Synthetic sample dossier
What an AI Workflow Assessment hands you, in the shape it arrives in. Every figure, name and finding below is synthetic — created to show the structure and depth of the deliverable, not to present a client engagement or a measured result.
Read this as a template, not evidence. The real dossier is built from your workflow, your artefacts and your numbers; it states which properties were actually tested, which were not, and what the residual risks are.
Engagement summary
| Workflow | RFP response drafting and assembly (synthetic example) |
|---|---|
| Owner | Bid Desk (synthetic) |
| Duration | 10 business days |
| Assessment question | Is this workflow ready to run under review, and what does it cost per accepted proposal? |
1. Threat model and control analysis (excerpt)
Trust boundaries are mapped first: everything the workflow reads but does not control — inbound e-mail, retrieved documents, tool results. Each tool's blast radius is assessed as if the agent could be manipulated into using it. All findings synthetic.
| Threat | Severity | Control | Residual |
|---|---|---|---|
| T1 — Inbound e-mail contains text addressed to the agent, not the human (prompt injection via untrusted input) | High | Instruction-bearing content is stripped and quarantined before the model sees it; the model is never the parser | Injection via attachment content not yet covered by the filter — open finding |
| T2 — Pricing lookup tool returns a value outside the approved model when the catalogue is stale | High | Pricing facts are re-validated against the approved model at the review gate, not inside the prompt | Staleness window of one business day accepted by the workflow owner |
| T3 — Service account can send e-mail even though the workflow only drafts | Medium | Send permission removed from the drafting role; sending requires a separately-granted release authority | None once the second authority is exercised only by the release step |
| T4 — Retrieved bid-library documents carry claims with expired provenance | Medium | Provenance must be present and current or the citation is dropped at assembly | Historical bids without provenance metadata are excluded (coverage loss, documented) |
| T5 — Reviewer approves under deadline pressure without seeing the exception flags | Medium | Review gate shows exception flags before sign-off is possible; flags cannot be dismissed in bulk | Human vigilance remains the last control — monitored via sampled audits |
2. Evaluation plan and results (excerpt)
Properties the workflow must hold, the method used to test them, and the result on synthetic data. A test suite is evidence over the tested cases and properties — nothing more, and the dossier says exactly that.
No released proposal contains a price outside the approved model
Method: Property-based test generating adversarial drafts (prices in tables, footnotes, paraphrases)
Result: 3 paraphrase violations found by the harness on a synthetic corpus; held at the review gate (synthetic figure)
A submission happens at most once
Method: Retry with idempotency key replayed against a double-fault injection
Result: No duplicate submissions in 10,000 synthetic retries (synthetic figure)
Every exception reaches a named human within one business day
Method: Fault-injection across confidence thresholds; queue audit
Result: 2 of 40 synthetic exceptions routed to an unmonitored inbox — control gap, fixed before recommendation
3. Acceptance criteria and release gates (excerpt)
What "good enough to ship" means, in testable terms. The fifteenth failing test is the reason the recommendation below is "redesign first" — a demo does not clear a gate.
- 14 of 15 release tests passing (the fifteenth: bulk-flag dismissal edge case — blocker until fixed)
- Exception rate at or below 8% on a 4-week window (synthetic baseline)
- Cost per accepted proposal at or below ₹X against the baseline (synthetic figure)
- Drift check on: weekly re-run of the release suite against production samples
- Named gate owner: Bid Desk lead, with authority to halt release
4. Economic model (excerpt, synthetic figures)
Baseline vs agentic path, per unit of accepted work. Real dossiers carry the assumptions block: loaded cost rate, volumes, exception handling, model pricing date, integration cost — every one stated, none hidden.
| Measure | Baseline | Agentic path |
|---|---|---|
| Volume (proposals per quarter) | 240 | 240 (same demand assumed) |
| Hours per proposal (human) | 6.5 | 2.1 (review-weighted) |
| Exception rate | — | ~7% reach a human (synthetic run) |
| Model spend per accepted proposal | — | ₹X (synthetic figure) |
| Loaded cost per accepted proposal | ₹A | ₹B — a material reduction (synthetic figures) |
₹X, ₹A and ₹B are deliberately unpopulated: a synthetic document must not imply a real price or saving. Your dossier carries real figures with their assumptions.
5. Recommendation and decision criteria (excerpt)
- Redesign first, then pilot: the drafting quality is strong; the boundary, pricing-evidence and escalation controls are not yet release-grade.
- Pilot entry criteria: the three control gaps above closed and re-tested.
- Pilot exit criteria: two consecutive 4-week windows inside the release criteria, with drift checks passing.
- Decline trigger: if exception rate cannot be brought under the bar, the workflow should stay in drafting-assist mode with mandatory human assembly.
The current-state and target-state maps, the full abuse-case catalogue and the assumptions block are omitted from this excerpt for length — the delivered dossier includes them.
This is the shape of the decision you receive.
Your workflow, your artefacts, your numbers — with the limitations and residual risks stated in writing.