Build Client-Approval Workflows for AI Use That Survive an Audit

Workflow area: Policy & Governance · Tier: Advanced · Hands-on time: ~45 minutes

The nightmare scenario is mundane: an engineer uses a perfectly good, internally approved assistant on client code, and six months later the client's audit asks for your AI controls on their engagement. If the answer is a shrug, you are renegotiating trust mid-contract. The fix is boring and powerful — decide per client, write it down, keep the evidence — and it turns AI governance from a sales liability into a differentiator, because most of your competitors cannot answer the audit either.

When to use — and when not to

• Use it if any code, data, or deliverable belongs to a clientservices firms, GCCs serving a parent company, agencies, and anyone with source code under NDA in their repos.
• Use it at deal time, not delivery timethe AI conversation belongs in the MSA/SOW negotiation, where it is a clause; mid-engagement it is a confession.
• Use it when clients start sending AI questionnairesthey already are; a pre-built evidence pack turns a two-week fire drill into a same-day reply.
• Skip the per-client apparatus for pure product workif the only data is your own, pb-27/28/34 cover you; this playbook's cost is only justified where client consent is the gating fact.
• Skip silent defaults in either directionneither "use AI everywhere until told otherwise" nor "ban it everywhere to be safe" survives contact with reality; the first fails audits, the second loses the productivity and drives usage underground (pb-32).

The recipe

1. 1.

   Build the approved-tool matrix. One table, one row per active engagement: client / engagement / AI tools permitted (specific tools and tiers, not "AI") / data classes allowed into those tools (client code? client data? never credentials) / contractual basis (clause reference, written consent, or none-yet) / review date. Three standard postures keep it manageable — Green (client consented in writing, listed tools allowed on client material), Amber (AI allowed on your own scaffolding and generic code, nothing client-confidential enters any tool), Red (no AI touching the engagement at all). Every engagement gets a posture today, even if today's honest answer is Amber-by-default pending the client conversation.

2. 2.

   Assemble the evidence pack once, reuse per client. A folder your account leads can send within a day: your AI policy and review standard (pb-27), CI enforcement summary (pb-28), agent governance (pb-29/30), incident classes and notification commitments (pb-31), plus vendor documentation pulled from trust centers — enterprise terms showing no training on your inputs, data residency, retention windows, certifications (SOC 2, ISO 27001). Refresh vendor docs quarterly; trust-center PDFs age fast and an outdated one in an audit reads worse than none.

3. 3.

   Draft the MSA/SOW AI addendum with counsel. Have Claude produce the negotiating skeleton, then hand it to legal — the model drafts structure, lawyers make it binding:

   Copy

   Draft a skeleton AI-use addendum for a software services MSA,
   for review by counsel — mark every clause [LEGAL REVIEW].
   Cover: (1) definition of AI-assisted development; (2) the
   approved-tool list as a referenced, updatable schedule, not
   inline; (3) data-handling commitments — what client material
   may enter which tool class, no-training and retention terms;
   (4) human-review warranty for AI-assisted deliverables;
   (5) incident notification for AI-related data exposure, with
   timeframe placeholder; (6) audit rights scoped to AI controls;
   (7) change process when we want to add a tool.
   Plain language, no invented legal citations.
   

   The load-bearing choice is (2): the tool list lives in a schedule you can update by notice-and-consent, so adopting a new tool does not mean reopening the MSA.

4. 4.

   Wire the matrix into delivery. A posture that lives in a spreadsheet governs nothing. Green/Amber/Red flows into onboarding checklists per engagement, repo-level configuration (Red engagements get assistant integrations disabled and pb-28 gates set to block on AI provenance labels), and agent scoping (client-scoped credentials per pb-30; agents cap at the rung the addendum permits, per pb-29). The test: a new engineer joining an engagement learns its AI posture in their first hour without asking.

5. 5.

   Run the client conversation proactively on your top accounts. For each Amber account, take a one-pager to the client: what you want to use, what never leaves your boundary, what they get (speed, and the evidence pack). Many clients say yes to a specific, evidenced request who would say no to a vague one — and the ones who say no have told you something you needed on the matrix anyway.

Make it yours

• Small consultancy: the matrix is one sheet, the evidence pack one folder, and the addendum one template your lawyer blesses once — the discipline is refusing to start work before the posture is set.
• Large services firm or GCC: the matrix becomes a system of record owned jointly by delivery and legal, with the parent company (for GCCs) treated as your most demanding client — their group policy is your floor; per-engagement audits get a named responder and the evidence pack under version control.
• Client mix with regulated industries: bank and healthcare clients bring their own regulator's AI expectations — map each regulated client's questionnaire once, keep the answers in the pack, and expect Red postures to be permanent on some accounts; price engagements accordingly rather than eroding the posture quietly.

Quality check

• Every active engagement has a posturean engagement missing from the matrix is Amber in your head and Green in practice, which is how audits are failed.
• The matrix matches the toolingspot-check a Red engagement's repos for assistant activity and provenance labels; a Red posture with Green telemetry is your most urgent finding.
• Evidence-pack response time is under 48 hourstime-trial it with a fake client questionnaire; slow evidence reads as absent evidence.
• The addendum has actually been signed somewherea template no client has executed is a hypothesis; get it into the next two renewals and learn what gets redlined.
• Postures get reviewed at renewalclients loosen and tighten; a matrix last touched at signing is stale by year two.