A defensible 6-week AI MVP ships one narrow feature plus an eval set, a no-AI fallback, observability, and a handoff bundle — and nothing else. Anything bigger is a 12-week build mislabeled. The 6-week scope is real in 2026 for teams that hold the line on a single capability, defer the admin dashboard and the second feature to MVP-2, and pay roughly $60K to $100K for the bundle. Founders who lose the timeline lose it three ways: scope creep, eval inflation, premature optimization. This piece walks the week-by-week artifact map, names the in- and out-of-scope lines, and prices the build against defensible 2026 ranges.
The 6-week scope is the cost-side floor of C1 — one tier lighter than the $150K idea-to-product MVP and presuming the discipline argued in the case for the smallest-possible AI feature. Longer-horizon economics live in the AI MVP economics playbook. The end-to-end engagement frame is the idea-to-product manifesto; the validation discipline upstream is the idea-validation playbook.
What a 6-week AI MVP actually is
A 6-week AI MVP in 2026 is a single-feature production build wrapped in the five layers a serious AI system needs: an eval set, a no-AI fallback, observability, prompt versioning, and a documented handoff. One senior engineer paired with a planner can scope, build, harden, and hand off the bundle in 30 working days. Anything broader is a 12-week build compressed on paper that will decompress on the calendar.
The window is feasible in 2026 in a way it was not in 2023 for three reasons. Frontier models — Claude Opus 4.8, GPT-5, Gemini 2.5 Pro — handle a wider slice of capability out of the box, removing weeks of fine-tuning earlier builds carried. AI-assisted engineering productivity has compounded: GitHub Octoverse 2025 data is consistent with senior engineers shipping 1.5× to 2× the code per day of the 2023 baseline. Eval-harness tooling has matured: a 50-to-200-case eval set in CI takes hours to wire, not days.
What 6 weeks is not: a multi-feature suite, a multi-tenant platform, an enterprise deep-integration, or a fine-tuned-model product. The scope is a wedge for one capability proven in production. MVP-2 picks up the rest.
Week-by-week structure with named artifacts
The 6-week calendar collapses into five phases. Every phase ships a named artifact. If the artifact is missing at end-of-week, the calendar is already slipping and the founder should renegotiate scope rather than push the milestone right.
| Week | Phase | Hours | Artifact that ships by Friday |
|---|---|---|---|
| 1 | Planning + capability map | 30–40 | Capability map, thin-slice prototype, signed scope |
| 2 | PRD + eval set | 30–40 | Eval-first PRD with 50–100 labeled test cases on day one |
| 3 | Build sprint 1 | 50–60 | Feature wired end-to-end against staging, first eval pass |
| 4 | Build sprint 2 | 50–60 | Feature production-deployed, eval re-run, fallback live |
| 5 | Hardening | 40–50 | Security pass, observability instrumented, runbook drafted |
| 6 | Handoff | 25–35 | Knowledge transfer session, runbook signed, eval cadence agreed |
The hour bands assume one senior engineer paired with a planner-PM. A solo engineer extends the calendar by roughly two weeks; a team of three compresses hardening but not planning, which is not parallelizable.
Week 1: planning and capability map
The output: a capability map (“can the current model reliably do this?”), a thin-slice prototype that exercises the riskiest step, and a signed scope. The thin-slice is load-bearing — it retires user-value risk before any production code. The discipline is in the case for the smallest-possible AI feature. If scope is not signed by Friday, week 2 should not start.
Week 2: PRD and eval set
The eval-first PRD is the artifact most 6-week builds skip — and the reason most slip. It names the feature, the user journey, the input distribution, and 50 to 100 labeled test cases that define “working.” The harness wires those cases into CI so every prompt change or model swap re-runs the suite. Without the eval set, week 5 hardening has nothing to harden against. A defensible 2026 PRD follows the framework in the AI PRD explained.
Weeks 3 and 4: build sprints
Ten working days of paired senior engineering. Sprint 1 wires the feature end-to-end against staging and runs the first eval pass — the gate that decides whether the feature is real. Sprint 2 production-deploys, re-runs the eval, wires the no-AI fallback, and ships the first observability dashboard. If eval pass rate is under 80% at end of week 3, scope shrinks until pass rate clears 80%. Founders who push past that line own the consequences in week 5.
Week 5: hardening
Security review against the OWASP LLM Top 10. Observability wired: input/output capture, eval-replay, cost dashboards, latency percentiles. Runbook drafted: what does the founder do at 2 a.m. when the model API errors, the drift detector fires, costs spike? The runbook is what distinguishes a $60K demo from a $100K product.
