Most founders negotiate AI partnership IP the same way they negotiate a SaaS contract — and lose six figures of value before they ever sign. A 2026 AI MVP engagement produces six distinct IP artifacts, not one bundle of “deliverables.” Each has a vendor default the partner will propose, a founder counter, a specific SOW clause, and a walk-away threshold. This is the playbook for the founder a week from signing.
The mistake is not in any single clause — it is in treating the SOW as the negotiating instrument. The real negotiation happens in a numbered appendix — Schedule B — that lists the founder’s commitments by artifact, signed by both sides, attached to the contract. This piece walks through how to build that Schedule B line by line, and which three vendor responses justify walking away rather than red-lining for another two weeks.
This script extends the founder-AI-partner operating manual inside the idea-to-product manifesto. This is a founder-facing operating script, not legal advice — validate every clause with counsel.
Why the SOW is not the negotiation — Schedule B is
A typical 2026 AI partnership SOW has a 30-page main body and a half-page “Intellectual Property” section. It says, “Founder owns work product; Partner retains background IP; transfers shall be effected at termination.” Founders sign and assume it covers everything.
It covers two things — source code and visual deliverables. The other four categories are governed not by the SOW’s IP section but by where the partner stores the artifact. Prompts live in the partner’s Cursor history. Eval runs live in the partner’s Weights and Biases project. Fine-tuned adapters live on the partner’s Hugging Face account. Production inference runs through the partner’s OpenAI enterprise contract. None are addressed by “Founder owns work product.”
The fix is Schedule B — a numbered appendix attached to the SOW listing six commitments by artifact: who owns it, where it lives, what format it transfers in, by what date. Schedule B is more enforceable than the MSA because it is bespoke, signed, and operational. The next six sections give each category’s vendor default, founder counter, Schedule B language, and walk-away threshold.
The six IP categories you are actually negotiating
Each row in the table below is one Schedule B line item.
| Category | Vendor default | Founder counter | Walk-away if… |
|---|---|---|---|
| 1. Source code | Founder owns at termination | Founder owns from commit one | Partner refuses founder-owned repo from day 1 |
| 2. Prompts | Partner’s Cursor history | Versioned in founder repo | Partner refuses to mirror prompts into the repo |
| 3. Eval set | Partner’s W&B project | Plain-text in founder repo, one-command runnable | Partner refuses to deliver eval set in re-runnable form |
| 4. Fine-tuned weights | Partner’s Hugging Face account | Founder’s HF org, base-license + license-of-delta documented | Partner claims fine-tuned adapter as proprietary IP |
| 5. Customer data | Partner stores, processes, may use for training | Founder-controlled storage, GDPR processor only, no training use | Partner refuses GDPR Art. 28 processor clause |
| 6. Model API account | Partner’s OpenAI / Anthropic enterprise account | Founder’s enterprise account, direct billing | Partner refuses to migrate API account to founder before launch |
Order matters. Code is the easy concession — asking for it builds momentum. The API account is the hardest and most valuable; surface it last, after five concessions establish the negotiation pattern. McKinsey’s 2024 State of AI report identifies evaluation infrastructure as the largest pilot-to-production gap, which is why the eval set sits in the middle of the table — the artifact founders most often underweight, and the one partners most quietly retain. If you have not yet run the diagnostic version of this with the partner, see the IP and weights conversation every founder should have on day 1 — that 7-question kickoff script is the input to this 6-category negotiation.
Category 1: Source code
Vendor’s default: “Founder owns the source code at the conclusion of the engagement, upon final payment.” Translation: partner controls a private repo until termination; founder gets a zip file with no commit history and no proof the code runs.
Founder’s counter: Source code, commit history, and CI/CD configuration are committed to a founder-owned GitHub or GitLab organization from commit one. The partner gets revocable team-member access.
SOW Schedule B language: All source code, repository history, and CI/CD configuration shall be hosted in a Founder-owned VCS organization from the first commit. Partner shall be granted revocable team-member access. A RUNBOOK.md describing the steps to run the system on a clean environment within four hours shall be committed before any milestone payment is released.
Walk-away: Partner refuses founder-owned repo from day 1. Either the partner retains IP through the access channel or their internal tooling only works inside their own monorepo. Code is the easy negotiation; if the partner cannot concede here, the rest will not improve. See how AI development partnerships handle IP and code ownership for the broader IP-category primer.
