A founder walks into a board meeting with a $150K AI MVP quote. The CFO asks a single question — “what is that money buying us, line by line?” — and the meeting ends three minutes later with the budget shelved for “more research.” The quote was not wrong. The decomposition behind it was missing. Most AI MVP proposals in 2026 still arrive with a 2018 SaaS shape: one line for design, one for development, one for QA, and a vague “AI integration” bucket. That shape no longer maps to where the money goes on a frontier-model build. The result is a budget that cannot survive a CFO conversation. This piece names the five lines a defensible 2026 AI MVP budget runs on, sizes each, and shows what skipping each line costs downstream.
This article extends the AI MVP economics playbook, the definitive guide to AI MVP build economics, itself a chapter of the idea-to-product manifesto. The economics playbook sets the principle. The five-line decomposition below is the operational form a founder uses to defend that budget in a board meeting.
Why the 2018 SaaS cost template fails for a 2026 AI MVP
Open most AI MVP proposals in 2026 and the budget shape is inherited: design, development, QA, deployment, sometimes “AI integration” as a separate bucket. That template was calibrated against a 2018 SaaS build — deterministic behavior, stable runtime substrate, one-time QA pass. None of that holds for a frontier-model build. Behavior shifts when a model alias updates. Quality is enforced by a graded eval suite, not a green CI check. The runtime substrate ships breaking-quality changes multiple times per year. QA is continuous, not a phase.
The result is a structural under-budgeting pattern. McKinsey’s State of AI reporting has documented for two years that roughly 80–85% of AI pilots stall before reaching production at scale. The stalled pilots are not failing because the engineering was bad. They are failing because the budget did not fund the lines the production build actually requires — the eval suite was never funded, the on-call period was never named, and infra was point-estimated when it should have been banded.
The five lines below replace the 2018 template with one that maps to where the money goes on a 2026 AI MVP. Each line has a 2026 dollar range, a percentage of total budget, and a skip-cost. The skip-cost is the half that makes the line defensible to a CFO.
Line 1: Planning + PRD
What it is. The week (sometimes two) before any engineer writes code. Customer interviews to validate the ICP. A one-page brief that names the workflow being automated and the three to five capabilities that minimum-prove the idea. A 6–10 page PRD with an eval-contract appendix — the test set the feature will be graded against, the rubric for each capability, and the threshold the build must clear.
2026 range. $12K–$25K. 25–55 hours of paired founder-engineer work at a blended $200–$350 per hour, over 5–10 calendar days.
% of budget. 8–12% on a $150K–$250K AI MVP.
What skipping costs. The downstream budget gets set against a feature pitch instead of a capability list. The eval set in Line 2 has nothing to anchor against. Build hours in Line 3 are estimated against a vague target, then balloon when “what does correct look like?” surfaces in week 4 instead of week 0. Founders who skip Line 1 typically discover the cost as a 15–30% overrun on Line 3.
Line 2: Eval engineering
What it is. Building, curating, and running the eval suite that proves the MVP ships at quality. A test set of 200–800 representative inputs against the buyer’s actual workload distribution. A rubric that turns each capability into a graded sub-criterion. An eval harness that runs on every prompt change and model alias update. Annotation work by a senior partner or domain expert — not a scale-labeler.
2026 range. $35K–$75K. Senior engineering at $200–$300 per hour for harness work, domain-expert annotation at $150–$250 per hour, 4–8 calendar weeks overlapping with the build.
% of budget. 20–30%.
What skipping costs. The build ships against an eyeball test instead of a graded eval. Demos pass against cherry-picked inputs. Production hits the workload distribution and surfaces failure modes the team never tested. The fix is a six-to-twelve week post-launch eval-engineering retrofit at 1.5–2x the cost of building the suite during the MVP — because the team is fighting live regressions while building the suite they should have built upfront. The decoding AI project TCO piece lays out the seven recurring TCO lines that an under-funded eval suite turns into scope creep.
Line 3: Build
What it is. Shipping the capability — prompt scaffolding, retrieval, agent loops, API integration, the user-facing surface (web, Slack, API), basic auth and access controls. Senior AI engineers paired with a frontend or platform engineer. The build path branches sharply on architecture — prompt-only, retrieval-augmented, agentic, and fine-tuned-model are four genuinely different paths with three-to-five-fold cost ranges between them.
