. Empire Teas has a review window of [review window, e.g. five business days] to accept or to raise defects against the criteria. If no response is received within the window, the phase is deemed accepted. Payment milestones (Section 17) trigger on acceptance. Quality gates (linting, automated tests, and a review of the work against its specification) apply throughout, ahead of each demo.
7. Indicative Timeline
The figures below are Altrium’s engineering estimate for building the features. They are a range on measured delivery baselines, not a fixed commitment, and they exclude client-response time, UAT, and training, which are shown separately.
| Phase | Scope | Build estimate |
|---|---|---|
| Foundation | Repositories, CI/CD, authentication, core data layer, environments | 2 to 3 weeks |
| Phase 0 | Core CRM, email in the CRM, global search, Pipedrive migration | 2 to 3 weeks |
| Phase 1a | Tea-domain workflows, products catalog and pricing, pipeline config, alerts | 2 to 3 weeks |
| Phase 1b | AI layer and outreach sequences | 2 weeks |
| Phase 1c | Calendar and contact sync, list views, attachments, custom fields, import/export, ERP handoff | 1.5 to 2 weeks |
| Feature-complete | approximately 11 to 13 weeks | |
| UAT and stabilisation | Regression, client UAT, fixes, sign-off | 3 to 4 weeks |
| Signed-off V1 | approximately 3.5 to 4 months |
Three things govern the schedule:
- Week 1 begins on contract signature.
- Late client dependencies extend the schedule day for day. Where a named dependency in Section 14 (for example Pipedrive access, the pricing confirmations, the certification spreadsheet, or a milestone acceptance) is delayed, the affected milestone and everything after it move out by the length of the delay.
- The estimate is a range on Altrium’s delivery baselines. This project has no measured history of its own yet, so the numbers carry real variance and are treated as a range, not a promise.
8. Cutover and Go-Live
The move off Pipedrive is managed so the team is never without a working system:
- Parallel run. For an agreed period before go-live, the new CRM runs alongside Pipedrive. Pipedrive stays live and usable throughout, so there is always a fallback.
- Data freeze and final migration. A short, agreed freeze window is used to run the final migration so no records are created in Pipedrive mid-copy and then lost.
- Rollback position. Until Empire Teas signs off, Pipedrive remains the system of record and the team can fall back to it at any point.
- Decommission on sign-off. Pipedrive is only cancelled after UAT sign-off and an agreed stable period on the new system, at Empire Teas’ decision. Altrium provides the migration report so Empire Teas can verify nothing was left behind before cancelling.
9. Training and Documentation
- Administrator training on configuration: users and roles, pipelines and stages, reference data, the products catalog and pricing, email templates, and custom fields.
- End-user sessions by role for the sales team, delivered remotely to cover all offices (Sri Lanka, Kenya, Netherlands, Shanghai, Spain).
- A concise user guide and an administrator guide, delivered as part of handover.
Training is scheduled into the UAT and stabilisation window so the team learns the system on the build they are about to accept.
10. Deliverables
At the close of the engagement Empire Teas receives: the source-code repositories; running development, staging, and production environments; the one-time migration report; the administrator and user documentation; the AI data-handling document approved at contract stage; and a handover session with the team.
11. Team and Governance
The work is delivered by Altrium under a named delivery lead who is Empire Teas’ single point of contact for the engagement, backed by Altrium’s delivery process, AI-assisted engineering tooling, and quality gates on every change. Altrium stands behind the schedule as a company; all work lives in a repository Empire Teas can access and is documented as it is built, so the engagement is never dependent on a single individual’s memory.
Governance cadence. Empire Teas sees progress through a demonstration at the end of each phase (Section 6) and a monthly checkpoint covering progress against plan, upcoming dependencies, and any risks. This gives Empire Teas a standing view of whether delivery is on track and removes the possibility of a late surprise.
12. Ownership, Hosting and Data Protection
Ownership. Empire Teas owns the source code and all of its data. The code lives in a repository Empire Teas can access from the start of the build. Empire Teas is free to appoint another party to maintain or extend the system in future; there is no lock-in. On termination or completion, Empire Teas receives the full source, a complete database export, and all stored files.
