Most non-engineer founders pick their AI partner from a 90-minute pitch and three case studies, then wonder why the contract surprises them in week 6. They asked about portfolio depth and pricing. They did not ask about evals, model-weights ownership, regression-suite handoff, or the named-lead substitution clause. The 11 items below are the contract clauses and artifacts a founder can demand inside a single 30-minute conversation. Each names what good looks like (the answer that signals a partner who will survive the build) and what hand-wave looks like (the boilerplate that signals a partner who will not). The checklist is a contract instrument, not a personality test.
Table of Contents
- Why a checklist and not a vibe check
- The 11 must-haves
- 1. Eval-bound SOW
- 2. Fixed-price clarity under variable token spend
- 3. IP ownership clauses — including weights and prompts
- 4. On-call window for model failures
- 5. Regression-suite handoff as a contracted artifact
- 6. Weekly cadence written as artifact list
- 7. Kill-clause at a named milestone
- 8. Day-1 IP conversation, not week-9 IP scramble
- 9. Named technical lead with a no-substitution clause
- 10. Reference clients that match your stage
- 11. 30-day post-launch contractual window
- How to run the checklist in 30 minutes
- Frequently Asked Questions
- Closing
Why a checklist and not a vibe check
The founder-vendor literature is full of advice on “trust your gut” and “cultural fit.” Both are real, neither is sufficient. In a SaaS-era engagement a founder could discover a poor partner choice in week 8 and still have a working product to renegotiate around. In an AI MVP engagement the founder discovers it on day 1 of production when the eval scores collapse and nobody owns the regression. McKinsey’s State of AI 2024 and Bain’s 2024 generative-AI survey both find that the operational variance between high-performing and low-performing AI deployments sits in cadence, governance, and eval discipline — not in model selection. A vendor that treats evals as decoration will produce an MVP the founder cannot defend in front of an investor or a customer.
This article sits just beneath the founder-AI-partner operating manual, which covers the operating rhythm after the contract. The companion first-90-days playbook is the calendar from week 0 onward. Both descend from the idea-to-product manifesto.
The 11 must-haves
The list is ordered by the cost of getting each item wrong, highest first. Items 1-5 are the AI-specific contract clauses; 6-9 are operating-model commitments; 10-11 are post-signature protections. None is optional. A partner who hedges on more than two of the 11 is not a partner to sign.
| # | Item | Cost of getting it wrong |
|---|---|---|
| 1 | Eval-bound SOW | “Done” becomes a personality opinion |
| 2 | Fixed-price clarity under variable token spend | Surprise token bill at week 8 |
| 3 | IP ownership including weights and prompts | Founder rents the prompt library back |
| 4 | On-call window for model failures | Production failure at 2 a.m. and nobody answers |
| 5 | Regression-suite handoff | Founder cannot rerun evals post-handoff |
| 6 | Weekly cadence as artifact list | “We’ll meet weekly” produces vibes, not artifacts |
| 7 | Kill-clause at named milestone | Trapped in a failing $200K SOW |
| 8 | Day-1 IP conversation | Week-9 IP scramble delays launch |
| 9 | Named lead, no-substitution clause | Senior pitched, junior ships |
| 10 | Reference clients matching your stage | Fortune-500 logos mask zero seed-stage experience |
| 11 | 30-day post-launch window | Drift and first-cohort traffic become founder problems |
1. Eval-bound SOW
The single most important clause in any AI MVP contract.
What good looks like. The SOW names a specific eval set — at least 30 cases drawn from real or representative user inputs — and a numerical threshold (for example, 90 percent task-completion or 85 percent faithfulness) above which the product is contractually “done.” The partner agrees the eval set is the acceptance criterion and commits to publishing the eval pass/fail report every Friday. The eval set is co-owned: the founder seeds it from real customer language, the partner expands it.
