What to Put in an AI Development Contract (and What to Never Sign)

The most expensive mistake a non-technical founder makes isn't picking the wrong stack. It's signing an AI development contract that doesn't say who owns the fine-tuned model, doesn't define done, and quietly ties them to the vendor's cloud account.

We've been on both sides of these contracts. This is the checklist we use at Augere Labs — the one we'd want a founder to run against our own SOW before signing.

The ten checks that matter

1. Who owns the code, and when does ownership transfer?
2. Who owns the prompts, evals, and fine-tuned weights?
3. Whose cloud account does this run in?
4. What's the definition of done — in measurable terms?
5. What happens when the model gets worse (regression clause)?
6. Who pays for third-party services (models, storage, vector DB)?
7. What's the confidentiality and data-use scope?
8. What's the exit clause?
9. What's the payment schedule tied to?
10. What's excluded — and how do change requests get priced?

Every AI SOW should answer these before it's signed. If any one is missing, negotiate before you sign — not after.

1. IP ownership on the code

Standard clause: "All work product created under this agreement is a work for hire and is the sole property of the Client upon payment in full."

Two things to watch for:

• "Upon payment in full" — if final payment is disputed, ownership stays with the vendor. Consider staged transfers on milestone payments, not one big transfer at the end.
• Carve-outs for the vendor's "background IP." A vendor with reusable components is legitimate, but the clause should be specific: "generic utility libraries and pre-existing frameworks." Not "any code we've written before."

2. Ownership of prompts, evals, and fine-tuned weights

This is the clause most contracts miss entirely. Prompts are the crown jewels of an AI product. Evals are the tests that prove it works. Fine-tuned weights are potentially the moat.

Explicit language we recommend: "All prompts, eval sets, fine-tuned model weights, and vector database contents created or curated under this agreement are Client-owned deliverables and shall be transferred at each milestone."

Ask for prompts in the repo, not in the vendor's private LangSmith. Ask for eval sets in a Google Sheet you own. Ask for weight files in a bucket you control.

3. Whose cloud account?

A pattern we see too often: the vendor builds everything in their AWS or Supabase account "for convenience." Six months later the founder is locked in — the whole product runs on infrastructure they don't control.

Non-negotiable clauses:

• The client's cloud, hosting, database, and model provider accounts are used from day one.
• The vendor gets scoped IAM access, not root credentials.
• All secrets and API keys are provisioned by the client.

If the vendor pushes back on this, that's the signal. A senior team wants your account because they've been burned by handovers too.

4. What "done" means

"The AI feature will work" is not a definition. Definition of done for an AI project has to include:

• Which endpoints exist and what they return
• What the eval set is and what pass rate is required (e.g. "≥85% pass at temperature 0 on the 40-example eval set in Appendix A")
• What p95 latency is acceptable
• What the monthly inference budget is
• What the UX acceptance criteria are per screen

See our AI spec guide for the format. The spec should be an appendix to the contract, not a separate doc.

5. Regression clause

AI features degrade. Models change. The internet changes. A contract without a post-launch regression clause is a contract that ends the moment the vendor invoices the last milestone.

Reasonable clause: "For 30 days after launch, the vendor will address any regression in the eval set pass rate at no additional cost. Beyond 30 days, a monthly maintenance retainer applies."

6. Third-party costs

Model bills, hosting, and vector DB fees can be a real number. The contract must say who pays for them during the build and after handover.

Standard shape: "Third-party service fees (OpenAI, Anthropic, hosting, database) are the Client's responsibility and are billed to the Client's accounts. The Vendor is not liable for cost overruns arising from Client-directed usage."

7. Confidentiality and data use

Two clauses matter here — one obvious, one not.

• Standard NDA: vendor cannot share client data or code externally. Standard.
• Model provider data use: the vendor commits to using zero-retention endpoints (OpenAI's -latest tier with data-use opt-out, Anthropic's default no-training). This matters if your users' data flows through the model.