Category 2: Prompts and prompt registry
Vendor’s default: Silence — the SOW does not mention prompts. The partner keeps them in Cursor history, a LangSmith project, or a notebook on their workstation. Prompts encode 20–30% of the engagement’s behavior; rebuilding at offboarding takes weeks and the rebuilds behave differently because iteration history is lost.
Founder’s counter: All production prompts live in /prompts/ in the founder-owned repository, one file per prompt with structured metadata (version, author, dataset, eval-set hash). Externally stored prompts mirror to the repo within one business day.
SOW Schedule B language: All production prompts shall be stored in /prompts/ in the Founder-owned repository, with version metadata, author attribution, and the eval-set commit hash against which each prompt was validated. No production prompt may live exclusively outside the repository.
Walk-away: Partner refuses to mirror prompts, citing “iteration speed” or “internal tooling.” Translation: the partner intends to keep iteration knowledge non-transferable. Founders who concede this own the codebase but cannot reproduce its behavior — see how to evaluate an AI vendor’s eval discipline before signing.
Category 3: Eval set and rubric
Vendor’s default: “Evaluation is performed by Partner using industry-standard tools. Results will be reported in quarterly business reviews.” Translation: eval set, rubric, and historical data live in the partner’s Weights and Biases (or Braintrust, or Langfuse) project. The founder sees scores in a slide deck and cannot run the suite themselves.
Founder’s counter: The eval set is plain-text files (JSONL, CSV, or YAML) in the founder-owned repo. The rubric is documented in markdown. A single command — make evals — runs the full suite on a fresh laptop in under 30 minutes. Historical results export at every milestone.
SOW Schedule B language: The complete evaluation set, evaluation rubric, and historical evaluation results shall be committed to the Founder-owned repository in plain-text formats. A documented command (e.g., make evals) shall run the full evaluation suite locally on a clean environment. Eval-set exports shall accompany every milestone payment.
Walk-away: Partner refuses to deliver the eval set in re-runnable form, citing proprietary tooling. Translation: the partner has reusable eval IP they intend to retain across clients — a legitimate model, but not a partnership, and the SOW price should reflect that.
The eval set is the highest-value artifact in a 2026 AI engagement. Source code can be rewritten in weeks; the eval set encodes months of negotiated definitions — what counts as a correct answer, an acceptable hallucination, a real edge case. Re-deriving it is a six-to-twelve month rebuild. Bain’s 2024 Beyond Hype report notes more than half of enterprise AI projects stall at handoff, and the eval-set gap is the most common proximate cause.
Category 4: Fine-tuned model weights
Vendor’s default: “Partner retains rights to fine-tuned model artifacts as part of Partner’s background IP and methodology.” Translation: if the engagement fine-tunes Llama 3.3, Mistral, or any open-weight model, the LoRA adapter sits on the partner’s Hugging Face account under whatever license the partner chooses. The founder paid for the fine-tune and cannot deploy it without the partner’s permission.
Founder’s counter: Any fine-tuned weights, adapters, or LoRA artifacts upload to a founder-controlled Hugging Face organization under a license selected by the founder. The base-model license (Llama 3.3 community license, Mistral Apache-2.0) carries through. Training script, dataset snapshot, and hyperparameters are committed to the founder-owned repo so the founder can re-run the fine-tune if the artifact is lost.
SOW Schedule B language: Any fine-tuned model weights, adapters, LoRA, QLoRA, or full-parameter checkpoints produced under this engagement shall be uploaded to a Founder-controlled Hugging Face organization, under a license selected by the Founder. The base-model license shall be preserved and documented. The training script, dataset snapshot, and hyperparameter configuration shall be committed to the Founder-owned repository.
Walk-away: Partner claims the fine-tuned adapter as proprietary IP. This is the clause where partners push back hardest — many genuinely believe their fine-tuning recipe is what they sell. The correct response: the recipe (training script, hyperparameters) can be jointly licensed; the artifact (the weights) cannot, because the founder paid for the compute. Conceding means the founder rented a model they bought — see the hidden Y problem in AI agency contracts.
Category 5: Customer data and training-data carve-out
Vendor’s default: “Customer data may be used for model improvement, evaluation, and methodology refinement, subject to anonymization.” Translation: the partner reserves the right to use end-user data — anonymized — to train models, build eval sets, or improve their internal tooling. Some of that damages the founder if the partner trains a competitor product on the founder’s user behavior.
