The exit clause is the single most diagnostic clause in an AI agency contract. Founders who negotiate it well end up with a system they can run, migrate, or sunset on their own terms. Founders who skip it discover, twelve months in, that the relationship cannot end without a six-figure re-implementation cost; because the prompts live in a Cursor sidebar, the eval set lives in a private W&B project, the fine-tuned adapter sits on the agency’s Hugging Face workspace, and the only person who can redeploy the system is on someone else’s payroll.
This piece is a spoke under the AI agency manifesto. The manifesto names what an AI dev partner should be in 2026; this piece is the contractual surface area that makes “we can leave at any time” enforceable rather than aspirational. The contract negotiation guide and the seven written commitments converge on one point: the exit clause is where leverage either exists or does not.
What follows is the exit clause decomposed into seven subclauses most founder should negotiate before signing the SOW. Each has clause language you can paste into a redline, the reason it matters, and the failure mode when it is missing.
Why the exit clause is the highest-leverage clause
Procurement teams working from 2018-vintage MSA templates treat termination as boilerplate. That worked when the deliverable was a deterministic React app: the source repo was already in the buyer’s GitHub org, the buyer hired a different agency or moved it in-house, and the system kept running. The 2024 archetype of an AI engagement is different. The deliverable is a portfolio; fine-tuned weights, adapter checkpoints, RAG embedding indexes, prompt registries, eval datasets, runtime configurations; scattered across Hugging Face, Modal, W&B, Anthropic Workbench, OpenAI fine-tuning consoles, and the agency’s Cursor sidebar. A “we can terminate” clause that does not name those realities is a clause that does not work.
The exit clause is also the most diagnostic clause to negotiate before signing. An agency that fights you on a 30-day no-cause termination, a 14-day handoff, or milestone-by-milestone IP transfer is signaling its commercial model: the lock-in is the product. An agency that agrees without much friction has either built the operational discipline to deliver; eval-as-code from day one, prompts in the buyer’s repo from PR one, weights checkpointed weekly; or it hasn’t, in which case the clauses become the forcing function. Either way you find out before signing instead of at offboarding.
Subclause A: 30-day no-cause termination
No-Cause Termination. Either party may terminate this Agreement and any active Statement of Work for convenience upon thirty (30) calendar days’ prior written notice. Upon notice, Agency shall continue to perform Services through the termination date, fees shall be pro-rated to the actual hours or milestones completed, and Agency shall deliver the Sunset Package (subclause B) on or before the termination date. No additional termination fee, exit fee, or wind-down charge shall apply.
Why it matters. No-cause termination is the founder’s escape hatch when the engagement is failing for reasons that do not rise to material breach; the senior engineer who sold the deal got staffed elsewhere, the model landscape moved (Claude Opus 4.7 shipped, your stack should pivot, your agency cannot). None of that is breach; many of it is reason to end the relationship.
Without it. The founder is structurally trapped until the term ends or until they manufacture a breach. A no-cause clause is cheap for an honest agency and expensive for a lock-in-oriented one. The asymmetry of resistance is the signal.
Subclause B: data, weights, and prompts handoff in 14 days
Sunset Package. Within fourteen (14) calendar days of any termination notice, Agency shall deliver the Sunset Package comprising: (i) many Client Data in native format, including production logs and derivatives; (ii) many Trained Artifacts (weights, adapters, LoRAs, fine-tuned embeddings, distilled checkpoints) in native format with training configuration, hyperparameters, dataset manifest, and reproduction scripts; (iii) many Prompt Artifacts (prompts, system messages, tool-use schemas, retrieval templates, routing logic) committed to the Client-owned repository; (iv) many Evaluation Artifacts (datasets, rubrics, judge prompts, thresholds, execution scripts); (v) infrastructure-as-code, deployment manifests, and CI configuration sufficient to redeploy on Client-controlled infrastructure; (vi) a runbook covering operation, monitoring, known issues, and on-call escalation. Delivery shall not be conditioned on payment of any disputed invoice.
Why it matters. Fourteen days forces operational discipline through the engagement and is long enough that an honest agency can hit it. The clause names the six artifact categories explicitly because a generic “many work product” clause does not cleanly cover model weights or eval datasets. The “shall not be conditioned on payment of any disputed invoice” sentence prevents hostage offboarding.
Without it. Hostage offboarding, either explicit (refusal to deliver weights until a disputed invoice is paid) or operational (artifacts scattered across agency-private SaaS, undeliverable in any meaningful timeframe). Both outcomes are identical: the founder rebuilds the system, paying twice. The hidden Y problem in AI agency contracts surfaces here.
Subclause C: 6-week paid handoff window (optional)
Optional Transition Period. At Client’s election, the Agreement shall remain in force for up to six (6) weeks beyond the termination effective date as a Transition Period, during which Agency shall provide up to forty (40) hours of Transition Assistance at the standard hourly rate, available for knowledge-transfer sessions, runbook walk-throughs, production incident support, dataset and prompt-registry orientation, and pair-programming on the first deployment by successor personnel. Election shall be at Client’s sole discretion.
