A founder finishes a 45-minute vendor pitch with a one-page “investment summary” in hand: $185K, six lines, two of them labelled “engineering” and “AI integration.” Three weeks later the same founder is at a whiteboard back-deriving what on-call coverage, a prompt registry, and a model-upgrade re-eval would have cost — because the vendor said all three were “in scope” and now they are change orders. The summary was not wrong. It was too compressed. Six lines hide where AI MVP money goes; eleven lines do not. This piece names the eleven, sizes each against 2026 rates, lists who owns each, and gives the one sentence that defends every line when finance asks “can we cut this?”
This worksheet builds on the AI MVP economics playbook, itself a chapter of the idea-to-product manifesto. The playbook sets the principle. The eleven-line worksheet below is the artifact a founder uses against a real vendor quote on a real Tuesday.
Why 11 lines, not 5 or 7
A defensible AI MVP budget has five to eleven lines depending on what the founder needs it to do. Five lines decomposes a quote for a board meeting; seven lines defends it against CFO follow-ups. Eleven lines is what the founder needs to plan the budget before any vendor sees it.
The four marginal lines — model-upgrade buffer, prompt registry, on-call coverage, contingency — are the ones a vendor quote almost never breaks out. They get rolled into “engineering” or “support” and surface as change orders in months two through four. McKinsey’s State of AI has tracked across two cohorts that roughly four in five AI pilots stall before production. The stall point is rarely the model — it is the unbudgeted work. Eleven lines names that work before it becomes a change order.
The 11-line worksheet
| # | Line | 2026 range | % of $150K | Who owns | Skip-cost |
|---|---|---|---|---|---|
| 1 | Idea validation + ICP interviews | $4K-$10K | 3-7% | Founder + paired senior eng | PRD targets the wrong workflow; build retargets in week 4 |
| 2 | PRD with eval contract | $8K-$18K | 5-12% | Senior AI eng + founder | Eval suite in line 5 has nothing to grade against |
| 3 | Architecture pick (prompt vs RAG vs agent) | $5K-$12K | 3-8% | Tech lead | Build cost balloons 2-5x at the wrong architecture |
| 4 | Eval set curation + rubric | $12K-$25K | 8-17% | Senior eng + domain expert | Demos pass, production fails on the long tail |
| 5 | Eval harness engineering | $18K-$35K | 12-23% | Senior AI eng | No regression signal when models update |
| 6 | Capability build | $40K-$80K | 27-53% | 2 senior engineers | This is the visible work — under-funded only by under-decomposing it |
| 7 | UI / API surface | $8K-$20K | 5-13% | Frontend or platform eng | Capability ships behind a CLI; demo days slip |
| 8 | Infra band (inference + retrieval + observability) | $3K-$12K (90-day MVP slice) | 2-8% | DevOps / platform | Quote a point; CFO discovers the band in month 3 |
| 9 | Model-upgrade re-eval buffer | $6K-$14K | 4-9% | Senior AI eng | First model alias update breaks the demo; team scrambles |
| 10 | Hardening + on-call coverage (first 90 days) | $12K-$28K | 8-19% | Senior eng (rotated) | First production failure has no on-call; reputation cost is uncapped |
| 11 | Contingency (10-15% of the build cost) | $10K-$22K | 7-15% | Founder, governed by change-order rule | First scope amendment gets paid out of the wrong budget |
| Defensible total | $126K-$276K | 100% at $150K worked example |
Every column is load-bearing. The dollar range is the 80% band against 2026 senior-AI-engineering rates ($200-$300/hr blended). The percentage anchors to a $150K total — the modal MVP in C3’s sample. “Who owns” tests whether your team has the role internally or has to source it. The skip-cost is the sentence the founder repeats back to a board member or vendor PM asking why the line exists.
Lines 1-3: scoping and capability definition
Line 1 — Idea validation + ICP interviews ($4K-$10K, 3-7%). Twelve to twenty buyer-persona conversations in week zero. Output is a one-page ICP brief naming the workflow, the trigger event, and three observable success signals. Who owns: founder paired with a senior AI engineer. Skip-cost: the PRD targets a workflow the buyer does not actually do on Tuesdays.
Line 2 — PRD with eval contract ($8K-$18K, 5-12%). A 6-10 page PRD with an appendix naming the eval contract — test set, rubric per capability, threshold the build must clear. Sounds like QA, is actually scope. Who owns: senior AI engineer drafting with the founder; the domain expert reviews the rubric. Skip-cost: line 4 has nothing to grade against; the build ships against an eyeball test that demos well and fails in production.
Line 3 — Architecture pick ($5K-$12K, 3-8%). A 1-2 week spike naming the build path: prompt-only, retrieval-augmented (RAG), agentic, or fine-tuned-model. The four paths have three-to-fivefold cost deltas — a prompt-only build at $30K, the agentic equivalent at $90K. Who owns: tech lead with the senior AI engineer who will own the build. Skip-cost: line 6 starts on the wrong substrate and rewrites in week three at a 2-5x multiplier.