What hand-wave looks like. “We’ll set up evals when we get there.” “Evals are part of our quality process.” A partner unwilling to make the eval set a contractual artifact is a partner who reserves the right to declare the product done on a vibes basis. See the eval-bound SOW and the role of evals in your weekly partner relationship.
2. Fixed-price clarity under variable token spend
Fixed-price is the right shape for an AI MVP contract — but only if it is honest about the variable cost line nobody discusses until week 6.
What good looks like. The fixed-price covers engineering labor, planning, eval-set authoring, handoff, and the 30-day post-launch window. Token spend is named as a separate, founder-borne line with an estimated monthly ceiling in the SOW, plus a contractual obligation that the partner notifies the founder within 24 hours if production traffic projects past 120 percent of the ceiling. The partner provides a cost-per-call estimate by week 3 and a real reading by week 6.
What hand-wave looks like. “Token costs are included in the price” — usually means the partner has underestimated them and will book a change order in week 8. “Don’t worry about token costs at this stage” — means the partner is not modeling unit economics. See fixed-price AI MVP contracts: what’s in scope vs. what’s not and the hidden cost of AI token bills.
3. IP ownership clauses — including weights and prompts
Most SaaS-era IP clauses leave the founder with the source code and the partner with everything that mattered.
What good looks like. The IP clause names four artifact classes — source code, prompt libraries, fine-tuned model weights or adapters, and eval sets — and assigns full ownership of each to the founder upon final payment. Background IP and open-source dependencies are listed explicitly. The partner retains no claim on prompt engineering as “methodology” and no claim on fine-tune artifacts as “tooling.” A clause specifically covers production prompts that change weekly.
What hand-wave looks like. “You’ll own the code, of course” — without specifying prompts, weights, or evals. “Our prompt library is our IP” — means the founder is renting the most valuable artifact in their own product. See the IP and weights conversation and how AI development partnerships handle IP.
4. On-call window for model failures
AI products fail differently than SaaS products. Model deprecations happen on the vendor’s calendar, not the founder’s. Token bills can 10x overnight. Eval scores can drift without code changes.
What good looks like. The contract names an on-call window — usually 90 days post-launch — during which a named partner engineer responds inside 4 business hours to one of three events: a production eval-score drop below threshold, a token-bill projection exceeding the agreed ceiling, or a model-deprecation announcement from a major vendor (OpenAI, Anthropic, Google, Meta). On-call coverage is included in the fixed-price; emergency engineering past 90 days is a named hourly rate.
What hand-wave looks like. “We’re always available on Slack” — not a contract. “Bug fixes are covered for 30 days” — wrong surface area; the failure modes above are not bugs. “Model deprecations are out of scope” — means the founder owns a forced migration when GPT-5 deprecates. See why your AI project budget should have a model deprecation reserve.
5. Regression-suite handoff as a contracted artifact
The eval set the partner used to ship the MVP is the artifact that survives the engagement — but only if the contract treats it that way.
What good looks like. The handoff deliverable list names a regression suite that runs locally on a founder-owned machine, with a documented make eval or equivalent one-command interface, a README that explains what each eval case tests, and a confidence interval on each metric. The suite includes the production eval cases, a smoke-test subset that runs in under 60 seconds, and a CI-ready format. The partner trains a founder-side engineer to run the suite in a 60-minute handoff session.
What hand-wave looks like. “Evals live in our internal tooling” — means the founder cannot rerun them. “We can rerun evals as a service” — means the founder pays monthly for access to their own quality artifact. “The repo has tests” — unit tests are not evals. See the handoff process — what you get when an AI MVP ships.
6. Weekly cadence written as artifact list
“We’ll meet weekly” is not a contract clause. It is a vibe.
What good looks like. The contract names the weekly cadence as a list of artifacts the partner produces every Friday: a 10-minute Loom of that week’s demo, an eval pass/fail report by category, a change-order register update, a cost-per-call reading, and a one-paragraph risk note (on-track, at-risk, deferred). The founder receives the bundle Friday afternoon, reviews async, and runs a 30-minute live review on Monday or Tuesday. See the weekly founder-partner cadence for each artifact.
