Most engagement contracts have thirty clauses and three decide whether you are entering a partnership or a dependency. The rest is decoration.
This post names those three and gives you the redline text. Hand it to counsel. Demand it from anyone trying to sell you AI services.
Why three
A contract is a load test of intent. Warm sentences about "collaboration" won't hold weight when the relationship is tested. The clauses that hold weight are the ones that change what the supplier physically does every week.
There are three (Class C — from our own live engagements). Each shifts a specific power dynamic. Together they make the supplier structurally unable to lock the client in.
Clause 1 — Shared Ledger Access from Day One
The biggest hidden cost in an AI engagement is the supplier's private log of what actually happened. Status decks are summaries written for the buyer; the operational truth lives in someone else's database. When the relationship ends, that database goes with them.
Redline language:
Client shall have read-only credentialed access to the engagement's full operational ledger — work performed, model calls, prompts, evaluations, deployments, errors, and costs — from the first business day onward, current to the minute, with no roll-up and no redaction. Supplier shall not maintain any internal record of engagement activity that is not also visible to Client through this ledger.
Two observable tests:
- A client staff member can pull yesterday's ledger at 7 a.m. without asking permission.
- There is no second ledger. If a supplier engineer would ever consult records the client cannot see, the clause is being violated.
A supplier that resists this clause is telling you what the relationship is. Our own ledger posture is on /anti-extraction-engagement-model.
Clause 2 — Exit-Ready Artifacts, Delivered Weekly
The standard pattern: artifacts — code, prompts, configs, dashboards, runbooks — live in the supplier's systems until the contract ends, at which point the client gets a zip file and a phone number that stops being answered. The partnership pattern inverts this.
Redline language:
At the end of each calendar week, Supplier shall deliver to a Client-controlled storage location an "Exit Pack" containing: (a) every artifact produced or modified that week, including source code, prompts, configuration, datasets, model weights or pointers, dashboards, and runbooks; (b) a signed SHA-256 manifest; (c) a plain-English summary of what is now in production and what would have to be rebuilt if the engagement ended on Friday. Client may terminate for convenience at any week boundary and shall owe no further fee.
The word doing the work is weekly. A monthly or quarterly exit pack is a vendor pretending to be a partner. Weekly means the off-ramp is never more than five business days old. Termination-for-convenience is what makes the exit pack real rather than ceremonial — without it, the artifacts are theatre.
This clause is cheap to honor if your build is healthy and expensive if it is not. That asymmetry is the point.
Clause 3 — Falsifier Published With Every Load-Bearing Claim
Every proposal contains performance claims. "Cuts resolution time by 40 percent." "Halves the error rate." "Handles 90 percent of inbound." These claims do most of the selling and almost none of the proving. The partnership move is to require, in writing, before work begins, the specific observation that would falsify each one.
Redline language:
For every quantitative or qualitative performance claim Supplier makes about the delivered system — in proposals, status reports, dashboards, or marketing — Supplier shall publish, in the shared ledger and on a Client-accessible artifact page, the corresponding falsifier: a specific, measurable observation that would, if recorded, disprove the claim. Claims without a published falsifier may not be invoked in pricing, scope, or renewal discussions.
This kills three failure modes in one move: vanity metrics (no falsifier means no claim), retroactive goalpost-shifting (the falsifier is timestamped before the result), and the soft handwave that fills industry case studies and proves nothing.
Stating in advance what would disprove a claim is older than any of this technology (Class E — the falsifiability criterion is standard in scientific methodology and survives because nothing else reliably separates a real claim from a marketing one).
What these clauses cost a real partner
Nothing. If you already run a shared ledger, weekly exit packs are a script. If you already ship exit packs, falsifiers are a paragraph at the bottom.
What they cost an extractive supplier is everything. The whole business model of opaque AI services rests on the client not seeing the ledger, not being able to leave on a week's notice, and not being able to check the claims. Take any of the three away and the model breaks.
That is the test. Sign all three on Monday and deliver the first Exit Pack that Friday — partner. Spend three weeks negotiating which "doesn't apply in this case" — vendor.
Where to go next
- The longer five-clause version is /partnership-vs-vendor-contract, which adds dependency-inventory and evidence-class disclosure on top of these three.
- The standing engagement posture is /anti-extraction-engagement-model.
- To walk through these clauses against contracts you have right now — your current vendor, a proposal on your desk, the renewal in your inbox — that is what the /workshop session is for. Bring the documents. We will read them together.
— Michael