Founder’s counter: Customer data lives exclusively in the founder’s own cloud (AWS, GCP, Azure, or managed Postgres on the founder’s bill). The partner is a GDPR Article 28 data processor, not a controller. A DPA is signed before any production data touches the system. No retention, copying, export, or reuse — including for training, fine-tuning, or eval-set construction.
SOW Schedule B language: All end-user data shall be stored exclusively in Founder-controlled infrastructure under accounts billed directly to Founder. Partner shall act as a data processor under GDPR Article 28 and shall execute a Data Processing Agreement before any production data is processed. Partner shall not retain, copy, export, or use end-user data for any purpose outside this engagement, including model training, fine-tuning, or evaluation-set construction. The carve-out applies to anonymized and aggregate data unless separately licensed in writing.
Walk-away: Partner refuses GDPR Article 28 processor language, or tries to retain “anonymized aggregate” rights to user data. A partner who has run production AI systems knows the controller-vs-processor distinction is non-negotiable. Resistance signals inexperience or an undisclosed business model that depends on data reuse.
Category 6: Model API account
Vendor’s default: “Partner shall provide model inference infrastructure under Partner’s enterprise vendor agreements, with usage billed through Partner’s account at cost plus a 20% management fee.” Translation: the production API account is in the partner’s name. The partner can revoke access on 24 hours’ notice. On a $200K annual inference bill, the 20% margin is $40K per year for routing traffic.
Founder’s counter: Production API traffic runs through the founder’s own enterprise account with each vendor, from day 1. The founder signs OpenAI’s enterprise terms, Anthropic’s commercial terms, and any model-hosting contract directly. The founder pays vendors directly. The partner gets revocable team-member access. No pass-through billing, no margin layer.
SOW Schedule B language: Founder shall hold, in Founder’s legal name, all production enterprise accounts with model API vendors (including OpenAI, Anthropic, Google, and any model-hosting providers in use). Founder shall pay each vendor directly. Partner shall be granted revocable team-member access. Partner shall not act as a reseller, pass-through, or margin layer on any vendor invoice. All BAAs, DPAs, and custom enterprise terms shall be in Founder’s name.
Walk-away: Partner refuses to migrate the API account to the founder before production launch, or insists on a margin layer. Translation: the partner depends on the inference-margin layer for engagement economics; the founder is being charged a markup to be locked in. This is the single most decisive clause in a 2026 AI engagement — concede it and the founder has bought a managed service, not a partnership.
The three walk-away clauses
A founder can negotiate any clause. The decision to walk should be triggered by exactly three vendor responses — each one signals a structural problem no SOW clause can fix.
Walk-away 1: Refusal to migrate the model API account to founder name before launch. If the partner will not move the production OpenAI or Anthropic account to the founder’s name before traffic flows, the partner has structured the engagement to depend on their vendor relationships. That dependence is the lock-in. Walk.
Walk-away 2: Refusal to license partner-curated synthetic or eval data perpetually and irrevocably to the founder. If the partner curates synthetic training data or eval examples and insists on retaining them as background IP, the founder paid for an asset they cannot keep using after offboarding. The lost asset value typically equals 15–25% of the engagement. Renegotiate the dollar number down or walk.
Walk-away 3: Refusal to write the negotiated commitments into a signed Schedule B appended to the SOW. A partner who agrees verbally but refuses to write commitments into a signed schedule is signalling the verbal commitments will not hold. No schedule, no enforceable agreement. Walk.
The cost of walking on day 0 is one polite letter. The cost of walking on day 91 is six figures, three months, and a half-built system that cannot be transferred without a complete rebuild — see the graceful exit guide for offboarding mechanics if you are past day 0.
How to run the negotiation in five working sessions
The six categories take roughly five working sessions to negotiate cleanly. The partner’s technical lead must be in the room — not just the account executive — for sessions 1 through 3.
- Session 1: Code, prompts, eval set. Walk through categories 1–3 in order. Capture answers verbatim. Send a draft Schedule B for these categories within 24 hours.
- Session 2: Fine-tuned weights, customer data. Categories 4 and 5. The partner will resist hardest on weights. Hold ground — the recipe is licensable; the artifact is not.
