A founder receives a 6-week AI MVP proposal that reads “fixed price, $75K, ships in 6 weeks.” Two pages, three milestones, no calendar, no exclusion list. By week 3 the build is two features wide and the eval set is a screenshot. By week 6 the founder owns a demo, not a product. The window is not the problem. The specification is. A defensible 6-week AI MVP sprint exists in 2026, but only against a day-by-day calendar, a written exclusion list, and a $60–90K band that the buyer can hold the vendor to. This piece names what a 6-week sprint includes (one capability + eval + fallback + observability + handoff), what it explicitly does NOT include (a second feature, multi-tenant auth, polish, on-call), the day-by-day artifact calendar across all 30 working days, and the hand-off package without which the engagement is not complete.
This piece sits under the AI MVP economics playbook, the reference on AI MVP build economics, and inside the idea-to-product manifesto, the master guide for non-engineer founders shipping AI products in 2026.
Why the 6-week sprint is a specification, not a window
Most vendor proposals treat “6 weeks” as a marketing window — a price tag attached to a calendar with no internal structure. The buyer signs against a window, the vendor scopes against a window, and by week 3 the project has drifted into a second feature, a second integration, or a third meeting cadence. McKinsey’s State of AI reporting documents that roughly 80–85% of AI pilots stall before production. A meaningful share of that stall is the failure to specify, not the failure to build.
A 6-week sprint becomes defensible when it is treated as a specification: a day-by-day calendar of artifacts the buyer can point at, a fixed inclusion list, a longer exclusion list, and a handoff package the buyer can refuse to accept the engagement as complete without. The window itself does not deliver value; the specification does. The rest of this piece is the specification.
The five capabilities a 6-week sprint must include
A defensible 6-week AI MVP ships with five capabilities — not five features, five capabilities. Each one is load-bearing. A sprint missing any one is not a defensible MVP, regardless of what the calendar reads.
| # | Capability | What it is | Why it is non-negotiable |
|---|---|---|---|
| 1 | One named capability | A single AI-driven workflow step end-to-end (input → AI work → output, in production paths) | Two capabilities in 6 weeks ships neither correctly |
| 2 | A graded eval set | 100–200 founder-curated representative inputs scored against a written rubric | Without it, “done” means “the demo worked once” |
| 3 | A no-AI fallback | A deterministic path that runs when the model refuses, times out, or breaches a confidence threshold | Without it, the first model regression is a customer outage |
| 4 | Observability | Structured logs for prompts, completions, latency, errors, and refusal events | Without it, the founder cannot debug the first failure mode |
| 5 | A handoff package | Runbook, prompt files, eval CSV with grades, README, 60-minute walkthrough | Without it, the founder owns a black box |
A vendor proposal that names fewer than five capabilities is either silently absorbing the missing ones into the build (which means cutting eval iterations) or planning to ship without them (which means the MVP is not defensible). Either way, the founder is buying a demo. The why your AI MVP should ship with 3 evals not 30 piece argues why fewer, sharper evals beat the temptation to inflate the eval set in week 4.
The 6-week day-by-day calendar
The sprint runs across 30 working days. Each week ends with a named, file-extensioned artifact the founder can refuse to advance without. The vendor’s standup, demo, or status update is not the artifact — the file is.
Week 1 — Discovery + eval contract
| Day | Activity | Output |
|---|---|---|
| Mon | Kickoff working session (founder + senior engineer + fractional eval partner) | Kickoff memo (1 page) |
| Tue | Capability scoping; one-line capability statement; user-action it replaces | Capability spec (capability-spec.md) |
| Wed | PRD drafting (input/output contract, success criterion) | Draft PRD v0 |
| Thu | Eval rubric drafting; sample 20 representative inputs with founder | Rubric v0 (rubric.md) |
| Fri | PRD + eval contract sign-off; founder reviews exclusion list | Signed PRD + eval contract |
Named weekly output: PRD + eval contract (the artifact a vendor cannot proceed without).
Week 2 — Eval set + architecture
| Day | Activity | Output |
|---|---|---|
| Mon | Founder + eval partner curate 80–180 more inputs against the rubric | Eval set v1 (eval-set.csv) |
| Tue | Architecture decision record: prompt-only, retrieval-augmented, or thin agent | ADR (adr-001.md) |
| Wed | Eval harness scaffold (loader, runner, grader stub) | Harness skeleton |
| Thu | First end-to-end dry run (10 inputs, no quality bar) | Dry-run output CSV |
| Fri | Week-2 review with founder; eval threshold proposed (e.g., 80% pass @ rubric) | Eval threshold signed |
Named weekly output: Eval set + threshold contract — the quality bar the build will be graded against.