Hosting. The system runs on AWS (containerised application on Fargate, a PostgreSQL database on RDS, file storage on S3, background jobs on SQS, and monitoring on CloudWatch). The AWS account arrangement (held by Empire Teas, or held by Altrium and transferred at the end) is confirmed in the commercial terms. The AI features call the Anthropic Claude API from the backend.
Data protection and security. Buyer contact data is personal data, and some buyers are EU persons, so it is handled with care: encryption in transit and at rest, automated database backups with an agreed retention period, a defined disaster-recovery position, role-based access control matching the roles in Section 2, and an agreed uptime expectation. The AWS region is chosen to suit the user base and data-protection considerations and confirmed at contract stage. Only the data needed for a given AI task is sent to the AI service, and the specific data-handling arrangements are documented for Empire Teas’ formal approval before the AI features are built.
Detail for budgeting and security approval. The infrastructure sizing assumptions, expected AI consumption and models, recommended AWS security services, backup and disaster-recovery design, AI data handling and retention architecture, third-party tools, and the resulting anticipated monthly operating cost are set out in Appendix A.
13. Assumptions
The proposal is built on the assumptions below. Most were confirmed in the 3 August 2026 requirements meeting (listed as such); the accepted approaches are settled with a defined fallback; and one item remains pending your input. Where a still-open item is later confirmed to be different, it is handled through the change process in Section 16 and may change scope, timeline, or cost.
Accepted approaches (decided, with a defined fallback):
- ERP handoff. V1 produces the manual handoff document. Automated integration against the current ERP (IFS) is a separately-priced follow-on, offered once IFS API availability is confirmed. If it is later commissioned and the APIs turn out not to be available, the handoff remains the manual document.
- AI hosting. The AI features run on secure, reputable cloud infrastructure. There is no on-premises or self-hosted model in V1. The specific data-handling details are presented for formal approval at the contract stage.
- AI review. Every AI-drafted action always requires human approval before it sends or saves. Any “automatic versus manual” option applies only to how work is routed to a reviewer, not to whether a person approves.
Confirmed in the 3 August 2026 requirements meeting (now part of scope):
- Bulk price structure. The priced unit is a client-specific buying standard with its own unique running number; a client buys several standards, each carrying its own weekly price.
- No volume tiers. Order size does not drive price in any structured way (only nominal movement), so there is no tier schema to build.
- Destination-market pricing. Price differs by destination market, which the model supports directly.
- Pricing approval. Representatives quote freely at or above a required gross-profit margin; any quote below that margin is approved by the Managing Director. The system holds cost and computes the margin.
- Payment terms. LC, Open Account, DA, CAD, and advance payments, set per customer by finance (held in the ERP) and shown as a field on the quote.
- Sample rounds. Typically one to two when matching an existing sample and six to seven when developing a new standard, rarely beyond ten.
- Roles. The six roles in Section 2, including a dedicated Tea Room pricing team and an Executive read-only view.
Still pending your input:
- Certifications. A product-to-certification mapping is needed for the certification-mismatch check. How it is stored is being confirmed on your side (with Kavi Amali and the QC team). Until it is provided, that one feature cannot be finalised.
14. Client Dependencies
Timely input from Empire Teas keeps the timeline on track; late input extends it day for day (Section 7). We ask for:
- Pipedrive access (an export or an administrator walkthrough) at the start, so the audit can confirm current pipelines and stages, custom fields, reports, automations, add-ons, record volumes, and everything the migration must preserve.
- The product-to-certification data (the one pre-kickoff item still outstanding after the 3 August meeting), so the certification-mismatch check can be built. Being confirmed on your side with Kavi Amali and the QC team.
- IFS details if the automated ERP integration follow-on is commissioned: whether its APIs or an export format are available.
- Mailbox connection. Each user connects their own Microsoft 365 mailbox to the CRM (a standard secure sign-in), so email, calendar, and contacts sync.
- Formal approval of the AI data-handling arrangements at the contract stage.
- A nominated point of contact and named acceptance signatory, and availability of the right people (sales, pricing, IT, compliance) as each phase requires.