8. Exit clause

What happens if you fire the vendor? A clean exit clause includes:

• Access to all repos, prompts, evals, weights, and infra credentials within 48 hours of termination
• Documentation handover (a runbook, not a Loom)
• Optional paid transition support (rate defined up front)

If the vendor won't put an exit clause in writing, they're planning to make you dependent. That's a red flag.

9. Payment schedule

Never pay 100% upfront. Never pay 100% on completion. Somewhere in between, tied to milestones the client can independently verify.

The shape we prefer for a 30-day MVP:

• 25% on signing (locks in calendar)
• 25% on spec sign-off (end of week 1)
• 25% on staging deploy (end of week 3)
• 25% on production launch + eval pass (end of week 4)

For longer projects, monthly milestones tied to concrete artifacts (a shipped feature, a passing eval, a deployed service) beat time-based milestones every time.

10. Exclusions and change requests

Every SOW should have an explicit "Not Included" section. Common items: mobile apps, SSO, SOC 2 compliance work, migrations from legacy systems, third-party integrations not listed by name.

Change requests need a price and a process. "$X per hour" is fine. "TBD" is not. Even better: "Change requests will be scoped and quoted within 3 business days; work begins after written approval."

Clauses to negotiate hard (or refuse)

Red flags we've seen and pushed back on:

• Auto-renewing retainers. A contract that renews unless you cancel 60 days in advance is designed to punish inattention.
• Non-compete on client's business. Occasionally acceptable; usually overreach.
• Vendor-owned Github org. Never. The repo lives on your Github, from day one.
• Unlimited revisions. Sounds generous. It's actually a scope trap.
• IP transfers "in the vendor's discretion." Ownership either transfers on schedule or it doesn't. No discretion clauses.

Fixed-price vs T&M for AI projects

Fixed-price contracts protect founders when scope is clear. T&M (time and materials) contracts protect vendors when scope is fluid. AI projects with a signed spec should be fixed-price — the scope is knowable. AI projects labeled "exploratory" almost always come in over budget on T&M.

Our default at Augere Labs: fixed-price for the MVP, month-to-month retainer for post-launch iteration. That combo keeps incentives clean.

What to bring to the negotiation

• Your AI product spec (see the linked guide)
• Your eval set — even a starter 20-example version
• A list of the third-party services you already use (auth, hosting, model provider)
• A specific launch date, not a "soon"

Vendors write better contracts against a founder who's done homework. Signal quality up front and you'll get better terms.

A common misconception

"The contract is boilerplate — the relationship is what matters." Half true. The relationship matters most while both sides are happy. The contract matters most the day one side isn't. Design for that day.

Frequently asked questions

Should I use a lawyer for a $10K AI project?

For a $10K project, a lawyer might cost more than the risk. Use a solid template (Bonsai, Bloom, or a vetted MSA) and negotiate the specific IP, exit, and definition-of-done clauses yourself. For projects above $30K, hire a lawyer — an hour of their time is cheap insurance.

Who owns the AI model at the end of the project?

Whatever the contract says. Default state under most SOWs is ambiguous, so make it explicit: prompts, evals, fine-tuned weights, and vector data are all client-owned deliverables. Get that in writing.

What's a fair payment schedule for an AI project?

Four milestones of 25% each, tied to signing, spec sign-off, staging deploy, and production launch. Avoid 50% upfront (vendor loses urgency) and 100% on completion (vendor eats the cash flow).

What if the vendor uses AI to write the code?

Fine, and increasingly universal. The clause that matters is the same: the resulting code is a work for hire, and the client owns it. Whether a human or an AI produced it doesn't change the IP status.

How do I protect against a vendor disappearing?

Client-owned cloud accounts, repo on the client's Github, weekly commits (not batch drops), and a written exit clause. If all four are in place, a vendor disappearing is inconvenient — not fatal.

Want the template we use?

At Augere Labs we hand every client our SOW template with the ten clauses above pre-filled. If you're about to sign with any vendor, book a 30-minute call — we'll read the contract with you and flag anything unusual.

Related: our agency hiring guide covers the earlier decision — which vendor to sign with in the first place.