A founder with three viable AI MVP partner shortlists is not under-informed — they are stalled. The longer the comparison runs, the more partners look the same, the more calendar bleeds, and the more the eventual choice gets made on tiredness rather than evidence. The remedy is not more research. It is a calendar-bound sprint that produces one named partner, one countersigned SOW, and one kickoff date in seven days.
This rubric operationalizes the founder AI partner operating manual within the broader idea-to-product manifesto. Seven days, seven deliverables, one scoring matrix, one decision memo, one signed contract. The rubric assumes the criteria and shortlist work is already done — that is the prerequisite, not the article.
Why a seven-day rubric
Three structural facts about AI MVP partner selection in 2026 make a calendar-bound sprint the disciplined choice.
Comparison fatigue compresses signal. Three discovery calls per week for three weeks produces nine partial impressions, none of them comparable. Every extra week widens variance in the founder’s memory faster than it narrows variance between partners.
Frontier-model cadence dates evidence. Anthropic shipped Opus 4 through Opus 4.8, OpenAI shipped GPT-5, and Google shipped Gemini 2.5 inside the last twelve months. A case study from a year ago references a different model than the founder will build against.
Bench capacity has a half-life. A strong partner with a six-week kickoff window in May may not have one in July. Stalled decisions push the eventual kickoff into the next bench gap. The cost of a slow choice is capital sitting unbuilt.
The rubric below assigns each day one deliverable, one time budget, and one watch-out. By Day 7 the founder owns a written record they can defend to a co-founder, an investor, or a board.
Day 1 — Shortlist criteria check
Deliverable: a 10-must-haves pass-fail matrix scored across all three shortlists. Time budget: 90 minutes. Watch-out: do not let any partner enter Day 2 with more than two “fail” rows — re-shortlist instead.
The 10 must-haves are AI-MVP-specific. Generic agency checklists fail here because they treat AI engagements as ordinary software engagements.
| # | Must-have | Pass criterion |
|---|---|---|
| 1 | Named eval workstream | Eval set, harness, judge calibration, and pass-rate threshold appear as line items in their standard SOW |
| 2 | Frontier-model contract | Contract names a primary model (GPT-5 / Opus 4.8 / Sonnet 4.6 / Gemini 2.5) and a fallback |
| 3 | IP-by-artifact-class | Code, prompts, eval suite, and any fine-tuned weights are assigned separately, not bundled |
| 4 | Six-scope-layer SOW | Feature shell, eval set, prompts, model contract, observability, on-call window |
| 5 | Inference pass-through | Frontier-model inference billed through founder’s vendor account, no markup |
| 6 | On-call SLA structure | Window length, severity definitions, response and fix-time SLAs in writing |
| 7 | Three current references | Three engagements inside the last six months, willing to take a phone call |
| 8 | Eval-protected case study | At least one prior engagement with a published pass-rate threshold and post-launch eval delta |
| 9 | Comparable budget band | At least one prior engagement inside your budget band (typically $90K–$250K) |
| 10 | Kill-clause precedent | Their standard MSA includes a milestone-based kill clause or accepts one in redline |
Score each partner across all ten rows in a single sitting. A partner failing 3+ rows is not a finalist. A partner failing 0–1 rows enters Day 2. Deeper criteria framing lives in how to do due diligence on an AI MVP partner’s prior work.
Day 2 — Discovery calls
Deliverable: one discovery scorecard per partner (three total). Time budget: three calls × 60 minutes + one hour of scoring = four hours. Watch-out: do not extend any call past 60 minutes — extension is a partner-pace signal, not a depth signal.
The discovery call is a structured audit, not a sales conversation. Score each partner across six rows: problem restatement, risk surfacing, preliminary effort range tightness, eval depth, model-contract explicitness, and on-call clarity. Score 0–2 per row, 0–12 per partner. The discovery call vs paid pilot spoke covers what a well-run call produces; the scorecard is the comparative instrument.
Two tests separate strong from polished. First, did the partner volunteer one risk you had not surfaced? A partner who only lists opportunities is selling. Second, was the preliminary effort range tight — a 50% band like $110K–$165K rather than a meaningless $50K–$500K? Tight bands signal a partner who has done this engagement shape before. If two partners tie at the top, that is information — Day 3 will separate them.
