A founder sends a vendor a $75K budget and gets back a proposal that reads “AI MVP — fixed price, 6 weeks, $75K.” Two pages, three milestones, no line items. The founder cannot tell whether the quote is an honest lean build or a $150K project with five lines silently removed. The dollar amount is not the issue. The decomposition is. A defensible $75K AI MVP exists in 2026, but only against a narrow scope, a small senior team, and an exclusion list written down. This piece names the 12 lines that sum to ~$75K for a tight 6-week build, identifies what the budget does NOT cover, and gives the four diagnostics that separate an honest $75K from a vendor cutting corners.
This is a spoke under the AI MVP economics playbook, the cluster anchor for AI MVP build economics. The playbook sets the 6–12 week economics; this article zooms into the lean-defensible bracket where the dollar discipline gets tightest. It lives inside the idea-to-product manifesto, the master guide for non-engineer founders shipping AI products in 2026.
Why $75K is a real bracket, not a vanity number
$75K is the price point at which a senior, frontier-model AI build becomes possible — but only at a strict scope. Below it, the team shape collapses to one engineer alone, the eval contract becomes informal, and the deliverable resembles a prototype more than a defensible MVP. Above it, the budget starts funding a second integration, a longer on-call window, or a dedicated eval engineer. McKinsey’s State of AI reporting has documented for two years that roughly 80–85% of AI pilots stall before reaching production scale. The $75K bracket is engineered to confirm or kill an idea within six weeks against a real eval set — not to ship a polished product to enterprise customers.
The companion piece on what $50K, $100K, $250K buy you in an AI MVP places $75K between the single-capability proof and the eval-bound build. This article zooms into the $75K bracket itself: a defensible lean MVP with one capability, one integration surface, a senior engineer plus a fractional eval partner, and a 6-week elapsed window.
The scope this $75K budget assumes
Before any line item makes sense, the scope has to be named. A $75K AI MVP at 2026 market rates assumes:
| Scope dimension | $75K assumption |
|---|---|
| Capabilities | One — a single AI-driven workflow step end-to-end |
| Integration surfaces | One — a single API, document store, or app surface |
| Build path | Prompt-engineered + light retrieval-augmented grounding |
| Eval set | 100–200 founder-curated representative inputs, graded against a written rubric |
| UI | Minimal — a Next.js page or thin web app; basic auth optional |
| Team | One senior AI engineer (50–60% allocated) + fractional eval partner (10–20%) + founder co-author |
| Window | 6 calendar weeks elapsed, including handoff |
| Handoff | Runbook, prompt files, eval CSV, README; no on-call after week 6 |
If the scope drifts from these assumptions — two capabilities, two integrations, multi-tenant auth, HIPAA evidence — the budget does not stay at $75K. The most common cause of a $75K project shipping at $130K is unspoken scope drift in week 2 that the proposal never repriced.
The 12-line decomposition
The full $75K, in 12 named lines:
| # | Line | Who | $ amount |
|---|---|---|---|
| 1 | Discovery + scoping workshop | Founder + senior engineer | $4,500 |
| 2 | PRD + eval contract | Senior engineer (lead author) | $7,500 |
| 3 | Architecture decision record | Senior engineer | $2,500 |
| 4 | Eval set curation | Founder + fractional eval partner | $6,000 |
| 5 | Eval harness | Fractional eval partner | $5,500 |
| 6 | Prompt scaffolding + retrieval | Senior engineer | $14,000 |
| 7 | Integration layer | Senior engineer | $8,500 |
| 8 | Minimal UI | Senior engineer (frontend hat) | $6,000 |
| 9 | Eval grading + iteration loop | Senior engineer + fractional eval partner | $7,500 |
| 10 | Observability + logging | Senior engineer | $3,500 |
| 11 | Handoff package | Senior engineer | $3,500 |
| 12 | Inference + infra (6-week build window) | Vendor pass-through | $5,800 |
| Total | $74,800 |
Each line has a deliverable a founder can point at. A vendor quote that does not produce the artifact behind each line — the PRD, the eval rubric, the architecture decision record, the eval CSV with grades, the runbook — has cut the line, not delivered it. The next section unpacks what each line actually buys.