15. Risks and Mitigations
| Risk | Mitigation |
|---|---|
| Client-response and access delays stretch the calendar (the largest timeline factor). | Front-load the Pipedrive audit and the outstanding answers; treat them as named dependencies; the day-for-day clause keeps the impact transparent rather than absorbed silently. |
| The product-to-certification mapping does not exist in a usable form. | The certification-mismatch feature depends on it (Section 13, 14). If it cannot be provided, that feature is descoped or simplified to a manual checklist, confirmed with Empire Teas before the catalog is built. |
| Pipedrive data quality or volume is higher than expected. | AI triage cleans and classifies before import; migration scope and the migration boundary (Section 4) are confirmed against the real data during the audit. |
| IFS APIs are not available if the integration follow-on is commissioned. | Automated integration is out of V1; the manual handoff always works. The follow-on is only scoped once API availability is confirmed. |
| AI produces incorrect output. | Return-nothing-rather-than-guess prompting, structured validation, and human approval on every AI action. |
| Team adoption across five offices. | AI is an optional assist that reduces manual work; a familiar, seeded pipeline layout; and role-based training plus a user guide (Section 9). |
16. Change Requests
Any change to the agreed scope, whether from an assumption in Section 13 being confirmed differently or a new request, is handled through a simple loop: Altrium assesses the impact on scope, timeline, and cost in writing; Empire Teas approves it; and the plan and price adjust accordingly. Work proceeds on the agreed scope until a change is approved. Change-request rates are set out in the commercial terms.
17. Commercials
This section is to be completed by Altrium’s delivery lead with the commercial terms and figures.
- Pricing model: [fixed price for the V1 scope in Section 4, or time-and-materials, or a hybrid].
- Payment schedule: milestone-based, triggered by the phase acceptances in Section 6. Suggested split: an initiating payment on contract signature, a payment on acceptance of each delivery phase (Foundation, Phase 0, Phase 1a, Phase 1b, Phase 1c), and a final payment on UAT sign-off.
- Running costs (operational, ongoing): AWS hosting (Fargate, RDS, S3, SQS, CloudWatch, and the security services) plus AI provider usage, which scales with lead and email volume. AI cost is controlled by using the lower-cost model on high-volume paths and per-feature toggles. Sizing assumptions, the per-service breakdown, and the anticipated monthly total are in Appendix A. These costs are billed by AWS and the AI provider directly to the account holder and are not part of the build price. [Who holds and pays the infrastructure account, to be confirmed.]
- Automated ERP integration follow-on: [priced separately once IFS API availability is confirmed].
- Support and warranty: [warranty period after sign-off; how support is requested; response expectations by severity, to be defined]. Defects against the agreed acceptance criteria are fixed under warranty at no charge during the warranty period.
- Annual maintenance and support (AMC): an optional annual contract that begins when the warranty period ends. Scope, exclusions, and the commercial basis are set out in Appendix A.9; the annual fee is [AMC fee, to be inserted].
- Change-request rate: [to be defined].
- Included: the V1 scope in Section 4. Not included: the deferred items in Section 4 and any change raised after sign-off (handled under Section 16).
18. Acceptance and Next Steps
- Empire Teas reviews this proposal and confirms the assumptions in Section 13.
- Empire Teas provides the outstanding certification data and Pipedrive access (Section 14).
- Altrium finalises the commercial terms (Section 17) and issues the contract, including the AI data-handling arrangements for approval.
- On signature, Foundation begins (Week 1).
We would be glad to walk through any part of this proposal on a call.
Appendix A — Infrastructure, AI Consumption, Security and Operating Costs
This appendix supports Empire Teas’ internal approval and budgeting: infrastructure sizing, expected AI consumption and models, recommended security services, backup and disaster-recovery design, AI data handling and retention, third-party tools, the anticipated monthly operating cost, and the annual maintenance basis.
A.1 Basis of estimate
These are Altrium’s engineering estimates at the assumptions below, not a quotation: AWS and the AI provider bill on actual consumption, directly to whoever holds the accounts. The infrastructure is deliberately right-sized for a 15-user internal system and can be scaled up without redesign.
| Assumption | Value used |
|---|---|
| Named users / peak concurrent | 15 / 20 |
| Record volume | Low tens of thousands of accounts, contacts, deals, leads, activities |
| Environments | Two — development and production, in separate AWS accounts |
| Production availability | Business-hours availability, single availability zone. Recovery from an instance or zone failure is a restore from backup (see A.5), not an automatic failover |
| AWS region | To be confirmed; a Singapore or European region runs 10% to 25% above the mid-cost basis used here |
| Pricing basis | AWS and AI provider list pricing as at August 2026, before any committed-use discount |
A.2 AWS infrastructure
A.2.1 The shape proposed
The system is sized for what it is: an internal application for a single team, not a public service. That produces a deliberately compact estate.
