Constraint Extraction from MSAs & Compliance Addenda

Tier: Foundation · Hands-on time: ~25 minutes

In a services company, the architect is the client-constraint enforcement point: your signature on the solution determines which client data touches which systems — including the AI tools in your own workflow. The irony this playbook resolves: you need the constraint register before you can safely use AI on client material, so the extraction itself must run on a safe path. It is also the natural first playbook of any engagement — pb-01's requirements work and pb-02's estimation both check against the register this produces.

When to use

• •
  Engagement kickoff: an MSA, DPA, security addendum, or completed questionnaire needs to become working constraints the delivery team can actually follow.
• •
  A client adds a compliance addendum mid-engagement (DPDP, GDPR, sector rules) and you need the delta against current practice.
• •
  Someone asks "are we allowed to…?" and the honest answer is "let me check" — for the third time this month.

Skip AI when the contract is two pages — read it. The register format is still worth filling by hand.

Prerequisites

• The governing documentsMSA, data-processing agreement, security/compliance addenda, in text-extractable form.
• An approved processing path for legal documentsthis matters before the first paste: your firm's enterprise AI tenancy, a NotebookLM-class doc Q&A tool covered by your org's agreement, or a local model. A client's MSA is itself confidential material.
• A structured-output-capable LLM API (optional but recommended) — for a register your tooling can consume.
• A register templatecolumns: constraint ID, category, verbatim clause, source location, plain-English rule, affects-AI-usage (Y/N), owner.

The workflow

1. 1.

   Confirm the processing path for the contract itself. Decide where this extraction runs before pasting: enterprise tenancy, client-approved endpoint, or local model. If your firm has a standing agreement covering legal documents, use it; if in doubt, treat the MSA as the most restricted document in the engagement.

2. 2.

   Extract with verbatim anchoring as a hard rule.

   Copy

   You are extracting binding constraints from a Master Services Agreement
   and its security addendum for an engagement register. I am the principal
   architect for the Meridian Retail engagement.
   
   Rules:
   - Every constraint MUST carry a verbatim quote and its section number.
   - If a topic is not addressed, output "NOT ADDRESSED" — never infer a
     permission or a prohibition from silence.
   - Where two clauses conflict, list both quotes and flag CONFLICT.
   
   Extract into categories:
   1. Data handling — residency, classification, retention, deletion.
   2. Tooling & subprocessors — restrictions on third-party services,
      cloud regions, approval requirements for new tools.
   3. AI-specific clauses — any mention of machine learning, AI, model
      training, or automated processing of client data.
   4. Audit & evidence — client audit rights, logging duties, breach
      notification windows.
   5. Personnel & access — location restrictions, background checks,
     least-privilege duties.
   
   Output as a table matching: ID, category, verbatim quote, section,
   plain-English rule, affects-AI-usage (Y/N).
   ###
   [paste contract text]
   

3. 3.

   Verify every quote against the source. This is not a spot-check playbook — a constraint register with a fabricated clause is worse than no register, because people will comply with rules that do not exist and miss ones that do. Open the PDF, find each section number, confirm the quote. Twenty constraints take ten minutes; a wrong one in front of client legal takes a quarter.

4. 4.

   Interrogate the silences. Most pre-2024 MSAs never mention AI. "NOT ADDRESSED" on AI tooling is not permission — it is a question for the account team and client. Prompt: "List the questions the silence in category 3 creates, given we intend to use AI assistants for design documents and code." Route those to whoever owns the client relationship; record the answers as constraints with source "client email, [date]".

5. 5.

   Derive the AI-usage ruling. From the register, write the one paragraph every team member actually needs: what client material may touch which AI tools, what must be redacted first, what is prohibited outright. For Meridian this might read: "Client data and named-system architecture: approved tenancy only. Redacted/structural content: firm-standard tools. Anything with cardholder data: no AI processing." Default, when the register cannot support a ruling: no.

6. 6.

   File and wire it in. Register lives with the engagement charter; the AI-usage ruling goes into the team onboarding doc. Every subsequent playbook's restricted-environment variant starts from this ruling. Recheck on every contract amendment.

Verify before you trust it

• Quote-by-quote source check. Every clause verified against the actual document — full coverage, not sampling. Fabricated or subtly altered legal text is the failure mode with the longest blast radius in this track.
• Section numbers resolve. Models mangle references between drafts of the same contract. A right quote with a wrong section still fails an audit.
• Silences confirmed. For each NOT ADDRESSED, run your own keyword search (AI, model, subprocessor, residency) over the PDF. The model missing a clause and declaring silence is the inverse failure of fabrication — equally costly.
• Conflicts escalated, not resolved. If the model flagged a CONFLICT, the resolution comes from legal or the account team — never from the model, and not from you alone.

Restricted-environment variant

This playbook is the restricted-environment variant for everything else — but it still needs its own safe path. Preferred order: your firm's enterprise AI tenancy with no-training terms; a client-approved endpoint in the client's own subscription; local extraction with Ollama and a capable open model for firms with neither. If none exist, do the extraction manually with the same register template — the artifact matters more than the acceleration. Never run a client MSA through a free consumer chat tool: the document that defines your confidentiality duties should not be the first thing that breaches them.