A retainer is supposed to buy ongoing work. In the AI services market right now, a worrying number of them are buying nothing — and the contract is written so nobody has to admit it.
The pattern
A retainer starts honestly. Real work ships in month one and two — a model gets tuned, a workflow wired in, a dashboard shipped. Around month four something quiet happens: the deliverable is in production, mostly running itself, and the monthly hours start being filled with "monitoring," "optimization," and "strategy syncs."
By month nine the inbox receives a tweak every couple of weeks. By month eighteen the client is paying for a relationship and an option to call, not for work. By month thirty the vendor cannot remember which deliverables are still live.
Nobody lied. The contract was signed in good faith. The treadmill is what happens when nobody is required to write down whether the fee is still buying anything (Class C — observable in the engagement's own logs, when those logs exist).
Why the market produces this
Retainers are rational for the vendor in a way clients underestimate. Predictable revenue gets a vendor through quiet quarters. Auto-renewal is the path of least resistance for both sides. And the AI services category is new enough that "ongoing optimization" sounds plausible even when the system has not been touched in six weeks.
The hiring-side analysis pushes the other way: employers buying AI work in 2026 are increasingly paying for shipped, owned artifacts rather than ongoing operator relationships (Class E — drawn from public interviews with operators of AI talent marketplaces). The retainer model assumes the opposite. The gap is where the treadmill lives.
The signals
A retainer is sliding toward annuity status when:
- The monthly status report stops listing artifacts. Reports go qualitative — "continued tuning," "ongoing alignment," "strategic check-ins" — instead of named, versioned outputs.
- The team you signed with is no longer on the engagement. A junior took over six months ago and you did not negotiate that handoff.
- You cannot answer "what would break if we paused this for a quarter?" If the honest answer is "nothing visible for ninety days," the retainer is buying an option, not current work.
- The renewal conversation is shorter every year. Auto-renew is the tell.
None of these are accusations. They are diagnostics. A retainer can pass through all four and still be honest — if the vendor will say so when asked.
The clauses we write into our own contracts
We do not run open-ended retainers. When the work is genuinely ongoing — a system that needs monitored evaluations, periodic re-tuning against drift, or scheduled audits — we write the engagement with clauses that make the treadmill structurally impossible (Class C — these are in our standard template and reproduced in every signed scope).
A defined deliverable per month. Every billed month names the artifacts produced — a re-trained version, an evaluation run, a documented incident response, a runbook update. "Hours available" is not a deliverable. If a month has no artifact, it does not bill.
A written exit condition. The contract names the conditions under which the engagement should end: the system reaches a stable state for N consecutive months, the client team is trained to maintain it, the underlying model is retired. When the condition fires, we are required to raise it — in writing, in the shared ledger — within the same week.
A cap on consecutive renewals without re-scope. The engagement cannot auto-renew past a fixed window without a fresh scope conversation. The default is twelve months. The re-scope is not optional and the conversation is supposed to be uncomfortable — that is the point.
A handoff option at any time. The client can, at any month boundary, request a full handoff package — runbook, credentials list, dependency map, evaluation set, recent failure modes — and we deliver it within the next billing cycle without prejudice to the relationship. The handoff is a right, not a breakup.
A retainer-to-project conversion clause. If the work turns out to be episodic rather than continuous, the contract converts to per-project at the client's request. The retainer was a hypothesis about cadence; if the hypothesis is wrong, the contract follows the evidence.
What this is not
This is not a claim that retainers are wrong. Some work is genuinely ongoing and a retainer is the right shape. The narrower claim: a retainer without a named monthly deliverable, a written exit condition, and a renewal pause is structurally indistinguishable from an annuity, and the client deserves clauses that tell the two apart.
Questions to ask a vendor offering a retainer, in writing:
- What is the deliverable this month, named and versioned?
- What conditions would end this retainer, and who is required to raise them?
- If I pause for a quarter, what concretely fails?
- Show me the last twelve months of monthly deliverable lists.
If those four answers do not exist, the retainer is buying the option to be billed, not the work.
The longer counter-model lives in the anti-extraction engagement model. The contractual posture — partnership clauses, not vendor terms — is on partnership-vs-vendor-contract. If you want to walk an existing retainer through these tests, that conversation is the intake call.