Day 3 — Reference calls
Deliverable: written notes from one reference call per partner. Time budget: three calls × 30 minutes + scribe time = three hours. Watch-out: refuse references older than 12 months — the 2025 model landscape is not the 2026 model landscape.
Ask each partner for three recent references the day after the discovery call. Pick the one most resembling your engagement shape — same budget band, same feature class, ideally inside six months. The Day 3 call is not a vibe check; it is a primary-source audit.
A nine-question script lives inside how to do due diligence on an AI MVP partner’s prior work. The four questions that separate strong partners on Day 3: did the eval suite catch a real regression before launch? was there a mid-engagement change order, and how was it priced? was the IP handoff clean, including eval set and prompt library? what would the reference do differently? Vague answers are disqualifying. References unwilling to discuss specifics on the phone are not references — they are testimonials.
Day 4 — SOW request + paid senior review
Deliverable: a written technical red-flag report from an independent senior engineer. Time budget: 30 minutes to brief, 4 hours of reviewer time, 30 minutes to read. Cost: $1,000–$3,000. Watch-out: do not skip this on the assumption that you can read SOWs well enough. You cannot — neither can your co-founder.
Ask all three partners for a draft SOW by Day 4 morning. Send all three to an independent senior engineer — fractional CTO, trusted ex-colleague, or paid reviewer — with a 30-minute brief on problem, budget, and timeline. The deliverable is a one-page red-flag report per SOW: scope ambiguities, eval workstream gaps, IP-by-artifact omissions, on-call SLA gaps, inference markup. The mechanics live in how to read an idea-to-product SOW and what to negotiate.
A $2,000 review on a $150,000 engagement is 1.3% of contract value and routinely surfaces $20,000–$50,000 of scope risk. Founders read SOWs for tone and price; operators read them for execution friction. Cross-check the report against AI MVP partner pricing red flags. A partner triggering two or more patterns is no longer a finalist.
Day 5 — Side-by-side weighted scoring
Deliverable: a ranked finalist set with a defensible written rationale. Time budget: two hours. Watch-out: build the matrix before looking at any partner’s score. Designing the matrix while filling it is how founders rationalize the partner they already prefer.
The matrix has five criteria, each weighted by impact on outcome. Score each partner 1–5 per row; multiply by weight; total per partner. A 75-point ceiling produces enough separation without false precision.
| Criterion | Weight | Why this weight |
|---|---|---|
| Eval discipline | 5 | Strongest single predictor of ship-at-quality |
| SOW clarity & IP terms | 4 | Strongest predictor of zero-surprise execution |
| Reference quality | 3 | Highest-fidelity signal of partner repeatability |
| Cost discipline | 2 | Matters, but third or fourth in importance |
| Cultural fit | 1 | Real, but routinely over-weighted |
The most common founder error is reversing this stack — cultural fit at 5, eval discipline at 1. The pattern is uniform downstream: a partner the founder enjoyed, a project that ships a demo, no eval suite, no production confidence, and a six-month tail of unbilled bug triage.
After scoring, the top partner should lead by at least 7 points on a 75-point scale. Tighter means Day 6 is genuinely a coin flip and the gut-check carries weight. A 15+ point gap means the decision is essentially made and Day 6 is confirmation.
Day 6 — Founder gut check and decide
Deliverable: a one-page decision memo naming the chosen partner and the runner-up. Time budget: two hours of solo writing. Watch-out: a gut check is not permission to overturn the scoring matrix on feelings. It is a structured cross-check, not a veto.
Three gut-check tests. First, the worst-day test: when the engagement hits its worst day — missed milestone, failed eval, hostile change-order — which partner do you trust to handle it like an adult? Second, the rollback test: if you part ways at milestone two, which partner produces the cleanest handoff? Third, the recommendation test: if a peer founder asked tomorrow which to hire, which would you name? Lean on SFAI Labs vs fractional CTO if a finalist is a different procurement shape.
If the gut check agrees with the matrix, the decision is made. If it disagrees, write the disagreement in plain language and re-read it the next morning. Overruling a matrix on feel is legitimate — but only after articulating the disagreement.
