← Back to audit

Synthetic sample

Controlled AI audit example

Method proof, not a client deliverable. Fictional company, workflow run as a real bounded diagnostic.

Rivage Ops (fictional)B2B services · ~18 people · Bordeaux

Chosen boundary

One priority workflow: inbound form + email qualification → quote draft on Slack, never client send without a human.

Deterministic layer

The model can improvise. The workflow should not.

Interpretation on one side, authority on the other. Consequential actions (publish, CRM write, send a price) go through rules and a named human.

  1. 01

    Input

    Document, message or event inside a bounded boundary.

  2. 02

    Model

    Extracts, classifies, drafts. No final authority.

  3. 03

    Rules

    Allowlist, validation, states, retries, logs, cost cap.

  4. 04

    Human

    A named owner approves, corrects or stops before sensitive action.

Failure path: stop, log, escalate, recover without hiding the exception.

The five artifacts

  1. 01

    Bounded workflow map

    Owner: sales lead. Inputs: site form + email. Output: quote draft + Slack notify. Auto: CRM read. Confirm: stage write. Forbidden: send quote, delete, invent prices outside rate cards.

  2. 02

    Measurable baseline

    Before: estimated median form → first draft 2-3 h (n=12, internal 2-week sample). Design target: median < 20 min to HITL draft, 0 auto-send.

  3. 03

    Allowlist matrix

    Auto: classify project type, extract missing fields, read CRM status. Confirm: create/update opportunity, post draft. Stop: off-card pricing, dispute, undocumentated deadline promise.

  4. 04

    Observable failure register

    Representative (8), edge (4), known-bad (3): empty attachment, duplicate request, prompt injection in body, missing rate card. Each case has expected stop and owner.

  5. 05

    Go / no-go + pilot spec

    Go recommended for a 4-week pilot, n≥30 standard requests, mandatory HITL. No-go without versioned rate cards or a named confirm owner. Optional next: Quote Pilot-style agent build.

This sample shows

  • · Bounded audit method on one workflow
  • · Model / rules / human separation
  • · Artifacts fit for leadership review
  • · Explicit go/no-go criteria

It does not show

  • · ROI or accuracy measured at a real client
  • · Legal, formal GDPR or AI Act sign-off
  • · Production deployment
  • · Access to confidential data

Bring one workflow. Leave with a decision.

Fixed package (Independent AI engineer): €1,500 excl. VAT, one workflow, five artifacts, no build obligation.

Book 20-40 minFull audit pageStructured intake

To archive: browser print → save as PDF.