Lines 4-7: build and eval
Line 4 — Eval set curation + rubric ($12K-$25K, 8-17%). 200-800 representative inputs against the buyer’s workload distribution, annotated by a senior partner or domain expert at $150-$250/hr. Who owns: senior engineer pairing with the domain expert. Skip-cost: the harness grades cherry-picked inputs that fail on the long tail.
Line 5 — Eval harness engineering ($18K-$35K, 12-23%). Harness that runs the eval set on every prompt change, model alias update, and retrieval-index refresh — regression dashboards, threshold alerts, CI gating. Who owns: senior AI engineering at $200-$300/hr for 4-8 calendar weeks overlapping the build. Skip-cost: changes ship without a regression signal; the team discovers regressions via support tickets.
Line 6 — Capability build ($40K-$80K, 27-53%). The visible engineering — prompt scaffolding, retrieval, agent loops, API integration, basic auth. Two senior engineers at $225-$300/hr blended for 6-10 calendar weeks. Who owns: engineering. Skip-cost: the common failure is over-budgeting this line by rolling lines 4, 5, 9, and 10 into it and calling the bundle “engineering.”
Line 7 — UI / API surface ($8K-$20K, 5-13%). Web app, Slack bot, REST API, or admin console. Often deferred on “we’ll just ship a CLI” — a deferral that holds for 60 days then becomes the reason the demo slips. Who owns: frontend or platform engineer in parallel with the build. Skip-cost: first three stakeholder demos require a senior engineer at a terminal.
Lines 8-11: where the post-launch overruns hide
These four lines are the ones competitor worksheets compress into “infra,” “support,” or “contingency” and the ones that surface as change orders in months two through four.
Line 8 — Infra band ($3K-$12K, 2-8%). Inference, retrieval index hosting, vector store, observability — for the 90-day MVP slice. Founders quote it as a point; finance pushes back as a band. Cost varies 4-5x by architecture: prompt-only at 5K queries/month at the low end; agentic RAG at 50K queries/month at the high end. The cost-per-query framework is how to band this line. Who owns: DevOps or platform engineering. Skip-cost: CFO discovers the band on the month-three invoice.
Line 9 — Model-upgrade re-eval buffer ($6K-$14K, 4-9%). Frontier model labs ship breaking-quality changes multiple times per year — an alias update can shift eval scores 5-15 percentage points without warning. This line funds the hours to re-run the suite, triage regressions, and decide whether to pin or migrate. Who owns: a senior AI engineer retained for triage windows. Skip-cost: the first alias update after launch breaks a capability mid-demo; the team scrambles with zero budget for the work.
Line 10 — Hardening + on-call, first 90 days ($12K-$28K, 8-19%). Launch to steady-state — prompt registry, observability hooks, error-budget definition, and a senior engineer rotated onto a part-time on-call. Not a junior support hire. Who owns: a senior engineer rotated from the build team. Skip-cost: first production incident hits with no on-call; the team triages over Slack DMs at 11pm; the cost is credibility, not just dollars. The TCO seven-line decomposition shows how an under-funded on-call line cascades into year-1 TCO.
Line 11 — Contingency, 10-15% of build cost ($10K-$22K, 7-15%). The line the founder controls under an explicit change-order rule: any line that runs over by ≥10%, or any scope amendment, draws down here — never silently absorbed elsewhere. Who owns: the founder, with a one-page change-order log. Skip-cost: the first scope amendment gets paid from eval or hardening, which then run short and re-emerge as a different overrun three weeks later.
Worked example: a defensible $150K AI MVP
A retrieval-augmented internal copilot for a 35-person professional-services firm. ICP is internal — partners and senior associates. Eval set is 400 question-answer pairs annotated by two partners. Build is 8 calendar weeks, 2 senior AI engineers, surface is a web app + Slack bot.
| # | Line | Picked figure | % of $150K |
|---|---|---|---|
| 1 | Idea validation + ICP interviews | $6K | 4% |
| 2 | PRD with eval contract | $11K | 7% |
| 3 | Architecture pick (RAG, picked over prompt-only) | $7K | 5% |
| 4 | Eval set curation + rubric (400 pairs, 2 partner-annotators) | $18K | 12% |
| 5 | Eval harness engineering | $22K | 15% |
| 6 | Capability build (RAG over Confluence + Notion + Drive) | $48K | 32% |
| 7 | UI surface (web app + Slack bot) | $12K | 8% |
| 8 | Infra band (90-day slice, 8K queries/mo) | $6K | 4% |
| 9 | Model-upgrade re-eval buffer | $7K | 5% |
| 10 | Hardening + on-call, first 90 days | $15K | 10% |
| 11 | Contingency (12% of build cost) | $8K | 5% |
| Total | $160K | 107% |
The worked total is $160K against a $150K headline — a 7% gap closed by (a) deferring the Slack bot to a fast follow ($4K), (b) cutting the eval set from 400 to 300 pairs ($5K), or (c) renegotiating contingency from 12% to 8% ($3K). Which line gets cut is now a defensible conversation — three named trade-offs, not silent vendor compression.