- One ECS cluster per environment, running the frontend and the backend. The background worker runs inside the backend task rather than as a separate service, because it is the same codebase.
- Single availability zone, and no NAT gateways. Tasks run with tightly scoped security groups rather than behind a private-subnet-plus-NAT layer, which alone removes roughly $35 a month per environment.
- Two AWS accounts under one Organization. Separate accounts cost nothing and are what keeps development from ever reaching production data, so this isolation is kept even in a lean build.
- Development is scheduled off outside working hours and doubles as the UAT environment during acceptance windows. Day-to-day development runs on developers’ own machines at no cloud cost.
A.2.2 Sizing and indicative monthly cost
| Item | Production | Development |
|---|---|---|
| Backend application, including the background worker (Fargate, 0.5 vCPU / 1 GB) | $18 | $6 |
| Frontend application (Fargate, 0.25 vCPU / 0.5 GB) | $9 | included above |
Database (RDS PostgreSQL, single-AZ; t4g.small / t4g.micro) |
$31 | $14 |
| Load balancer | $19 | not used |
| S3, SQS, ECR, Route 53, Secrets Manager | $6 | $4 |
| CloudWatch and data transfer | $9 | $4 |
| Threat detection (GuardDuty) | $5 | $5 |
| Sub-total | approximately $97 | approximately $43 |
Total AWS: approximately $140 per month.
Storage grows with use: production starts at 50 GB and autoscales, adding roughly $6 per 50 GB per month. At the record volumes in A.1 this is not expected to be material in the first two years.
A.2.3 What the baseline includes, and what it defers
Included at the price above: encryption in transit and at rest, secrets held in AWS Secrets Manager with rotation, single sign-on with multi-factor authentication for administrative access, continuous threat detection, an immutable audit log of infrastructure actions, automated database backups with point-in-time recovery, and a build pipeline that fails on dependency vulnerabilities. Those are the controls Altrium treats as non-optional for a system holding buyer personal data.
The following are deliberately deferred rather than omitted by oversight, and each can be added independently at any point without rework:
| Option | What it adds | Monthly cost |
|---|---|---|
| Multi-availability-zone database | Automatic failover in under five minutes instead of a restore from backup taking one to two hours | +$30 |
| AWS WAF | Filters malicious web traffic at the load balancer using AWS managed rule sets plus rate limiting | +$15 |
| AWS Security Hub | Scores the accounts against the CIS AWS Foundations Benchmark and AWS Foundational Security Best Practices | +$10 |
| AWS Config | Records resource configuration over time and alerts on drift from the approved baseline | +$12 |
| Amazon Inspector | Scans container images for known vulnerabilities, complementing the build-pipeline scan | +$8 |
| Private subnets with NAT gateways | Removes public addressing from the application tier entirely | +$35 |
| Cross-region backup copy | Protects the backups themselves against the loss of an entire AWS region | +$8 |
| A separate staging environment | UAT no longer shares the development environment | +$45 |
Adding every item above brings the estate to roughly $300 per month. Altrium’s recommendation is to start at the baseline and add the multi-availability-zone database and WAF once the system is carrying real business — approximately $45 a month for the two largest risk reductions available.
A.2.4 Why this shape
The conventional enterprise configuration for a system like this — three environments, multi-availability-zone throughout, private subnets with NAT gateways, and the full security service suite — prices at approximately $520 per month. For a public-facing service carrying live transactions, that is the right answer.
This is an internal application used by fifteen people during business hours. At that profile the conventional build spends roughly $4,500 a year on redundancy and monitoring ahead of any evidence that the system needs it. Altrium’s position is that the sensible order is the other way round: start at the baseline, run the system on real data, and add the specific controls in A.2.3 as real usage justifies them. Nothing in the baseline forecloses that — every option listed is additive and can be enabled without rework or downtime.
