Most AI services contracts in 2026 are designed to keep the client paying after the work is done. We do not write contracts like that, and this post is the long-form version of why.
What extraction looks like in this market
Four patterns show up over and over in the AI services market right now. We treat them as a checklist of things our engagements must not do (Class C — configuration of the engagement itself, observable in the signed scope and the artifact list).
Seat licensing as the deliverable. The vendor sells "AI transformation" and the artifact the client walks away with is a per-seat subscription to a tool the vendor resells. Cancel the seats and the work disappears. The client never owns anything.
Opaque pipelines. A model is fine-tuned or a workflow is built, and the client receives a UI but not the prompts, the training data list, the evaluation set, the failure modes, the cost-per-call, or the versioned code. When something breaks six months later, only the vendor can diagnose it. That is not a delivery; that is a hostage situation.
Vendor lock by data gravity. The work is done inside a platform whose export format is deliberately useless. The client cannot leave without a re-implementation project, which the vendor is happy to quote.
Retainer treadmills. A monthly fee continues forever for "ongoing optimization" that is, on inspection, an inbox that occasionally receives a tweak. The treadmill is the product.
Each of these is rational for the vendor and corrosive for the client. The fact that they are everywhere is not an accident — it is what the incentive structure of the market currently selects for.
What we do instead
Our engagement model is built around four counter-commitments. Each one is a clause you can read, not a slogan.
A shared ledger. Every claim, decision, dependency, and cost in the engagement is logged in a document the client owns from day one. Evidence classes (A = empirical-in-session, B = code/inspection, C = configuration/integration, E = expert citation, F = falsifier present, U = unverified) are tagged on non-trivial claims. The ledger is the canonical source of truth — not a sales deck, not a status email. If a thing is not in the ledger, it did not happen.
Exit-ready artifacts. At the end of every engagement the client receives a runnable artifact: code, prompts, eval set, deployment scripts, and a written runbook that describes how to operate it without us. We test this by having someone outside the engagement re-run the artifact from the runbook before sign-off. If they cannot, the artifact is not ready (Class F — the falsifier is concrete and we have failed it before).
No hidden dependencies. Every external service, API key, model version, and library the deliverable depends on is listed, with a monthly cost estimate and a substitution note ("if this provider disappears, the replacement is X, cost delta Y"). Clients have a right to know how their thing will die.
Partnership clauses. The contract includes explicit clauses we will not waive: the client owns all artifacts and outputs; we will not insert dependencies on tools we resell without disclosure and consent; and if the engagement is going badly we are required to say so in writing, with the falsifier, by the end of the week we noticed it. The clause that triggers the disclosure is in the contract because intent without a clause is not a commitment.
What this maps to on the hiring side
Jay Kumar Chimata, who runs a 22k-user AI talent platform, did an interview with Themesis covering what employers in 2026 are actually willing to pay for versus what candidates are studying: Meet Jay Kumar Chimata: JobFirst.ai and the Real AI Job Market (Class E — primary-source interview with the operator of a hiring marketplace).
Our reading, in our own words: hiring managers are paying for people who can ship and explain a working artifact end-to-end, and they are not paying premium for someone who can only operate a vendor's UI. That maps directly to how we scope an engagement. We do not staff seat-operators; we staff people who can hand the client a runbook and walk away.
Themesis also has a useful split of AI work into three career tiers — tradesperson, engineer, innovator — and what depth of study is actually load-bearing for each: The Biggest Problem for AI Workers Is "Focus" (Class E). We use that same tiering during our intake call, because the right engagement scope for a tradesperson team is not the same as the right scope for an innovator team, and selling the same package to both is one of the quieter forms of extraction.
What this does not promise
We are deliberately not promising that this model is cheaper in month one. Sometimes it is more expensive up front, because writing a runbook your client can actually run takes longer than handing them a UI and a login. The savings show up later, in two places: the client is not paying a retainer that has stopped delivering, and the client is not paying a re-implementation cost in year two when they want to leave.
We are also not promising that we always succeed. Engagements fail. Falsifiers fire. When that happens, the partnership clause requires us to write it down, in the ledger, by Friday. The point of the clause is that the bad week is the test of the model, not the good week.
The standing offer
If you are evaluating an AI services engagement — ours or anyone else's — these are the questions worth asking the vendor in writing:
- What artifact do I own at the end? Can someone outside the engagement re-run it from a runbook?
- What does this depend on that I am not being shown?
- What is the falsifier that ends the engagement, and who decides it has fired?
- Show me last quarter's ledger.
A vendor whose contract cannot answer those four questions is, by our definition, extractive. That is not name-calling — it is the test.
The plain-language version of the working hypothesis behind our practice lives on /science. The exact wording of what we do and do not claim is on /standard/what-we-do-not-claim. The partnership clauses described above are reproduced in the /partnership-not-vendor page, and the running ledger format is described on /transparency.
If you want to test any of this against your own situation, the intake call is on /workshop. It is the same call whether the engagement is the right fit or it is not — we would rather tell you in week one than bill you for a year of theater.