How to use the worksheet against a vendor quote
Lay the vendor’s quote side by side with the eleven lines. For every compressed line, ask one of two questions:
- “Is this in scope, and if so, which line of your quote pays for it?” — for lines 4, 5, 9, 10, 11.
- “If it is not in scope, what is the change-order rate when it surfaces?” — same five lines.
Vendors who answer both in twelve seconds per line have priced the post-launch work explicitly. Vendors who cannot have a six-line quote against eleven-line work — the difference gets paid in change orders, on-call gaps, or a stalled pilot.
The worksheet runs in reverse for internal teams: hand the eleven lines to engineering leadership before any sprint plan. The year-1 cost curve extends the build window into a 12-month run cost; the $250K-buys piece shows how the eleven lines reshape at a higher total.
Frequently asked questions
Why eleven lines and not nine or thirteen?
Eleven is the minimum that names the four post-launch lines (8-11) separately. Nine collapses the model-upgrade buffer into hardening, and the founder loses the lever for a model-alias-driven re-eval conversation. Thirteen over-decomposes the build for a CFO-facing artifact. Eleven is the count where every line is independently defensible and no line is purely tactical.
What if my MVP total is $80K, not $150K?
The eleven lines compress proportionally — every line shrinks except lines 2 and 3, which stay roughly fixed because they are scoping work, not engineering work. Compression usually shows up as a smaller eval set in line 4 (200 pairs instead of 400), a single-engineer build in line 6, and a deferred surface in line 7. Lines 9, 10, and 11 should never compress to zero — that is the failure mode this worksheet exists to prevent.
What if the vendor refuses to break their quote into eleven lines?
That is the signal. A vendor who decomposes into eleven lines has priced the post-launch work explicitly. A vendor who insists on six lines is either (a) genuinely scoping all eleven lines and absorbing the risk in margin — in which case the quote runs 20-30% above the decomposed version — or (b) under-scoping the four post-launch lines and planning to recover via change orders. The eleven-line decomposition distinguishes (a) from (b) before signing.
How does this differ from a generic SaaS MVP cost worksheet?
A SaaS MVP worksheet has four to six lines — design, development, QA, deployment. QA runs at 5-10% of total. None of lines 8-11 exist as a separate category. A 2026 AI MVP worksheet moves QA to 20-30% as eval engineering (lines 4-5), adds a discrete hardening + on-call line (10), bands infra (8) instead of point-estimating it, and treats the model-upgrade buffer (9) as a known expense rather than a contingency event. The structural delta is 30-80% on the total at comparable scope.
Should the contingency line ever exceed 15%?
Rarely. Above 15%, the founder is hedging in contingency rather than fixing an under-decomposed line elsewhere. The exception is a regulated industry MVP (healthcare, finance, legal) where compliance review can compress unpredictably — there contingency can defensibly land at 18-22%, with the regulatory uncertainty named explicitly as the reason.
Can the founder own the eval set annotation?
For the first 100-200 pairs, often yes — it is one of the highest-return uses of founder hours because the annotation surfaces rubric ambiguities the PRD missed. Past 200 pairs, founder annotation becomes a bottleneck. The hand-off point is roughly when the rubric stops changing weekly.
What is the right ratio of build (line 6) to eval (lines 4+5)?
Roughly 1.0 to 1.4 — eval work lands at 60-80% of the capability build cost on a 2026 frontier-model MVP. A 4:1 build-to-eval ratio is a 2018 SaaS worksheet, not a 2026 AI MVP worksheet, and vendors who quote that ratio are usually under-funding eval and recovering margin via change orders in months 3-6.
Where does the infra band actually land?
$3K-$12K for the 90-day MVP window. The 4x range collapses by architecture — prompt-only at 5K queries/month lands at $3K-$5K; agentic RAG at 50K queries/month with a managed vector store and observability lands at $10K-$12K.
Key takeaways
- Eleven lines is the planning artifact founders use before a vendor quote arrives; five and seven-line versions defend a quote after the fact.
- Lines 8-11 (infra band, model-upgrade buffer, hardening + on-call, contingency) are where post-launch overruns hide. Name them explicitly or pay them as change orders.
- The skip-cost column is the sentence a founder repeats back to anyone asking why a line exists.
- Anchor every percentage to a worked total — $150K here — so the percentages are decisions, not decoration.
- A vendor who decomposes into eleven lines has priced post-launch work explicitly. A vendor who insists on six lines has not.
- The eleven lines are the format. The dollar figures are 2026 heuristics — validate against your own vendor proposals, regulatory context, and finance model before signing.
Arthur Wandzel