Week 6: handoff
A live debug session walks the founder through the codebase, the eval harness, the observability stack, and the runbook. The handoff is not documentation; it is a working session where the founder runs an eval, deploys a prompt change, and reads the dashboard with the engineer present. The longer-engagement version of this kickoff dynamic lives in anatomy of an AI agency engagement: the first 14 days.
What’s in scope: 1 feature plus 5 supporting layers
The defensible 6-week bundle is one feature wrapped in five layers. Any of the five missing and the bundle is a demo, not a product.
| Layer | What it covers | Why it is in scope |
|---|---|---|
| The feature | One narrow AI capability the user touches | Must be one, not three |
| Eval set | 50–100 labeled cases in a CI-wired harness | Defines “working”; survives model swaps |
| No-AI fallback | A deterministic path when the model errs or times out | Without it, every model failure is customer-facing |
| Observability | Input/output capture, latency, cost, drift dashboards | What the founder reads at 2 a.m. |
| Prompt versioning | Versioned prompts in source control with rollback | Enables safe iteration without regressing the eval set |
| Handoff | Runbook + live transfer session | Converts a recipient into an operator |
The bundle is mutually reinforcing. The eval set guards the feature; the fallback survives eval failures in production; observability detects drift the eval missed; prompt versioning lets the founder fix what observability surfaces. Remove one and the rest weaken.
What’s out of scope: the exclusion table
The exclusion table is the part the founder most needs in writing before signing.
| Excluded | Why it is out of MVP-1 | Defer to |
|---|---|---|
| Admin dashboard | 1–2 weeks of UI for non-end-user roles | MVP-2 |
| A second AI feature | Doubles eval scope, fallback surface, on-call load | MVP-2 |
| Multi-tenant data isolation | 2+ weeks of schema and auth work | MVP-2 or pre-Series-A |
| Fine-tuning | 2–4 weeks of data prep and eval re-baseline | After product-market fit |
| Polished design system | Multi-week iteration past the MVP shell | Post-MVP-1 |
| Mobile app shell | A separate platform build | Post-MVP-1 |
| Real-time streaming | If async is acceptable, async is the scope | Add only if eval demands it |
| SSO / enterprise auth | Days of integration per identity provider | Pre-enterprise-customer-1 |
| Two-language support | Doubles the eval set and prompt surface | Post-launch |
Each item is a real piece of engineering that breaks the 6-week budget if slipped into scope without renegotiation. A founder who wants the admin dashboard moves to a 9-week scope at roughly $90K to $130K. That is a defensible trade. Pretending the dashboard fits the 6-week budget is not.
The 3 founder traps that turn 6 weeks into 12
Every slipped 6-week AI MVP we have audited slipped one of three ways.
Trap 1: scope creep disguised as “small additions”
Week 3, the founder watches the first demo and asks “can we also add X?” X is a settings page, an export button, a second user role. None is large. The trap is that three smalls is a medium, and a medium adds two weeks because every addition multiplies the eval surface, the fallback paths, and the handoff. Defense: a written change-request form. Any addition past week 1 is named, sized in days, and either swapped for an in-scope item or deferred to MVP-2.
Trap 2: eval inflation
Week 2, the team starts the eval set at 200 cases and the founder pushes for 500 because “more is better.” Scope balloons. By week 3 the team is still labeling instead of building. Build sprints compress, hardening is skipped, the product ships untested. Defense: 50 to 100 cases on day one, drawn from the actual input distribution. A 100-case set across the right distribution out-performs a 500-case set across noise. The eval set grows in MVP-2 against real production traffic.
Trap 3: premature optimization
Week 4, the team notices inference cost is higher than expected. The next 10 days dissolve into prompt golf, model-switching experiments, and a partial vendor migration. The feature ships late, the eval set has not been re-run against the new model, and the founder owns a system whose quality is unknown. Defense: do not optimize cost in MVP-1. Ship on whichever frontier model passed the eval set in week 2. Cost optimization belongs to MVP-2, after the eval baseline is locked.
Real 2026 cost: $60K to $100K, defended
A 6-week AI MVP at defensible 2026 senior-engineer rates costs $60K to $100K:
| Line item | Range | Share |
|---|---|---|
| Planning week | $8K–$14K | ~13% |
| PRD + eval engineering | $12K–$18K | ~18% |
| Build sprints (weeks 3–4) | $25K–$40K | ~42% |
| Hardening (week 5) | $10K–$16K | ~17% |
| Handoff (week 6) | $5K–$8K | ~8% |
| Inference + infra | $1K–$3K | ~2% |
| Total | $60K–$100K | 100% |
Defenders: $180 to $220 per fully-loaded hour for a US senior AI engineer (consistent with Stack Overflow Developer Survey 2025 compensation bands plus standard agency-side load); one senior engineer plus a 0.5–0.7 FTE planner-PM; 2026 AI-assisted productivity at roughly 1.5× to 2× the 2023 baseline (consistent with GitHub Octoverse 2025); eval-set runs in CI plus dev/staging traffic dominate the inference line. EU teams price 20% to 30% lower.