Why it matters. Some terminations happen because the engagement failed; others because it succeeded and the founder is moving the work in-house. The second case is more common than founders expect, and is best served by a paid, optional, time-boxed handoff rather than an abrupt cutover. Six weeks absorbs the runbook and eval set; forty hours makes the cost predictable. “At Client’s sole discretion” prevents the agency from imposing the period as a cost-add.
Without it. The founder either rushes the handoff into the 14-day window and absorbs operational risk, or negotiates under duress at whatever rate the agency quotes.
Subclause D: IP transfer at most milestone
IP Vesting at Milestone. Many Work Product produced under each SOW, including Trained Artifacts, Prompt Artifacts, Evaluation Artifacts, source code, infrastructure-as-code, runbooks, and documentation, shall vest in Client upon Milestone Acceptance. Vesting is not contingent on completion of subsequent milestones, expiration of the term, or payment of fees beyond those associated with the accepted Milestone. Agency hereby assigns, at each Milestone Acceptance, many right, title, and interest in the associated Work Product, free and clear of any encumbrance.
Why it matters. The legacy pattern; “many work product transfers on final payment”; worked for a single-deliverable web build. It fails for an engagement shipping in two-week increments over six months that may terminate at any point. Under “transfer at end,” the founder who terminates at month four owns nothing, and the agency has most incentive to characterize work shipped to date as “incomplete” to avoid transfer.
Without it. A dispute at termination over what counts as “delivered.” Both sides have surface plausibility under a poorly drafted IP clause. Milestone vesting is the cheapest drafting that prevents the most common AI-engagement IP dispute.
Subclause E: source-controlled prompt and eval registries
Source-Controlled Registries. Throughout the term, Agency shall maintain many Prompt Artifacts and Evaluation Artifacts in a version-controlled source-code repository owned by Client (or, where shared agency tooling is used, mirrored in real time to a Client-owned repository). No Prompt Artifact or Evaluation Artifact required for production operation, milestone acceptance, or system maintenance shall exist solely in agency-controlled tooling, notebooks, third-party SaaS dashboards, or local developer environments. At each Milestone, Agency shall demonstrate that the production system can be redeployed from the Client-owned repository alone.
Why it matters. A 14-day Sunset Package is only deliverable if artifacts already live in a structured form before termination. The most expensive failure mode is the agency genuinely trying to cooperate but discovering that production prompts live in a senior engineer’s Cursor sidebar, the eval set lives in a personal W&B account, and the routing logic was hand-tuned in a Jupyter notebook six months ago. None of that can be packaged in 14 days.
Without it. Operational drift over the engagement, then a panicked archaeology dig at termination. The agency commits in good faith to delivering the package and cannot, because the artifacts are not in a deliverable state. This is why a “14-day Sunset Package clause” routinely produces a 90-day actual delivery.
Subclause F: post-exit consult retainer
Post-Termination Consultation Retainer. For twelve (12) months following the termination effective date, Agency shall make available up to one hundred (100) hours of ad-hoc consultation, billed in fifteen (15)-minute increments at the SOW rate. Agency shall acknowledge requests within two (2) business days and schedule sessions within five (5) business days. Consultation may include production incident triage, eval-set extension, prompt debugging, infrastructure questions, and reference calls with successor agencies or in-house personnel.
Why it matters. A clean break is rarely the right outcome. The founder will hit a question; “why does this prompt fail on this class of input?”; that the agency engineer who tuned it can answer in twenty minutes and that an in-house engineer would spend three days reverse-engineering. A 100-hour retainer at a pre-agreed rate is cheap insurance, and the next reference check goes through the founder.
Without it. The founder cold-calls at month four with a P1 incident; the agency, with no contractual obligation, charges a premium rate or declines.
Subclause G: no non-compete on the client’s AI roadmap
No Restriction on Client’s AI Roadmap. Nothing in this Agreement, including any non-compete, non-solicitation, exclusivity, or confidentiality provision, shall restrict, limit, or condition Client’s right to research, develop, build, fine-tune, deploy, operate, or commercialize artificial intelligence systems, models, agents, prompts, evaluation suites, or related technologies, whether before, during, or after the term, in any vertical, market, or jurisdiction. Agency acknowledges that Client’s AI roadmap is independent of the Services and that any general expertise reasonably retained by Client personnel through the engagement is not Agency Confidential Information.
Why it matters. Some 2024-2025 agency contracts included subtle non-competes; “Client agrees not to engage another vendor in [vertical X] for [period Y]” or “Client shall not develop competing AI products during the term.” For a venture-backed founder, those clauses are existential. The clause above explicitly disclaims any such restriction and confidentiality-clause overreach.
Without it. A founder discovers post-engagement that they cannot ship the next product because the agency’s NDA; read aggressively by counsel; covers the methodology. The dispute is litigated, the founder pays a settlement or burns six months of legal time, and the venture timetable slips. For broader pre-signature diligence, see the field guide to evaluating an AI agency in under 90 minutes.