Line-by-line: where the money actually goes
Line 1 — Discovery + scoping workshop ($4,500)
Two half-day working sessions between the founder and the senior engineer. The deliverable is a 2-page scoping memo naming the one capability the MVP will prove, the user-action it replaces, the success criterion, and the explicit exclusion list. Vendors most often skip this and replace it with a 30-minute kickoff call. Twenty hours of senior time at a blended $225 per hour.
Line 2 — PRD + eval contract ($7,500)
A 6–8 page PRD with an eval-contract appendix. The PRD names the capability, the workflow, the input/output contract, and the rubric the build will be graded against. The eval contract is what distinguishes a defensible MVP from a demo. Roughly 30 hours of senior engineering at $225–$250 per hour. If the PRD is two pages of marketing language with no eval rubric, this line has been cut.
Line 3 — Architecture decision record ($2,500)
One page naming the build path — prompt-only, retrieval-augmented, agentic, or fine-tuned — with trade-offs and cost band of each. A $75K MVP almost always lands on prompt-engineered or thin retrieval-augmented. The decision matters less than the fact that it was made deliberately. Ten hours of senior time.
Line 4 — Eval set curation ($6,000)
100–200 representative inputs sampled against the buyer’s actual workload distribution, not cherry-picked. The founder owns the sampling; only the founder knows what “representative” means for their workflow. The fractional eval partner builds the spreadsheet, the rubric, and the grading scaffold. 25 hours of partnered work at a blended $240 per hour.
Line 5 — Eval harness ($5,500)
The runnable script that loads the eval set, calls the build, captures outputs, and writes graded results to a CSV. At the $75K bracket, the harness is intentionally light — no LLM-as-judge loop, no CI gate, no regression suite across model aliases. 22 hours of senior eval-engineering time. The stop-budgeting-AI-projects-in-story-points piece explains why eval-runs is the unit a 2026 build should be sized in.
Line 6 — Prompt scaffolding + retrieval ($14,000)
The largest single line. Prompt design, retrieval index construction (if applicable), structured-output schemas, error handling for the model layer. Senior AI engineering at $250 per hour for roughly 55 hours, spread across weeks 2–5.
Line 7 — Integration layer ($8,500)
The single API, document store, or app surface the MVP talks to. Authentication, request/response shaping, error handling, rate limiting. Senior engineering at $225 per hour for roughly 38 hours. If the scope drifted to two integrations in week 2, this line doubles — and the budget should be repriced.
Line 8 — Minimal UI ($6,000)
A Next.js page or thin web app where a user can enter input, see the model output, and inspect intermediate state. No design system, no settings, basic auth optional. 26 hours of frontend-capable senior engineering at $230 per hour. A vendor pitching a designed UI at this price is borrowing hours from another line.
Line 9 — Eval grading + iteration loop ($7,500)
Three iteration cycles across weeks 3–5: run the eval set, grade outputs against the rubric, fix the worst failure modes, re-run. The fractional eval partner runs the grading; the senior engineer adjusts the build. 32 hours of partnered work at a blended $235 per hour. This is the line that earns the “defensible” in defensible MVP.
Line 10 — Observability + logging ($3,500)
Structured logs for prompts, completions, errors, and latency — enough that the founder can debug the first production regression. 15 hours of senior time. The day-one observability stack piece covers tools and retention for builds that need more.
Line 11 — Handoff package ($3,500)
A README, a runbook covering the three most likely failure modes, prompt files, the eval CSV with current grades, and a 60-minute handoff call. 15 hours of senior engineering. Without it, the founder owns a black box.
Line 12 — Inference + infra ($5,800)
Frontier-model inference cost across the 6-week build window — Claude Opus 4.8, GPT-5, or Gemini 2.5 Pro, depending on the architecture decision. Vector store and embeddings if retrieval-augmented. Application hosting on a standard cloud. Observability log storage. This is a pass-through line — the vendor invoices actuals, not a marked-up estimate. Inference dominates: $4,000–$5,000 of the line is API calls during eval iteration.