Further levers, once real usage is known: a one-year AWS Savings Plan on steady-state compute (20% to 30% on those lines), and right-sizing after the first two months in production.
A.3 Expected AI consumption, providers and models
A.3.1 The provider is a configuration choice
Every AI feature calls the model through one internal provider interface, so the application does not know which model answers. Changing provider is configuration plus a round of quality testing, not a redesign.
| # | Option | What it means | Relative cost | Best when |
|---|---|---|---|---|
| 1 | Anthropic Claude API (recommended default) | The backend calls Anthropic directly. Simplest to operate. | Baseline | Empire Teas is comfortable with a reputable external AI provider under commercial terms |
| 2 | Claude on Amazon Bedrock | The same models served inside Empire Teas’ own AWS account and region, under the existing AWS agreement. No separate AI vendor. | Baseline, plus roughly 10% | Data governance is the deciding factor |
| 3 | Google Gemini (direct or via Vertex AI) | A different provider at lower token prices; adds a Google Cloud footprint or a second vendor. | Roughly one third of baseline | Token cost is the deciding factor, or a Google Cloud relationship already exists |
Published prices per million tokens (verify at contract stage): Claude Haiku 4.5 $1 / $5; Claude Sonnet 5 $3 / $15; Gemini 2.5 Flash-Lite $0.10 / $0.40; Gemini 2.5 Flash $0.30 / $2.50; Gemini 2.5 Pro $1.25 / $10.
Altrium will run a side-by-side quality comparison on real Empire Teas inquiries during Phase 1b and present the results before the provider is locked.
A.3.2 Cost per AI interaction
AI is billed on use, so rather than assume Empire Teas’ volumes, the table gives the cost of each individual AI action. Multiply by expected usage to budget.
| AI action | Model | Why this model | Cost per interaction |
|---|---|---|---|
| Triage and score one lead | Haiku 4.5 | Classification against fixed criteria; correctness is checkable | 0.15 US cents |
| Extract the fields from one inquiry email | Haiku 4.5 | Named fields into a form, confirmed by a person before saving | 0.25 US cents |
| Read one business card (capture app) | Haiku 4.5 | Mechanical, and verifiable at a glance | 0.5 US cents |
| Draft one email or sequence step | Sonnet 5 | Must write in Empire Teas’ voice to a real buyer | 1.5 US cents |
| Assist with one quote | Sonnet 5 | Reasons across readiness, certifications, and pricing; errors reach a buyer | 1.7 US cents |
| Brief one account | Sonnet 5 | Synthesis across an account’s full history | 2.1 US cents |
| Draft one product brief | Sonnet 5 | Must flag what is missing rather than invent it | 2.1 US cents |
| One dashboard morning briefing, per user per day | Sonnet 5 | Synthesis across the pipeline; first thing the team reads daily | 2.6 US cents |
| Research and rank one trade-show lead | Sonnet 5 | The reasoned “why this lead is worth your time” judgement | 2.0 US cents |
On Gemini, each figure is roughly one third of the above. The top tier (Anthropic Opus, Google Pro) is deliberately not used: at roughly five times the cost, nothing in this workload needs it.
To put the numbers in context, at Altrium’s assumed volumes — 15 users, 40 new leads, 87 inquiries, 880 drafts, and a daily briefing each — the total is approximately $25 per month. A trade-show lead costs about 4 US cents end to end across all three of its AI calls, so a 150-lead exhibition adds roughly $6. Every AI action in the system costs between a fifth of a cent and three cents; there is no usage pattern at this team size that produces a surprising bill.
Recommended budget: $50 per month, with a billing alert at $150. Controls: per-feature toggles an administrator can switch off without a deployment; the cheaper model on every high-volume path by design; prompt caching on repeated instruction text; per-call cost, token, latency and outcome logging so spend is attributable; and a configurable monthly ceiling that disables AI features rather than allowing an unbounded bill.