Pricing outside the range is a signal. Below $60K is undisclosed offshoring, an inexperienced team, or a quote missing one of the five supporting layers. Above $100K is a 9-week scope mispriced as 6-week, or padded margin. Ask “what is not in this number?” at proposal stage, not at week 4.
The founder pre-signing checklist
Before signing, a non-technical founder should hear a written answer to every item:
- One feature, named in writing? Not a family. One.
- Eval set sized and harness named? 50 to 100 cases in CI.
- No-AI fallback path explicitly in scope? The most-skipped layer.
- Observability — prompt capture, drift, cost, latency — included? All four.
- Runbook is a deliverable, not “post-launch”? Signed by end of week 6.
- Handoff is a live session, not a doc dump? Founder runs an eval with the engineer present.
- A written exclusion list? Item by item.
- A change-request process? A 15-minute written form, not a Slack ping.
- A model fallback plan if the chosen frontier model is deprecated? A defensible 2026 engagement has one.
- The MVP-2 conversation cadence? Roughly week 5.
A proposal that cannot answer in writing is either under-scoped or mispriced. Thirty minutes of pre-signing questions saves three weeks of mid-engagement renegotiation.
Book a 30-min idea review
If your 6-week scope is being priced — or you have a proposal in hand and want a second opinion on whether the bracket matches the scope — book a 30-min idea review. We walk the week-by-week artifact map against your scope, name the missing layers, and produce a defensible build plan.
FAQ
Is 6 weeks enough to build a real AI MVP in 2026?
Yes — for one narrow feature wrapped in the five supporting layers (eval set, fallback, observability, prompt versioning, handoff). Anything broader is a 12-week build mislabeled. The 2026 frontier-model capability and AI-assisted engineering productivity make the 6-week scope feasible in a way it was not in 2023.
What does a 6-week AI MVP not include?
A second AI feature, an admin dashboard, multi-tenant isolation, fine-tuning, a polished design system, a mobile shell, SSO, custom analytics, real-time streaming, or two-language support. Each defers to MVP-2 or to a 9-to-12-week scope.
How much does a 6-week AI MVP cost in 2026?
$60K to $100K at defensible US senior-engineer rates: $8K–$14K planning, $12K–$18K PRD and evals, $25K–$40K build sprints, $10K–$16K hardening, $5K–$8K handoff, $1K–$3K inference. Below $60K signals offshoring or missing layers. Above $100K signals a 9-week scope or padded margin.
What ships at the end of each week?
Week 1: capability map, thin-slice, signed scope. Week 2: eval-first PRD with 50–100 test cases in CI. Week 3: feature wired end-to-end on staging, first eval pass. Week 4: production-deployed feature, eval re-run, fallback live. Week 5: security pass, observability instrumented, runbook drafted. Week 6: signed runbook, live handoff, eval cadence agreed.
Why is the eval set so heavily weighted?
It is the instrument that defines “working.” Without it, the build has no defensible quality measurement, prompt changes cannot be regression-tested, and model swaps are blind. A 6-week build without an eval harness shipped a demo, not a product.
What if my idea needs more than one AI feature?
Then it needs more than 6 weeks. The honest scope is 9 to 12 weeks at $90K to $150K, or a phased plan: MVP-1 at 6 weeks covering the highest-risk feature, MVP-2 at 4 to 6 weeks adding the second. Two AI features in a 6-week budget is the most common cause of slip.
How do I know if my scope is 6 weeks vs. 12 weeks?
If you can name one user journey and one AI capability in one sentence each, and accept the exclusion table without negotiation, it is 6 weeks. If any of those breaks, it is 9 to 12 weeks. The bracket-exclusion question surfaces the answer in 10 minutes.
What is the biggest risk to a 6-week timeline?
Scope creep disguised as small additions during weeks 3 and 4. A written change-request process — every addition named, sized, and either swapped or deferred — is the discipline that protects the calendar.
Can a solo founder DIY a 6-week AI MVP with Cursor or Claude Code?
You can build something inside 6 weeks for the cost of your time. The output is a thin-slice prototype or a working demo, not the bundle above. The five supporting layers require deliberate engineering work most solo founders skip. DIY is appropriate for validation, not shipping to paying users.
Where can I read about the broader engagement model?
The idea-to-product manifesto describes the program end-to-end. The idea-validation playbook is the C1 anchor upstream. The AI MVP economics playbook is the cost-side companion at the C3 anchor tier. The $150K decomposition lives in how much an AI MVP costs in 2026.
Arthur Wandzel