2026 range. $45K–$90K. Two senior engineers at $225–$300 per hour blended for 6–10 calendar weeks. The range collapses or expands sharply on the architecture decision in Line 1.
% of budget. 30–40%. Still the largest line, but no longer dominant the way “engineering” used to be on a 2018 SaaS budget.
What skipping costs. This line cannot be skipped, but it can be under-funded. The common failure: estimating against a prompt-only architecture when the capability requires retrieval or agentic loops. Discovery in week 3 or 4, the architecture decision gets re-litigated, the $50K estimate becomes a $90K reality. Stack Overflow’s 2024 Developer Survey reported 76% of developers using or planning to use AI tools while only 43% trusted their accuracy — the gap closes when the architecture decision is made deliberately, not by default.
Line 4: Infra
What it is. The runtime substrate. Inference cost on the chosen model tier (Claude Opus 4.8, GPT-5, Gemini 2.5 Pro, or an open-weight model on a managed inference provider). Vector store and embeddings infrastructure if retrieval-augmented. Observability — trace storage, eval run history, prompt-and-completion logs — sized for 90–180 days of retention. Application hosting on a standard cloud. Auth, secrets, per-customer isolation if multi-tenant.
2026 range. $8K–$25K for the 6–12 week build window, then $1.5K–$8K per month recurring. The recurring number is a band, not a point estimate — inference cost varies 2x in either direction within 9 months as usage grows, token prices fall, or the team upgrades to a higher tier.
% of budget. 6–12% of the build phase. Recurring infra sits in the 24-month TCO.
What skipping costs. Two failure modes. Point-estimating inference — the founder budgets $2K per month, usage triples in six months, the bill is $7K, and the CFO discovers a 3.5x miss the team did not flag. Skipping observability — the team logs nothing structured, then cannot debug the first production regression in week 10, then ships a fix-by-guessing that compounds the issue. The day-one observability stack piece covers tools and retention policies that survive real production load.
Line 5: Hardening + on-call
What it is. The 30–60 days after the MVP ships. Senior engineering coverage during business hours for production regressions. A weekly eval re-run with triage of regressed cases. A 14-day SLA on responding to frontier model alias updates. Documentation and a handoff runbook for whoever takes over operations after the on-call window closes.
2026 range. $15K–$40K. 20–40 hours per week of senior engineering at $225–$300 per hour plus a small reserve for surge incidents.
% of budget. 10–15%. Small relative to the build, disproportionately important — the bridge between “the demo passed” and “the system runs.”
What skipping costs. The first regression in week 10 has no owner. The founder discovers it from a customer email, files it as a bug, and waits five days for a triage that should have happened in five hours. The customer trust hit is permanent. The fix costs $5K; the relationship damage is the real cost. Founders who skip Line 5 often pay 2–3x its size in emergency engineering plus customer-success time within the first 60 days post-launch. The hidden costs of AI development piece covers operational categories most agency proposals omit.
The one-page worksheet
Founders should be able to pull this worksheet out of the article, fill in the dollar amounts against a specific vendor proposal, and present it to a CFO in 15 minutes.
| Line | Name | 2026 range | % of budget | This proposal | Defensible (Y/N) |
|---|---|---|---|---|---|
| 1 | Planning + PRD | $12K–$25K | 8–12% | $___ | ___ |
| 2 | Eval engineering | $35K–$75K | 20–30% | $___ | ___ |
| 3 | Build | $45K–$90K | 30–40% | $___ | ___ |
| 4 | Infra (build window) | $8K–$25K | 6–12% | $___ | ___ |
| 5 | Hardening + on-call | $15K–$40K | 10–15% | $___ | ___ |
| Total | $115K–$255K | 100% | $___ |
How to use it.
- Match proposal lines to the five names. If a proposal has no separate eval engineering line, ask where eval work is funded; if the answer is “rolled into QA,” the build is structurally under-funded.
- Check each line against the 2026 range. Lines below the range need a defense — what is the proposal assuming that lets it under-spend the market?
- Ask the skip-cost question per line. “If this line were halved, what would the downstream cost be?” Vendors who cannot answer have not modeled the trade-offs.
