Sample Scope, Production Tier: A 700-Hour Project in Five Phases
This is a representative example of a larger engagement: from assessment to production operations to rollout across additional processes. Names and figures are illustrative; the structure, the acceptance criteria, and the hour estimate per phase are exactly what you get. The smaller counterpart (a single pilot, 144 hours) is the pilot sample scope.
700 hours in total · 4 to 5 months · 5 phases with acceptance criteria, each stoppable on its own · billed for the hours worked · entry and architecture proof: phase 1 in 60 hours · operations package from 8 hours per month
1. Starting point
A financial services company with 350 employees processes several hundred customer documents daily (applications, statements, correspondence). Three teams type contents into two core systems by hand. Last year’s AI pilot never left testing: no GDPR sign-off for production, because customer data would have gone to a US API.
2. Goal
Document triage and data extraction in production, GDPR-compliant on the company’s own infrastructure, then rollout to two more processes. The internal IT team operates the system after handover; an operations package covers response times.
3. Phases and hours
| Phase | Result | Acceptance criterion | Duration | Estimated hours |
|---|---|---|---|---|
| Phase 1: Audit & target architecture | AI inventory, data-flow analysis, target architecture, GDPR classification per processing step | Architecture review passed with IT and data protection | 2 weeks | 60 h |
| Phase 2: Core system production-ready | Triage + extraction on own infrastructure, approval UI, PII protection layer | Extraction quality above agreed threshold on 3 months of historical documents | 5 weeks | 200 h |
| Phase 3: Integration & hardening | Both core systems connected, monitoring, alerting, load test, audit log | Error rate and latency below agreed limits under load | 4 weeks | 160 h |
| Phase 4: Rollout, processes 2 and 3 | Transfer to two more document processes incl. team onboarding | Both processes live, legacy effort measurably reduced | 5 weeks | 180 h |
| Phase 5: Operations setup & enablement | Runbook, internal team training, handover, 30 days of assisted operations | Internal team resolves standard incidents without external help | 3 weeks | 100 h |
Total: 700 estimated hours · roughly 4 to 5 months. Billed for the hours worked, at the hourly rate agreed up front, phase by phase. Every phase is accepted on its own. You can stop after any phase and keep everything delivered up to that point.
4. Operations package after phase 5 (optional, monthly cancellable)
| Tier | Service | Response time | Hours per month |
|---|---|---|---|
| Base | Monitoring, security updates, incident fixes | 2 business days | 8 h |
| Standard | Base + 1 development day per month | 1 business day | 16 h |
| Plus | Base + 2 development days per month | 4 hours (business days) | 24 h |
5. What I need from you
- Access to 3 months of historical documents (anonymized or masked is fine for phases 1 and 2).
- One contact each from business, IT, and data protection for a weekly 30-minute sync.
- Infrastructure access (own hardware or EU cloud) from phase 2; sizing is a phase 1 deliverable.
6. Assumptions and exclusions
- The core systems offer usable APIs (verified in phase 1; otherwise the phase 3 estimate is revised before it starts).
- Processing exclusively on your infrastructure or in an EU cloud; no customer data to third-party APIs outside the EU.
- Business approvals (e.g. payouts) stay human, deliberately out of scope.
7. Decision
This document is written so you can forward it directly to whoever signs. After phase 1 you know whether the architecture holds, after 60 hours, not after 700. The answer you owe me is one of three words: yes, no, or “call”, and all three are fine.
The “call” is right here, no form required:
Your scope follows the same structure: hour estimates per phase, acceptance criteria, each phase stoppable on its own, within 24 hours.
Hour estimates per phase, or a clear no with reasons.
Your agents answer from whatever the retriever finds, and too often that is last quarter's truth. I build the context layer they answer and act from: a temporal knowledge graph that keeps every fact with its source and the time it held, reads with each person's own permissions, and writes nothing without a person's approval. On your own tenant, billed by the hour, step by step.