A compact assessment of value, feasibility, process change, cost drivers and risk. This is the consulting artifact behind the technical case study, not a claim that a personal deployment is already an enterprise programme.
Scores are an engineering hypothesis on a 1–5 scale, not user research. Risk runs in the opposite direction: five is highest. A client discovery workshop must challenge every score.
| Use case | Impact | Feasibility | Data readiness | Risk | Decision | Why |
|---|---|---|---|---|---|---|
| Deterministic site discovery | 5 | 5 | 4 | 2 | ship | Public geodata and explicit filters; every result is reproducible. |
| Tender constraint extraction | 4 | 3 | 4 | 3 | pilot + review | Document interpretation saves time, but low-confidence constraints need human confirmation. |
| Cited research assistant | 3 | 4 | 4 | 3 | pilot + refusal | Useful for orientation when evidence is visible; never the legal or financial record. |
| Autonomous investment or eligibility decision | 4 | 2 | 2 | 5 | reject | Missing private, contractual and site-survey facts make autonomy unsafe and economically misleading. |
The target process removes search and transcription work. It does not remove the professional decisions that require non-public evidence.
No payback claim is published because no representative user study exists. A pilot earns continuation through measured workflow outcomes.
Users complete the target workflow faster without a material rise in missed candidates, unsupported answers or review burden, and the measured value exceeds full operating and ownership cost.
Public-data gaps dominate the workflow, human verification erases the time saved, evidence coverage stays unreliable, or users need a different decision rather than a better research path.
The personal product is evidence that the mechanics work. A regulated client deployment needs a separate acceptance boundary.
| Concern | Public product today | Pilot acceptance | Enterprise trigger |
|---|---|---|---|
| Identity and data | Anonymous, public data, restricted CORS. | Entra groups, managed identity, no client secrets in prompts or traces. | Private endpoints, retrieval ACLs, tenant isolation and retention policy. |
| Quality | Deterministic gates plus pre-release model evals. | Labelled client cases, failure taxonomy and human-review sampling. | Independent validation, drift monitoring and model-risk approval. |
| Reliability | Single region, health smoke, source freshness. | Named SLOs, rollback exercise, backup/restore test and incident owner. | Contractual RTO/RPO, regional recovery and on-call rotation. |
| Delivery | PR gates and immutable main deployment; no permanent UAT stack. | Isolated integration data, candidate release and business acceptance. | Segregation of duties, approved promotion and audited change record. |
| Governance | Visible AI disclosure, citations and refusal. | DPIA screen, use-case owner, model/data cards and user training. | Formal AI inventory, Purview-class lineage and periodic control review. |
Each phase produces evidence for the next investment decision. Dates are illustrative until stakeholders and data access are known.
Map the current process, users, decisions, data rights, NFRs and baseline effort. Re-score the use cases.
Use representative cases, label outcomes, validate source coverage and exercise refusal and security threats.
Integrate one workflow, instrument cost and quality, train reviewers and run with a small accountable cohort.
Compare against baseline, close material controls, estimate full TCO and choose scale, narrow or stop.
The technical case study shows what runs and why. This page shows how I would frame the same system as a client mission: value first, bounded scope, testable controls, measured economics and a reversible roadmap.