What hand-wave looks like. “We do agile standups” — meaningless for a non-engineer founder. “We send a status email” — means the founder reads prose instead of inspecting evals. “We use Linear / Jira / Asana” — means the founder reconstructs what the tool was supposed to summarize. The companion piece anatomy of a great AI agency kickoff covers establishing the cadence on day 1.
7. Kill-clause at a named milestone
The most useful clause is the one the founder hopes never to use.
What good looks like. The contract names at least one milestone (typically the end of week 2 after the PRD and eval baseline, and again at the end of week 6 after the mid-point readout) at which either party can terminate with no penalty beyond work-to-date pro-rata. The clause requires written notice, a 5-business-day transition window, and IP and artifact transfer on the original terms. A partner that resists the clause is signaling that they intend to charge regardless of progress.
What hand-wave looks like. “We don’t normally include exit clauses; we earn the trust.” — translates to “we want the founder financially committed before the first hard week.” “Termination requires 30 days notice and full SOW payment” — means there is no exit. See the AI agency exit clause and the graceful exit.
8. Day-1 IP conversation, not week-9 IP scramble
The IP conversation that happens in week 9 — when lawyers surface model-weight and prompt-library questions just before launch — is the most expensive conversation in the engagement.
What good looks like. The first 60-minute conversation in week 0 (or in the discovery call) names the four IP artifact classes — code, prompts, weights, evals — and lands the ownership default for each before the contract is drafted. Any deviations are written into the SOW, not left in email. The conversation surfaces customer data, training data, and PII handling at the same time.
What hand-wave looks like. “Let’s deal with IP at the end; the lawyers will work it out.” — means the founder will negotiate at launch, when negotiating power is gone. “Standard IP terms; nothing to discuss.” — means the founder is signing a SaaS-era template that does not name AI artifacts.
9. Named technical lead with a no-substitution clause
The senior engineer who sold the pitch is not always the engineer who writes the code.
What good looks like. The SOW names a single technical lead by name, with a CV or LinkedIn link, and specifies a minimum allocation (typically 25 to 50 percent of their time during the build). A no-substitution clause requires the founder’s written consent for any change to the named lead. The named lead attends every weekly review and the mid-point readout.
What hand-wave looks like. “We staff dynamically based on project needs.” — means the senior engineer pitched, the junior engineer ships. “Our team is interchangeable; quality is uniform.” — almost never true. See the AI product studio — a different operating model for the staffing model that makes named leads structural.
10. Reference clients that match your stage
Three Fortune-500 references mask zero seed-stage experience. A solo founder needs to hear from a solo founder.
What good looks like. The partner provides three references whose engagement profiles match the founder’s stage, sector, and team shape. References are recent (within 18 months) and reachable. The partner is willing to introduce by email. Ideally one reference is an engagement that did not go as planned, with the partner naming what changed in their process as a result.
What hand-wave looks like. “We’re under NDA on most engagements.” — common, but if all three references are NDA-only, the partner is signaling weak portfolio breadth. “Here are the logos of clients we’ve worked with.” — not a reference. “Our case studies are on the website.” — case studies are marketing; references are diligence. See how to evaluate an idea-to-product partner’s prior work.
11. 30-day post-launch contractual window
Launch is not the end of the engagement; it is the moment the eval set meets reality.
What good looks like. The SOW names a 30-day post-launch window during which the partner is contractually responsible for production eval-score monitoring, regression bug fixes, change orders inside the original scope, and weekly readouts on real-user behaviour. The window is included in the fixed-price, not a separate retainer. At day 30, the partner produces a forward-looking memo — top-5 user surprises, recommended next iteration, ongoing-cost forecast, and a renew-scope-down-exit recommendation. See the 30-day post-launch period explained.