- Total and compare. A proposal at $80K either omits a line or under-funds eval engineering by 50%-plus. A proposal at $400K needs a named reason for the markup.
Worked example: defending a $150K AI MVP to a CFO
A founder is selling a $150K AI MVP to her board’s CFO. The CFO asks the line-by-line question. The script:
| Line | Allocated | Defense |
|---|---|---|
| Planning + PRD | $18K | ”Five days of paired founder-engineer work. Customer interviews against six target accounts. A 6–8 page PRD with an eval-contract appendix — the document we hold the build accountable to.” |
| Eval engineering | $45K | ”Builds the test set we grade the MVP against. 400 representative inputs from our buyer’s workload, rubric-graded by our domain expert at $200 per hour for 80 hours, plus the harness work by our senior engineer at $275 per hour for 100 hours. Skipping this is the McKinsey 80% pilot-stall pattern.” |
| Build | $55K | ”Two senior engineers for seven weeks at $250 per hour blended. Prompt scaffolding, retrieval against our knowledge base, Slack integration, admin surface. Retrieval-augmented architecture, not agentic — keeps this line at mid-range.” |
| Infra (build window) | $12K | ”Inference and observability for the 8-week build. Claude Sonnet 4.6 plus Claude Opus 4.8 for harder cases. Pinecone vector store. LangSmith trace storage with 120-day retention. Recurring post-launch cost sits in the 24-month TCO at $2K–$5K monthly.” |
| Hardening + on-call | $20K | ”30-day on-call window. 30 hours per week of senior coverage at $250 per hour plus a surge reserve. Covers the first model-alias update if a frontier provider ships one inside the window — eval suite re-runs within 14 days.” |
| Total | $150K | All five lines named with a number and a skip-cost. The CFO can push back on a specific line but cannot push back on the budget shape. |
The conversation that follows is structurally different from a vendor-quote conversation. The CFO is now negotiating which lines to fund and at what depth — not whether the project is real.
From the 5 lines to a 24-month TCO
The worksheet covers the MVP budget — going from a signed PRD to a launched system. It does not cover the 24-month TCO, which adds four recurring lines: ongoing inference and observability infra (post-launch monthly band), a maintenance retainer for eval suite updates, model-upgrade re-evals as frontier providers ship new versions, and prompt registry maintenance.
The what does $50K, $100K, $250K buy you in an AI MVP piece walks through three reference budgets and what each unlocks. The how much does an AI MVP cost in 2026 piece gives the broader market range. The decoding AI project TCO piece extends the five MVP lines into the seven recurring TCO categories CFOs typically miss across 24 months.
Frequently asked questions
Why five lines and not seven, or three?
Five maps cleanly to the work a 2026 AI MVP does in a 6–12 week window. Three (plan, build, ship) under-decomposes — it hides eval engineering inside build and hides hardening inside ship, which is exactly the pattern that produces the 30%-overrun board meeting. Seven over-decomposes for an MVP budget — the additional lines (model-upgrade re-eval, prompt registry, regression triage) are post-launch recurring categories that belong in the 24-month TCO decomposition. Five is the minimum that lets a founder defend the build budget without conflating phases.
How does this differ from a generic SaaS MVP cost breakdown?
A SaaS MVP breakdown has four shapes — design, development, QA, deployment — with QA at 5–10% of budget and no eval engineering line. A 2026 AI MVP breakdown moves QA to 20–30% as eval engineering, adds a discrete hardening + on-call line, and treats infra as a band rather than a fixed monthly figure. A comparable-scope SaaS MVP runs $80K–$160K in 2026; an AI MVP runs $115K–$255K. The 30–80% delta is the cost of the eval suite, the model-upgrade buffer, and the hardening window.
What if my vendor’s proposal has only three lines?
Ask the vendor to decompose into the five lines above before you sign. A three-line proposal is not necessarily bad — it may be a presentation convention. But if the vendor cannot produce a five-line decomposition on request, the proposal is likely under-funding eval engineering or hardening, and the gap will surface as scope creep in week 8 or 12.
Can a $50K or $80K AI MVP be defensible against this worksheet?