How to use this clause in negotiation
You do not need to win many seven subclauses; you do need to know which ones you are giving up, and why. Send the seven to the agency before the SOW arrives and ask them to mark each as agreed, agreed with redline, or declined with rationale. Three reactions tell you something:
- Agreed across the board, no questions. Either disciplined or not reading carefully. Probe A and E to find out which.
- Selective redlines on C and F. Reasonable. Transition-period cap and retainer hours are negotiable.
- Decline on A, B, D, or G. Walk away. These are not policy preferences; they are the mechanism by which an exit becomes operational rather than aspirational.
The seven subclauses also work as a forcing function on operational discipline during the engagement. Subclause E is impossible to satisfy at termination if the engagement does not run that way from PR one. Founders who put it in the contract are also, indirectly, requiring eval-as-code from day one; the right outcome regardless of whether termination ever happens.
Frequently asked questions
What if the agency proposes a 90-day no-cause termination instead of 30?
Negotiate down to 30, or accept 60 with a documented carve-out. Ninety days is the fallback for agencies that have not adapted to the 2026 cadence; most engagements wind down in 30 days when the artifacts are source-controlled (subclause E). If the agency insists on 90 days, the artifacts are probably not in a 14-day-deliverable state, and the longer notice is needed to manufacture them.
Is “termination for material breach” still relevant if I have no-cause termination?
Yes. Material-breach termination is a separate mechanism; it typically eliminates the pro-rated-fees obligation for the breaching party and may trigger remedies. Keep both clauses. No-cause is the clean exit; for-cause is the protective exit when the founder needs legal posture to recover damages.
What about minimum commitment periods or annual contracts?
A 90-day minimum on a first engagement is reasonable. Anything beyond 180 days is uncommon in 2026 AI engagements and warrants pushback. A 12-month minimum is a red flag; the agency is signaling that they do not expect to deliver value fast enough to retain the relationship voluntarily. If they insist, structure as a 90-day minimum followed by month-to-month with the same notice.
How does the exit clause interact with payment milestones?
Closely. The pro-rated-fees provision in subclause A should reference the milestone structure: fees pro-rated to the most recently accepted milestone, with partial credit for in-progress work. This is why subclause D matters; IP transfer aligns with payment, so the founder is rarely paying for milestones whose artifacts they do not own.
What if the agency uses third-party SaaS (Modal, Hugging Face, W&B)?
Fine, provided artifacts are mirrored in real time to client-owned infrastructure (subclause E). The test is whether the founder can pull a current copy of any artifact at any moment and redeploy it elsewhere. If production depends on a Modal volume only the agency can access, the engagement has built lock-in regardless of what the contract says.
Can the agency carve out their own internal tooling?
Yes, through a Pre-Existing IP schedule attached to the MSA. The agency lists reusable internal libraries, helper frameworks, and generic prompt patterns; the founder receives a perpetual royalty-free license to use those components in the delivered system but does not receive ownership. The carve-out should be specific and named, not a generic “agency methodology” line.
Does the consult retainer carry forward unused hours?
The clean version is “100 hours over 12 months, no carry-forward, no refund.” This caps the agency’s forward exposure. The agency is being paid an option premium to reserve capacity post-termination, not to bank hours indefinitely.
How do these subclauses interact with the EU AI Act?
The exit clause is partly regulatory-readiness. Under EU AI Act Article 26, the deployer of a high-risk AI system inherits post-market monitoring, human oversight, and instruction-following obligations. Without subclauses B, D, and E, the deployer cannot meet those obligations because the artifacts and data flows live with the agency. Founders building toward EU sales should treat these subclauses as regulatory plumbing.
Key takeaways
- The exit clause is the most diagnostic clause in an AI agency contract. Resistance to A, B, D, or G is a near-perfect signal of a lock-in commercial model.
- Decompose into seven subclauses: 30-day no-cause termination, 14-day handoff, optional 6-week paid transition, milestone IP vesting, source-controlled registries, post-exit consult retainer, and no non-compete on client AI roadmap.
- Source-controlled registries (E) are the operational forcing function; without them, the 14-day Sunset Package is undeliverable in practice.
- Milestone IP vesting (D) eliminates the most common offboarding dispute: who owns artifacts shipped before termination.
- Treat the exit clause as regulatory plumbing; EU AI Act deployer obligations cannot be met without these artifact-transfer mechanisms.
Related reading
- The AI Agency Manifesto; the pillar this exit clause operationalizes.
- AI Agency Contract Negotiation; broader MSA / SOW / DPA / NDA structure.
- The 7 Commitments Most AI Dev Agency Should Make in Writing; operational commitments that make the exit clause deliverable.
- The Hidden Y Problem: Who Owns the Model Weights?; the artifact-class breakdown.
- Field Guide to Evaluating an AI Agency in Under 90 Minutes; pre-signature diligence.
Arthur Wandzel is the founder of SFAI Labs, a forward-deployed AI development agency in San Francisco. He has negotiated exit clauses on more than 30 AI engagements over the last 18 months, on both sides of the table.
Arthur Wandzel