Week 3 — Build, integration, first eval pass
| Day | Activity | Output |
|---|---|---|
| Mon | Prompt scaffold + retrieval index (if applicable) | Prompt v1, retrieval v1 |
| Tue | Integration layer (one API or document store) | Integration code |
| Wed | First full eval run against the curated set | Eval results v1 (pass rate ~40–60%) |
| Thu | Failure-mode triage; founder reviews 20 worst outputs | Failure-mode log |
| Fri | Build-side fixes from triage; second eval run | Eval results v1.1 |
Named weekly output: First eval baseline — the build’s starting pass rate against the rubric.
Week 4 — Iteration + fallback + observability
| Day | Activity | Output |
|---|---|---|
| Mon | Prompt iteration (structured output, system prompt, few-shot examples) | Prompt v2 |
| Tue | No-AI fallback path (confidence threshold + deterministic route) | Fallback code |
| Wed | Observability instrumentation (structured logs, latency, refusal events) | Observability stack |
| Thu | Third eval run | Eval results v2 (pass rate ~65–80%) |
| Fri | Week-4 review with founder; decision on remaining gap | Go/no-go on threshold |
Named weekly output: No-AI fallback + observability — the two capabilities that move the MVP from “model works” to “system holds up.”
Week 5 — Minimal UI + final eval
| Day | Activity | Output |
|---|---|---|
| Mon | Minimal UI (input box, output panel, intermediate state view) | UI v1 |
| Tue | UI integration with build; basic auth (optional) | UI integrated |
| Wed | Final iteration on the worst 10% of eval cases | Prompt v3 |
| Thu | Fourth eval run (the deciding one) | Eval results v3 |
| Fri | Founder + vendor review eval v3 against threshold | Threshold pass/fail decision |
Named weekly output: Threshold-passing eval run (or, in honest sprints that miss, a documented gap analysis).
Week 6 — Handoff
| Day | Activity | Output |
|---|---|---|
| Mon | README draft (architecture, deployment, ops) | README v0 |
| Tue | Runbook (three most likely failure modes + recovery) | Runbook v0 |
| Wed | Founder walkthrough (60 min); founder runs eval end-to-end alone | Walkthrough recording |
| Thu | Final eval CSV with grades archived | eval-final.csv |
| Fri | Sign-off; vendor involvement ends | Handoff complete |
Named weekly output: Complete handoff package — the artifact set without which the engagement is not done.
What is NOT in a 6-week sprint
The exclusion list is longer than the inclusion list, and that is the point. Naming what is NOT in scope is what prevents week-3 scope drift from quietly becoming a $40K change order.
| Excluded item | Where it actually lives | Typical cost if added |
|---|---|---|
| A second AI capability | A 9–12 week engagement, or a follow-on sprint | +$25K–$40K |
| A second integration surface | A 9–12 week engagement | +$10K–$18K |
| Multi-tenant auth + customer isolation | Hardening + on-call engagement | +$12K–$22K |
| HIPAA / SOC 2 / GDPR evidence | Separate compliance engagement | +$25K–$70K |
| LLM-as-judge eval harness | $100K+ bracket | +$8K–$14K |
| Designed UI (component library, theme) | Separate design engagement | +$8K–$18K |
| On-call window after week 6 | Hardening + on-call retainer | +$15K–$45K |
| Customer success / onboarding flows | Separate engagement | +$6K–$15K |
| Admin console / settings UI | A larger engagement | +$8K–$18K |
| Performance optimization beyond the threshold | Hardening engagement | +$10K–$20K |
| Polish, animations, marketing-site integration | Separate frontend engagement | +$8K–$20K |
| 24-month TCO planning | Separate finance engagement | +$3K–$8K |
The exclusion list is the buyer’s friend. A proposal that includes only inclusions is one that absorbs every grey area into the labor lines and prices the change order in week 3. The 6-week AI MVP framework: what to defer piece extends this list with the deferral logic for each item.