- Session 3: Model API account. The closer. Treat as a single dedicated session because it surfaces the partner’s engagement economics. If the partner hedged earlier, this discussion will expose why.
- Session 4: Schedule B consolidation and counsel review. Combine the three drafts into a single Schedule B. Send to both counsels in parallel. Ask for redlines within five business days.
- Session 5: Redline reconciliation. A real partner returns minor language tweaks. A reluctant partner returns the high-stakes clauses (eval set, weights, API account) stripped. The redline itself is a diagnostic. Accept reasonable tweaks; hold ground on substance. Sign or walk.
The sequence takes 10–15 calendar days — small relative to engagement length, tiny relative to offboarding cost without a Schedule B. For a pre-built starting point, download the AI MVP scoping worksheet — it includes a Schedule B template ready to redline. If you want a second pair of eyes before signing, book an idea review — we will walk through your draft and flag the clauses most likely to be stripped.
Frequently asked questions
When in the sales cycle should I introduce Schedule B?
After the partner sends their standard SOW and before you sign it — typically the second or third sales call. Earlier, the partner has not committed enough to negotiate. Later, you have lost negotiating room. Schedule B is most effective while the partner is still in pursuit; once you have signed and paid, you are on the partner’s clock.
Will partners actually agree to all six categories?
A boutique partnership-model firm will agree to five and negotiate hard on one (usually weights). A traditional agency will agree to two or three. Neither is wrong, but the dollar number should reflect which model you are buying. A partnership earns its price by ceding more IP; an agency keeps more IP and should cost less for the same scope.
What if the partner says their lawyer needs to draft Schedule B?
Decline politely and send your own draft first. Whoever writes the first draft sets the framing. Your Schedule B should match this article’s structure with explicit founder-counter language. The partner’s counsel can redline — but they cannot reframe — what they are responding to.
Is the API account negotiation realistic for a founder without an existing OpenAI relationship?
Yes. OpenAI and Anthropic both accept new enterprise accounts at any usage level above the free tier with 5-minute sign-up and a credit card. The “we have a special enterprise discount you do not have access to” framing is rarely true under $500K in annual inference, and even when it is, the discount is rarely worth the lock-in. Ask for the specific discount percentage in writing — most partners cannot produce it.
Do I need a startup lawyer to negotiate this?
Draft Schedule B yourself using this article’s language. Send to the partner’s counsel. When their redline comes back, engage your own startup lawyer — typically 2–4 hours of counsel time. The framing is operational; the legal review is a sanity check at the end, not the substance.
What if I have already signed the SOW without a Schedule B?
Negotiate Schedule B as an amendment now, before the next milestone payment. Partners are most willing to amend when the next payment is contingent on it. Frame the amendment as “operational alignment for offboarding” rather than “renegotiation,” which removes the threat framing. If the partner refuses, treat it as a walk-away signal for the next renewal — see how to end an AI partnership when it’s not working for the structured exit path.
Should I share this article with the partner before the negotiation?
Yes — and the partner’s response is itself a diagnostic. A real partner reads it, agrees with most categories, and proposes Schedule B language in the next call. A reluctant partner deflects with “every engagement is different” or “our standard terms cover this.” The deflection is the signal.
Which single clause matters most if I can only negotiate one?
The model API account. Conceding it means every other category becomes negotiable on the partner’s terms because the partner controls the on/off switch. Holding it means every other category can be re-litigated at offboarding because the partner has no remaining hold over the system.
Key takeaways
- The IP and weights negotiation has six categories — source code, prompts, eval set, fine-tuned weights, customer data, and model API account. Each has a vendor default, a founder counter, specific SOW language, and a walk-away threshold.
- The negotiation does not happen in the SOW. It happens in Schedule B, a numbered appendix listing all six commitments with operational specificity. Schedule B is more enforceable than the MSA.
- The three walk-away clauses: refusal to migrate the API account, refusal to license partner-curated data perpetually, refusal to write commitments into a signed Schedule B. Each signals a structural problem.
- The model API account is the most decisive clause. The eval set is the highest-value artifact. Weights attract the hardest resistance. Negotiate in this order: code → prompts → evals → weights → data → API account.
- Five working sessions and 10–15 calendar days are enough to negotiate all six. Skipping costs six figures and three months at offboarding. Doing it costs one polite letter.
Arthur Wandzel