A.4 Security posture
Included in the baseline at the cost in A.2.2:
| Control | What it does here |
|---|---|
| AWS Secrets Manager | AI provider key, OAuth client secrets, and database credentials, with rotation; nothing sensitive in source control |
| Amazon GuardDuty | Continuous threat detection across account activity, network flow, and DNS |
| AWS CloudTrail | Immutable audit log of every AWS API call, to a central archive |
| AWS KMS | Encryption keys for database, storage, and backups |
| IAM Identity Center | Single sign-on with multi-factor authentication and time-limited roles; no long-lived keys for people |
| AWS Shield Standard | Baseline denial-of-service protection (automatic, no charge) |
| Account isolation | Development and production in separate AWS accounts, so development can never reach production data |
Standing controls: encryption in transit (TLS 1.2 or above) and at rest; database not reachable from the public internet; S3 public access blocked at account level; least-privilege security groups; and a build pipeline that fails on any dependency vulnerability at CVSS 7.0 or above, with exceptions requiring written justification and an expiry date.
Available as priced add-ons (see A.2.3): AWS WAF, Security Hub, Config, and Inspector. Altrium’s recommendation is to add WAF once the system carries real business, and to review the remainder at that point against Empire Teas’ own compliance requirements.
A security review before go-live covering authentication, session handling, permission enforcement, file upload, and the AI prompt boundary is included in the delivery process at no separate cost.
A.5 Backup and disaster recovery
| Scenario | Maximum data loss | Time to restore service |
|---|---|---|
| Accidental change or deletion by a user | 5 minutes | 1 to 2 hours |
| Database instance or availability-zone failure | 5 minutes | 1 to 2 hours (restore from backup) |
| Loss of the entire AWS region | Up to 24 hours | 4 to 8 hours, if cross-region copy is enabled |
These targets follow directly from the single-availability-zone baseline, and Empire Teas should confirm they are acceptable. Adding the multi-availability-zone database (+$30 per month, A.2.3) reduces the second row to under five minutes, automatic and is the single most cost-effective availability improvement available.
Design. Automated backups with point-in-time recovery at five-minute granularity, retained 14 days in production and 3 days in development. Daily snapshots. S3 versioning on file attachments, so an overwritten or deleted attachment is recoverable. Infrastructure defined as code, so an environment is rebuilt from the repository rather than from memory, which is what makes the regional recovery time achievable at all. A documented restore rehearsal before go-live and a restore test each quarter thereafter — a backup that has never been restored is not a backup.
A.6 AI data handling, retention and security
A.6.1 What is sent, and what is never sent
Only the minimum a task needs leaves the system, and only when a person triggers that task: the body of the one email being processed (extraction); the lead’s own fields (triage); a summary of one account’s activity (briefing); the relevant thread and account context (drafting); the quote’s line items and checklist state (quote assistance); aggregate counts and the user’s own open items (dashboard).
Never sent: credentials or API keys; bulk database exports; payment, banking, or letter-of-credit detail; any record outside the specific task. No blanket upload of the database ever occurs, and nothing is sent automatically in the background in V1 — every call is initiated by a person clicking an AI action.
A.6.2 Security of the AI path
Calls are server-to-server from the private network over TLS 1.2 or above, so the browser never talks to the AI service and the API key is never exposed to a user’s device. The key lives in Secrets Manager, is injected at runtime, is never in source control, and is rotatable without redeployment. Every call is logged with model, feature, user, token counts, latency, and outcome, and failures are alerted on. Every AI output is tagged with its source and marked as edited once a person changes it, so the system can always show whether a saved field came from a person or an approved AI draft. The audit trail records who approved each AI-generated action.
A.6.3 Retention
| Data | Retention |
|---|---|
| CRM records and file attachments | Life of the system; deletion on request under the data-protection process |
| AI call metadata (model, feature, user, tokens, latency, outcome) | Life of the system — this is what makes AI spend and behaviour auditable |
| AI prompt and response bodies | Not persisted in production logs; available in development only, for debugging |
| Application and infrastructure logs | 30 days, then archived or expired per the agreed policy |
| Database backups | 14 days production, 3 days development (A.5) |
A.6.4 The AI provider’s own terms
Under Anthropic’s commercial API terms, inputs and outputs submitted through the API are not used to train their models, and a zero-data-retention configuration is available to eligible organisations. Google offers comparable commercial terms for paid Gemini API and Vertex AI use. Altrium will obtain the current written terms and the data-retention position from whichever provider is selected and present them to Empire Teas as part of the AI data-handling document for formal approval before the AI features are built and before any real buyer data is processed. Empire Teas should treat this paragraph as Altrium’s understanding to be confirmed in writing at contract stage, not as a commitment from the provider.