The memo names the chosen partner, the runner-up, the three top reasons in order, and the one risk being explicitly accepted. A named runner-up shortens the recovery loop from two weeks to two days when bench capacity or terms fail.
Day 7 — Sign and schedule kickoff
Deliverable: a countersigned SOW and a kickoff date within ten business days. Time budget: a two-hour redline call, then signature. Watch-out: do not sign without a kickoff date in the same document — a signed SOW with a TBD start date is a contract with an exit hatch built in.
By Day 7 the chosen partner has the Day 4 redline points scoped down to the two or three that materially affect outcome — eval workstream specificity, IP-by-artifact-class assignment, inference pass-through, on-call SLA. Most redlines resolve inside one 60-90 minute call. If the partner pushes back on more than one of the four, the runner-up gets the call instead.
The Day 7 contract has six items the founder owns: signed and dated SOW, kickoff date inside ten business days, named day-one engineer on a calendar invite, week-one milestone in the SOW, kill-clause language verified, and the eval-set pass-rate threshold written as a number. The discovery week method covers what happens between signature and kickoff. A field guide to evaluating an AI agency in under 90 minutes covers the first sift; the rubric above is the next register up.
Frequently asked questions
Is seven days actually enough to choose between three AI MVP partners?
Yes, provided the shortlist work is done. The rubric assumes three viable partners are already named — it does not budget time for shortlisting. Compression works because each day produces one deliverable, not because decisions are rushed. A founder who still needs to find three partners should budget an extra one to two weeks.
What if all three partners look equivalent after Day 5?
That is information. A 0–3 point gap on a 75-point scale means partners are genuinely similar, and the deciding factor is correctly the Day 6 gut check. A tight matrix rarely means more analysis would have separated them — usually soft criteria carry more weight than expected.
Should I run the discovery calls in parallel or sequentially?
Parallel — all three on Day 2. Sequential calls let the second partner adjust to what the first said and let the founder anchor on whoever spoke first. The exception is timezones: if one partner is 8+ hours offset, run two on Day 2 and the third early Day 3.
Do I really need to pay $1,000–$3,000 for a senior SOW review?
Yes, unless you are a senior engineer yourself. The review surfaces scope ambiguities, IP-by-artifact omissions, inference markup, and on-call gaps a non-technical founder cannot see — typically 10–25x the cost in surfaced risk on a $150K engagement.
What if the partner refuses to send a draft SOW by Day 4?
Strong partners produce a draft SOW inside 72 hours of a serious discovery call. A partner who needs more than four working days is over-booked, under-systematic, or routing the request through a sales funnel that does not prioritize your budget band. Any of those is a Day 4 disqualifier.
Can I add a paid pilot inside the seven-day window?
Generally no — a paid pilot is a one-to-two-week engagement and does not fit. Run the sprint, choose, then propose a pilot inside the chosen partner’s engagement if capability uncertainty remains high. The sprint decides which partner; the pilot tests whether stated capability holds against your data.
Is the scoring matrix defensible to investors or a board?
Yes — that is partly why it exists. A founder with a written matrix, a senior red-flag report, and a one-page decision memo is exercising procurement discipline that maps onto the diligence frame investors apply to founder judgment.
What if my budget is well under $90K — does this rubric still apply?
The cadence applies; the criteria compress. At a $30K–$60K budget the partner field is smaller, the SOWs are shorter, and the paid senior review compresses to a $500–$1,000 two-hour engagement. The 10 must-haves stay the same.
What if I picked wrong and want out at milestone two?
That is what the kill clause is for. A milestone-based kill clause — in the standard MSA or added in Day 7 redlines — lets the founder terminate at a defined milestone with no penalty beyond unpaid scope. The named runner-up becomes the recovery option.
Closing
The seven-day rubric is the cheapest defensible way to turn three viable shortlists into a signed contract and a kickoff date in one calendar week. Founders who stretch the same decision over four to six weeks usually arrive at the same choice — but with a later kickoff and a shallower paper trail.
A founder ready to apply the rubric this week can book a discovery call with SFAI Labs as one of the three. The 60-minute call follows the Day 2 structure and produces the same scorecard inputs every other partner should produce.
Arthur Wandzel