Sometimes, with structural choices. A $50K AI MVP is defensible if the architecture is prompt-only (no retrieval, no agents), the integration surface is one channel, the eval set is tight (150–250 inputs), and the hardening window is 14 days. The trade-off is a narrower capability surface and a thinner on-call buffer. Below $50K, the math typically breaks at 2026 senior engineering rates.
How do I know if eval engineering is genuinely funded or just labeled?
Three checks. First, ask for the eval test set size and annotation methodology — “400 inputs, domain-expert graded at $200 per hour” is fundable; “we will build evals during the project” is not. Second, ask which engineer owns the eval harness — a build engineer who also “does evals on the side” rarely produces a defensible suite. Third, ask for the eval pass threshold the build must clear before launch — “85% pass on the test set with no critical-class failures” is a real threshold; “evals look good” is not.
How should infra be presented to a CFO if it is a band, not a point estimate?
Present build-window infra as a fixed number because the 6–12 week window is short enough that variance is bounded. Present post-launch recurring infra as a band tied to a named usage projection — “$1.5K–$8K per month, depending on whether we hit 50, 200, or 500 monthly active users by month three, with quarterly review against actuals.” CFOs accept bands tied to named external inputs. They reject bands that read as “we do not know.”
What primary sources back the dollar ranges in this article?
Three inputs. The Stack Overflow Developer Survey 2024 for senior engineering rates — US-based seniors in the 7–10 year band sit at $150K–$220K base, mapping to $200–$300 per hour blended on an agency engagement. Public pricing pages from Anthropic, OpenAI, and Google for inference cost ranges. Artificial Analysis for frontier model release cadence, which informs the model-alias update buffer in Line 5.
How does this connect to the broader idea-to-product manifesto?
The five-line worksheet is the operational form of the manifesto’s claim that non-engineer founders ship AI products by funding evaluation as the spec, not features as the spec. Line 1 produces the eval contract. Line 2 builds the suite. Lines 3, 4, and 5 execute against the contract. A founder who reads the manifesto and accepts the principle but cannot defend a five-line budget against a CFO has only half of what the manifesto promises.
What is the most common founder failure pattern when defending an AI MVP budget?
Two failures, one upstream of the other. Presenting a single-number quote (“$150K”) instead of a decomposition — the single number invites the CFO question that ends the meeting. Presenting a decomposition shaped like a 2018 SaaS MVP instead of the 2026 shape — it passes a casual read but fails a CFO read because it does not name eval engineering or hardening as discrete lines. The fix is to arrive with the worksheet filled in, skip-cost articulated per line, and a willingness to negotiate which lines to fund at depth.
Key takeaways
- A defensible 2026 AI MVP budget decomposes into five lines: planning + PRD ($12K–$25K), eval engineering ($35K–$75K), build ($45K–$90K), infra during build window ($8K–$25K), hardening + on-call ($15K–$40K). Total band: $115K–$255K.
- The 2018 SaaS template — design, development, QA, deployment — fails for AI MVPs because it bundles eval engineering inside QA (understating it 4–6x) and omits hardening entirely. That structural omission is the dominant pattern behind the 80%-plus pilot stall rate McKinsey reports.
- Each line has a skip-cost. Skipping Line 1 produces a 15–30% Line 3 overrun. Skipping Line 2 produces a six-to-twelve week post-launch retrofit at 1.5–2x cost. Skipping Line 4’s observability blinds the team on the first regression. Skipping Line 5 makes customer trust the first cost paid.
- Founders defending a budget should arrive with the worksheet filled in against a specific proposal, each line named with a number and a skip-cost. The CFO conversation shifts from “is this real?” to “which lines fund at what depth?”
- The five MVP lines extend into the seven recurring TCO lines on a 24-month horizon. A founder who has defended the MVP budget should expect to defend the TCO decomposition at the next board cycle.
Pulling a quote into a five-line worksheet is the cheapest single thing a founder can do to convert a vendor proposal into a board-defensible budget. It is also the move most often skipped, which is why so many AI MVP projects die in the budget conversation rather than the build.
Want the editable version? Download the AI MVP Scoping Worksheet — the five lines pre-formatted, with skip-cost prompts and a CFO defense column ready for your next board meeting.
Arthur Wandzel