Cost band: what $60K, $75K, and $90K buy
The $60–90K band is not negotiation room — each $15K step buys a structural element. A vendor quoting a single point estimate without naming what the buyer gets at each rung is collapsing real structural differences into a pricing slogan.
| Price point | What it buys |
|---|---|
| $60K | One senior AI engineer (50–55% allocation), founder owns eval set curation entirely, no fractional eval partner, minimal observability (structured logs only) |
| $75K | One senior engineer + fractional eval partner (15% allocation), founder co-curates eval set, observability includes refusal-event tracking, runbook covers 3 failure modes |
| $90K | Same staffing as $75K + an extra week of iteration (Week 4 doubles in eval-grading time), runbook covers 5 failure modes, observability dashboard included, 14-day post-handoff Slack-only support |
The anatomy of a $75K AI MVP decomposes the $75K bracket into 12 named lines. The $50K / $100K / $250K bracket piece sets the bands above and below this sprint.
Hand-off artifacts: the proof the sprint is complete
The sprint is not complete when the demo works. It is complete when the founder owns the eight handoff artifacts and has run the eval set themselves at least once without the vendor.
- README — architecture summary, deployment steps, environment variables, dependency list.
- Runbook — the three most likely failure modes with named recovery steps. Five for $90K.
- Prompt files — all prompts versioned, with the prompt-history log showing what changed between v1, v2, v3.
- Eval set CSV — 100–200 graded inputs with rubric scores, archived.
- Eval harness code — runnable script that loads the eval set, calls the build, captures outputs, writes graded results.
- Architecture decision record — the ADR naming the build path with trade-offs.
- Observability stack — structured logs schema, dashboard URL (if included), retention policy.
- 60-minute walkthrough recording — vendor walks founder through the system; founder runs the eval set end-to-end live.
A vendor that wraps up with a demo and a slide deck — and no eight artifacts — has not delivered a 6-week MVP sprint. They have delivered a 6-week demo. The stop paying agencies for documentation, pay them for evals piece frames why the eval CSV is the load-bearing artifact in the handoff package, not the README.
Founder time: 60–100 hours, unbilled but real
The line absent from every sprint proposal: the founder’s own time. A 6-week sprint requires 60–100 hours of founder engagement, heaviest in weeks 1–2 (eval set curation, PRD sign-off) and again in weeks 4–5 (iteration decisions, threshold call). At a $200/hour founder opportunity cost, that is $12K–$20K of unbilled labor on top of the $60–90K vendor invoice.
This is structural. BCG’s Build for the Future frames AI value capture as co-creation — the vendor cannot decide what “representative input” means for the founder’s workflow, and a vendor that offers to handle eval curation alone is either over-promising or quietly grading the build against a synthetic set that misses the actual workload. The honest sprint codes founder time as a structural feature, not an optional contribution.
Four diagnostics for an honest sprint proposal
Two proposals can both read “$75K, 6 weeks, AI MVP sprint” and mean very different things. Four diagnostics separate honest from theatrical.
Diagnostic 1 — Is there a day-by-day calendar? Honest proposal: 30-day calendar with named outputs per week. Theatrical proposal: three milestones across six weeks with vague titles.
Diagnostic 2 — Is the exclusion list longer than the inclusion list? Honest proposal: explicit “what is NOT in scope” with where each excluded item lives. Theatrical proposal: inclusions only, with “scope to be confirmed” attached.
Diagnostic 3 — Are all five capabilities funded? Honest proposal: capability + eval + fallback + observability + handoff are each line-itemed. Theatrical proposal: capability is a line; the rest are “included” or absent.
Diagnostic 4 — Is the handoff package enumerated? Honest proposal: eight artifacts named, each with a file extension. Theatrical proposal: “handoff documentation” as a single line.
A proposal that passes all four is one a founder can sign and hold the vendor to. One that fails two or more is a sprint window with a price tag — not a specification. The companion piece on stop budgeting AI projects in story points, budget them in eval runs frames why eval runs are the right unit for the sprint’s measurement contract. For the downstream cost discipline once the sprint is done, see decoding AI project TCO: 7 cost lines most CFOs miss.
Frequently asked questions
Can a 6-week AI MVP sprint really ship a defensible product?
Yes — for a single-capability, single-integration MVP with a founder-curated eval set, an exclusion list that prevents scope drift, and the five capabilities (capability + eval + fallback + observability + handoff) all funded. Not for a multi-feature MVP, a customer-facing production launch with compliance evidence, or any build that skips eval engineering.
Why six weeks specifically — not four, not eight?