What hand-wave looks like. “Post-launch is a separate engagement.” — means the founder pays a second SOW the moment production traffic surfaces problems. “Bug fixes are warrantied for 7 days.” — too short; production patterns take 2 to 3 weeks to surface.
How to run the checklist in 30 minutes
Fit the checklist into a single 30-minute discovery call or SOW review — not the legal review that follows. Three minutes per item gets through 10; the last is the wrap. Take notes in a shared doc and mark each item green, yellow, or red.
Opening prompt: “I’d like to walk through 11 items I check with every prospective AI partner. Each is a yes / no / how question. Ready?” Closing prompt: “Any items you’d push back on? I’d rather know now than in week 6.”
The decision rule is mechanical. Green on at least 8 of 11 — proceed to SOW. Yellow on more than 3 — escalate to a second conversation with the named lead. Red on any of items 1, 2, 3, 5, or 7 — do not sign. Those five are the structural ones. See how to pick an AI development partner when you’ve never built software.
Frequently Asked Questions
What if my partner is excellent on items 6-11 but weak on 1-5? Do not sign. Items 1-5 are the AI-specific structural clauses. A partner strong on cadence and named-lead but weak on evals, IP, on-call, or regression handoff builds SaaS apps with AI features — not AI MVPs.
Can I run this checklist on a freelance senior engineer instead of an agency? Most items transfer. Items 6, 7, 9, and 11 compress. Items 1, 2, 3, 5, and 8 — eval-bound SOW, fixed-price clarity, IP, regression handoff, day-1 IP — are still mandatory.
The partner agrees to eval-bound SOW but wants to defer authoring the eval set until week 2. Is that okay? Yes. Co-author in weeks 1-2, threshold agreed before the build half begins. What is not okay is deferring the clause itself; the contract commits to evals as the acceptance criterion before signature.
My budget is $50K, not $200K. Do all 11 items still apply? Yes. What scales with budget is the depth of each item, not its presence. A $50K engagement still names an eval threshold, still has an exit clause, still produces a regression suite at handoff.
How do I run the checklist with a partner only available by email? Send the 11 items as a document, ask for written responses within 5 business days, then schedule a 30-minute call. A partner who cannot make a 30-minute call before signing a $50K-$200K SOW is signaling that the founder is one of many accounts. See why nearshore is the new offshore.
What if the partner refuses to share their SOW template before signing? Walk away. Every reputable AI MVP partner has a standard SOW they will share under NDA at the discovery-call stage. Refusing is the loudest hand-wave on the list.
Is the kill-clause realistic, or is it a partner deal-breaker? Realistic. Every senior AI partner has seen engagements that needed to terminate at week 2 or week 6; a clause that names the exit makes the conversation manageable rather than catastrophic.
My partner offered a kill-clause at week 6 but not week 2. Is that enough? Acceptable only if the partner also agrees to a no-fault termination at the mid-point readout (week 5). Two exit gates inside the build half is the standard.
What is the single most common item partners hand-wave on? The eval-bound SOW. Most agencies are uncomfortable making evaluations the contractual definition of done because it shifts risk onto the partner. That is precisely why the clause matters.
Should the contract name the eval framework — Promptfoo, Braintrust, LangSmith, Inspect? No. The contract names the eval set, the threshold, the cadence, and the handoff format. The harness is a tool, not a contract artifact.
Closing
The checklist is one of two BoFu instruments the founder operates before signing. The other is the founder’s first-90-days playbook — the calendar for the engagement once the contract is signed. Together they replace personality-and-portfolio diligence with mechanical, contract-shaped diligence. None of the 11 items is exotic. All are routinely missed.
The next step: book a 30-minute conversation with the SFAI Labs team and run the checklist against our standard SOW. We answer all 11 items in writing before the call, walk them live in 30 minutes, and the founder leaves with either a signed engagement or clear reasons to keep looking. Book a 30-minute discovery call and bring the checklist.
Arthur Wandzel