A.6.5 Keeping the AI inside Empire Teas’ own cloud account
If the requirement is that buyer content must not reach a separate AI vendor’s service, that is achievable without changing the architecture, the user experience, or materially changing the cost. Claude on Amazon Bedrock serves the same models inside Empire Teas’ own AWS account and region, under the existing AWS agreement: no new vendor, no new region, no new agreement, and no new place for data to rest. It is Option 2 in A.3.1, at approximately $25 to $35 per month. Where data residency or vendor exposure is a requirement, this is Altrium’s recommended answer.
Consistent with Section 13, V1 assumes a hosted model on reputable cloud infrastructure and does not include an on-premises or self-hosted model. If Empire Teas’ review concludes that no hosted AI service is acceptable in any form, that is a change of assumption handled through Section 16.
A.7 Third-party tools and services
| Tool | Purpose | Supplied by | Cost |
|---|---|---|---|
| AI provider (Anthropic Claude, or the A.3.1 alternative) | The AI layer | Altrium sets up; on the Bedrock option no separate AI vendor account is needed | Consumption, per A.3 |
| Microsoft 365 | Mail, calendar, and contact synchronisation via Microsoft Graph | Empire Teas (already held) | No additional cost |
| GitHub with GitHub Actions | Source control and the build, test, scan, and deploy pipeline | Altrium | $0 to $50 per month |
| Domain and DNS | Application address | Empire Teas | Registrar fee; AWS certificates are free |
| OWASP Dependency-Check | Dependency vulnerability scanning in the build | Altrium | Open source |
Deliberately not required: no third-party automation platform, no integration marketplace, no separate error-tracking or analytics vendor. Every avoided vendor is an avoided subscription, an avoided contract, and one fewer place buyer data is copied to.
If the trade-show capture application is confirmed in scope, it runs on the same AWS platform: same account, same region, same security services, same database, same backup and recovery design. It introduces no additional hosting vendor. It carries the WhatsApp Business Platform for lead notifications ($5 to $25 per month at expected volumes) and Amazon SES for email (a few dollars), plus an enrichment data source still to be selected, and adds $25 to $40 per month of AWS infrastructure by reusing what A.2 already sizes.
A.8 Anticipated monthly operating cost
| Line | Indicative monthly cost |
|---|---|
| AWS infrastructure, both environments | approximately $140 |
| AI provider usage | approximately $25 at the volumes in A.3.2; budget $50 |
| GitHub | $0 to $50 |
| Anticipated total | approximately $165 to $215 per month |
Excluded: Empire Teas’ existing Microsoft 365 licences; the domain registrar fee; the capture application’s incremental costs if confirmed (approximately $30 to $65 per month, per A.7); any hardening options selected from A.2.3; and any change arising from the data-governance review in A.6.5.
Altrium’s recommendation is to run the baseline from go-live and add the multi-availability-zone database and AWS WAF (approximately $45 per month combined, per A.2.3) once the system is carrying real business, bringing the steady-state figure to approximately $210 to $260 per month. Against this, the annual Pipedrive licence is removed once Pipedrive is decommissioned.
A.9 Annual maintenance and support
An annual maintenance contract is recommended, beginning when the warranty period ends. The fee is in Section 17; this defines what it covers.
Included: corrective maintenance (diagnosis and fix of defects); security maintenance (dependency and container vulnerability monitoring, patching within an agreed window by severity); platform maintenance (runtime, framework, and AWS service upgrades); operational support (monitoring and alert response, backup and restore verification, the quarterly restore test in A.5); a defined allowance of [N] hours per month for minor enhancements, poolable within a quarter; and response and resolution targets by severity with a named contact.
Not included, quoted separately as change requests under Section 16: new features and integrations; the automated ERP integration follow-on; the deferred items in Section 4; migration to a new ERP; and the AWS and AI consumption charges in A.8, which are billed by those providers to the account holder.
Commercial basis. Annual maintenance for a system of this size is conventionally set at 15% to 20% of the build price per year, payable annually or quarterly in advance, commencing at the end of the warranty period. The figure and the enhancement-hour allowance are in Section 17.
Prepared by Altrium for Empire Teas (Pvt) Ltd. This is a draft for review.