What is NOT in this $75K budget
The exclusion list is as important as the inclusion list. A $75K MVP does NOT include:
| Excluded item | Where it lives | Typical cost if added |
|---|---|---|
| A second capability | A larger bracket ($100K+) | +$15K–$25K |
| A second integration surface | A larger bracket | +$8K–$15K |
| LLM-as-judge eval harness | $100K+ bracket | +$8K–$12K |
| Multi-tenant auth + customer isolation | $100K+ bracket | +$10K–$18K |
| On-call window after week 6 | Hardening + on-call line | +$15K–$40K |
| HIPAA / SOC 2 / GDPR evidence | Hardening or separate engagement | +$20K–$60K |
| Buffer for a frontier alias drop mid-build | Risk reserve | +$5K–$10K |
| Designed UI (vs minimal) | Separate engagement | +$8K–$15K |
| Customer-success or onboarding | Separate engagement | +$5K–$12K |
| 24-month TCO planning | Separate engagement | +$3K–$6K |
The exclusion list is the founder’s friend. A proposal that names what it excludes is one the founder can scope-control against. A proposal that lists only inclusions invites scope creep in week 2 — every “small addition” lands in the budget without a repricing conversation.
Honest $75K vs vendor cutting corners: the four diagnostics
Two proposals can both read “$75K, 6 weeks, AI MVP” and mean very different things. Four diagnostics separate the honest quote from the corner-cutter.
Diagnostic 1 — Is the eval contract a named artifact? An honest proposal includes a PRD + eval contract with a written rubric. A corner-cutting proposal mentions “eval” once in passing, with no rubric or test-set sizing. Hiding eval engineering at this price means the line gets absorbed into the build — which means the build ships against an eyeball test.
Diagnostic 2 — Are the engineers named and senior? An honest proposal names the senior engineer leading the build and lists their rate. The fractional eval partner has a name and a rate. A corner-cutter references “our team” with no named individuals, which usually means a junior engineer doing the bulk of the work at a senior bill rate.
Diagnostic 3 — Is the exclusion list explicit? An honest proposal includes a “what is NOT in scope” section roughly the length of the inclusion section. A corner-cutter lists only inclusions, leaving every grey area free to be billed as a change order in week 3.
Diagnostic 4 — Is the inference line a pass-through with a band? An honest proposal bands inference cost ($4K–$6K for a 6-week eval-iteration cycle) and invoices actuals. A corner-cutter point-estimates inference at $1,500, which means the proposal either under-budgeted eval iterations or plans to absorb overrun into other lines. Either way, the founder gets a surprise.
A proposal that passes all four diagnostics is one the founder can sign with a straight face. One that fails two or more is a $130K project mispriced as $75K — and the gap will show up as change orders or a quietly de-scoped deliverable in week 5.
The founder time line: $0 billed, 100 hours real
The line nobody puts in the proposal: founder time. A $75K AI MVP requires 80–120 hours of founder engagement across the 6-week window — heaviest in weeks 1–2 (scoping, PRD, eval set curation) and again in weeks 4–5 (eval grading, iteration decisions). At a $200 per hour founder opportunity cost — modest for a venture-backed or domain-expert founder — that is $16K–$24K of unbilled founder labor on top of the $75K vendor invoice.
This is a structural feature of the price point, not a hidden cost. Vendors at this bracket cannot afford to take over the work the founder uniquely can do — sampling representative inputs, owning the rubric, deciding what “acceptable output” means. A vendor offering to absorb all of that within the $75K is either over-promising or about to deliver an MVP graded against a synthetic eval set that does not reflect the real workload. BCG’s Build for the Future frames AI value capture as a co-creation problem — the $75K bracket is engineered around that co-creation, not white-glove delivery.
Is $75K right for your idea?
Five questions filter the idea against the bracket:
- Does it require more than one AI capability to prove? If yes, go up a bracket.
- Does it require more than one integration surface? If yes, go up a bracket.
- Will the MVP ship to paying customers in the 6-week window? If yes, the on-call line is required — go up a bracket.
- Can the founder commit 80–120 hours across the 6 weeks? If no, the vendor has to absorb that work, which means a bigger budget.
- Is the goal a fast feasibility confirm-or-kill against a real eval set, pre-PMF? If yes, $75K is the right call.