Six weeks is the smallest window that contains the eval iteration loop. Weeks 1–2 build the eval contract; weeks 3–5 iterate the build against it; week 6 handles handoff. Compressing to four weeks usually means skipping eval set curation or the iteration loop, which breaks the defensible property. Extending to eight weeks usually means the team is iterating against a poorly anchored eval suite — push the team to fix the suite, not extend the calendar. The companion 6-week AI MVP scope and milestones piece walks the same window from the angle of what is actually possible to fit.
What is the difference between a 6-week sprint and a 12-week build?
A 12-week build funds two capabilities, two integrations, a fuller UI, an LLM-as-judge eval harness, and a 14–30 day on-call window after handoff. A 6-week sprint funds one capability and stops at the lean defensible threshold. The AI MVP economics playbook walks the full 6–12 week economics.
How do I know the vendor will not absorb scope creep into the build?
Two signals. First, the proposal includes an exclusion list as long as the inclusion list, with where each excluded item lives. Second, the contract names a repricing clause that triggers when an excluded item is requested mid-sprint — even by a single line. Without both, scope creep is a near-certainty by week 3.
Is the founder really expected to spend 60–100 hours over six weeks?
Yes. Weeks 1–2 require ~25 hours (kickoff, PRD sign-off, eval set curation). Weeks 4–5 require ~25 hours (iteration decisions, threshold call). Weeks 3 and 6 require ~10 hours each. A founder who cannot commit this should buy the 9–12 week bracket, where the vendor absorbs more of the work.
What if the eval pass rate misses the threshold at week 5?
Honest sprints document the gap, ship what is defensible, and price a follow-on iteration sprint (typically 2–3 weeks at $25K–$40K). Theatrical sprints rewrite the threshold downward in week 5 and ship anyway. The handoff package’s eval CSV is what makes the gap visible — refuse to accept a sprint without it.
Does the sprint include any on-call period after week 6?
The base $60–75K band does not. The $90K band includes 14 days of Slack-only support — answers to operational questions, not new development. Any production on-call is a separate engagement, typically $15K–$45K for a 30–60 day window with response-time SLAs.
What models does a 6-week sprint typically use?
Frontier general-purpose models — Claude Opus 4.8, GPT-5, or Gemini 2.5 Pro — chosen during the Week 2 architecture decision based on the capability’s needs (long-context vs structured output vs tool-use). Inference cost lands in the $4K–$6K range across the 6-week build window, invoiced as pass-through actuals, not a marked-up estimate.
Can the sprint go straight to production with paying customers?
For an internal tool or a single design-partner customer who accepts rough edges, yes. For a multi-tenant SaaS launch, an enterprise customer with procurement expectations, or any regulated deployment — no. Production launch is a separate engagement that funds multi-tenancy, compliance, on-call, and polish.
What happens if the founder skips a weekly artifact sign-off?
The sprint loses its specification property and becomes a window again. The weekly sign-off is what prevents week-3 scope drift and what keeps the vendor accountable to the calendar. Skipping it is the single most common pattern behind a sprint that ships a demo instead of a defensible MVP.
Key takeaways
- A 6-week AI MVP sprint is a specification — a day-by-day calendar with named weekly outputs, an exclusion list, and a handoff package — not a marketing window.
- Five capabilities must be funded: one named capability + a graded eval set + a no-AI fallback + observability + a handoff package. A sprint missing any one is not defensible.
- The 30-day calendar produces a named, file-extensioned artifact each week: signed PRD + eval contract (W1), eval threshold (W2), first eval baseline (W3), fallback + observability (W4), threshold-passing run (W5), handoff package (W6).
- The exclusion list is longer than the inclusion list — a second capability, multi-tenancy, designed UI, compliance evidence, on-call all live in larger brackets or separate engagements.
- Cost band $60K–$90K: each $15K step buys a structural element, not negotiation room.
- Eight handoff artifacts (README, runbook, prompt files, eval CSV, harness code, ADR, observability stack, walkthrough recording) — without them, the engagement is not complete.
- Founder time is structural: 60–100 hours unbilled, heaviest in weeks 1–2 and 4–5.
- Four diagnostics separate honest from theatrical sprints: day-by-day calendar, exclusion list ≥ inclusion list, all five capabilities funded, eight handoff artifacts enumerated.
Ready to scope a 6-week sprint against a specification that defends itself at sign-off? Request an idea-review call and bring the next vendor proposal that lands in your inbox.
Arthur Wandzel