The $50K / $100K / $250K bracket piece walks the broader continuum. The founders’ MVP cost worksheet gives a related 11-line decomposition oriented to the $100K mid-bracket. For cost-line discipline that scales from MVP to TCO, see decoding AI project TCO — the seven lines most CFOs miss across a 24-month horizon.
Frequently asked questions
Is $75K really enough to build an AI MVP in 2026?
Yes — for a single-capability, single-integration, founder-co-authored MVP shipped against a graded eval set across a 6-week window. Not enough for two capabilities, a customer-facing production launch in week 6, or any compliance evidence. The bracket is honest under tight scope and silently insufficient if scope drifts even one feature wider.
What is the difference between a $75K and a $100K AI MVP?
A $100K MVP funds one additional structural element — usually a dedicated eval engineer with an LLM-as-judge harness, a second integration surface, or a 14-day on-call window after handoff. The $25K differential is not “more hours of the same thing”; it is one of three named additions that lift the build from feasibility-proof to defensible-against-regression.
Can a $75K AI MVP go to production with paying customers?
Sometimes — for an internal tool, a paid pilot with a single design-partner customer who tolerates rough edges, or a low-stakes consumer use case. Not for a multi-tenant SaaS launch, an enterprise customer with procurement expectations, or any HIPAA / SOC 2 / GDPR-scoped deployment.
How do I tell a vendor cutting corners from one quoting honestly at $75K?
Run the four diagnostics: (1) is the eval contract a named artifact with a rubric? (2) are the engineers named and senior? (3) is the exclusion list explicit? (4) is the inference line a pass-through with a band? A proposal failing two or more is a larger project mispriced.
How many engineers staff a $75K AI MVP?
One senior AI engineer at 50–60% allocation across 6 weeks, plus a fractional eval partner at 10–20%, plus the founder co-author. No PM, no designer, no dedicated QA. The team shape is intentionally lean because the scope is intentionally narrow.
How long does a $75K AI MVP take to build?
Six calendar weeks from kickoff to handoff. Week 0: discovery; weeks 1–2: PRD, eval set, architecture; weeks 3–5: build with eval iteration; week 6: handoff. Compressing below 6 weeks at this price requires cutting a line — typically Line 4 (eval set curation) or Line 9 (eval grading iteration), both of which break the “defensible” property.
Do I need to fund eval engineering on a $75K MVP?
Yes. Lines 4, 5, and 9 sum to $19,000, roughly 25% of the budget. Skipping eval engineering leaves the build graded against an eyeball test, which surfaces failure modes in production rather than in week 4. The piece on stop paying agencies for documentation, pay them for evals frames why eval engineering earns this price point.
What happens after the 6 weeks end?
Vendor involvement ends. The founder owns the runbook, eval CSV, prompt files, and the deployed application. If the MVP needs an on-call window, that becomes a separate engagement (typically $15K–$40K). The $75K bracket is engineered to be a clean exit, not a soft handoff into an open-ended retainer.
Is a $75K AI MVP a fixed-price contract?
Usually fixed-price for Lines 1–11 (labor) and pass-through with a banded estimate for Line 12 (inference). A fully fixed-price $75K that includes inference means the vendor either marked up inference 2x as a reserve or plans to cut eval iterations if inference runs hot. Banded pass-through inference is more honest.
Key takeaways
- $75K is a real, defensible AI MVP bracket in 2026 — but only against a tight scope: one capability, one integration, founder co-authorship, 6-week window, no on-call.
- The 12-line decomposition sums to $74,800 with named role, deliverable, and timing per line.
- Eval engineering (Lines 4, 5, 9) sums to ~$19,000 — roughly 25% of the budget. A proposal that omits these is structurally under-funding the defensible-MVP property.
- The exclusion list matters as much as the inclusion list. Name what is NOT in the budget so scope drift in week 2 triggers a repricing conversation rather than a silent change order.
- Founder time is unbilled but real — 80–120 hours across the 6 weeks, heaviest in weeks 1–2 and 4–5.
- Four diagnostics separate honest $75K from corner-cutters: named eval contract, named senior engineers, explicit exclusion list, pass-through banded inference.
Ready to scope your own AI MVP against a budget that can defend itself in a board meeting? Download the AI MVP Scoping Worksheet and lay it over the next vendor proposal that lands in your inbox.
Arthur Wandzel
