Most “starter package” pages from AI agencies in 2026 are not packages — they are price ranges with a button. A founder sees “$15K–$300K depending on scope,” fills a form, and receives a proposal with no defensible relationship to the floor of that range. No fixed scope, no named deliverables, no exclusion list. This piece names what an honest AI MVP starter package looks like in 2026: a $60–90K fixed-price contract, six weeks elapsed, one AI capability proven against a written eval set, one integration surface, observability and a fallback path baked in, and a 30-day post-handoff on-call window. Seven deliverables. A written exclusion list. One qualifier for whether to buy this tier or skip to the next.
It builds on the AI MVP economics playbook. The playbook frames the full 6–12 week cost space; this article fixes the contract at the bottom-defensible end. Both sit inside the idea-to-product manifesto, the master guide for non-engineer founders shipping AI products in 2026.
Why $60–90K is the defended floor in 2026
The floor sits at $60K because that is the dollar amount at which three structural lines — eval engineering, observability, fallback — survive without being silently cut. Below $60K the eval set turns into an eyeball test, observability collapses into a vendor’s default dashboard, and the fallback becomes a try/except block. The build still ships, but without the artifacts that distinguish a defensible MVP from a prototype demo.
The ceiling sits at $90K because at $91K the package is no longer a starter — it has funded a second integration, a longer on-call window, or a dedicated eval engineer. The $60–90K band leaves room for one honest variation: a slightly larger eval set, a slightly hairier integration, or a slightly longer founder co-author window. Outside that band, the package is mispriced.
The McKinsey State of AI survey has reported that 80–85% of AI pilots stall before reaching production. The starter is calibrated to be the cheapest engagement that does not stall — the smallest dollar amount at which eval contract, observability, and fallback survive into week 6.
The scope this package assumes
A starter is only defensible against a narrow scope. The seven dimensions:
| Scope dimension | Starter package assumption |
|---|---|
| Capabilities | One — a single AI-driven workflow step end-to-end |
| Integration surface | One — a single API, document store, or app surface |
| Build path | Prompt-engineered with light retrieval-augmented grounding |
| Eval set | 100–200 founder-curated representative inputs with a written rubric |
| Team | One senior AI engineer (50–60% allocated) + fractional eval partner (10–20%) + founder as co-author |
| Window | 6 calendar weeks from kickoff to handoff |
| Post-handoff | 30-day on-call for triage and bug fixes; no new feature work |
Two capabilities breaks the package. A hairier-than-single-API integration breaks the package. HIPAA, SOC 2, or GDPR evidence inside the window breaks the package. The starter is defensible because the scope is narrow — narrowness is what makes the price honest.
The seven deliverables
The starter package produces seven concrete artifacts. Each is a tangible deliverable a founder can name in a counter-sign meeting and ask for at handoff.
Deliverable 1 — PRD with eval contract appendix
A 6–8 page Product Requirements Document naming the capability, the workflow it replaces, the input/output contract, the success criterion, and the rubric the build is graded against. The appendix is what separates this from a 2018-style web MVP PRD — it specifies the 100–200 representative inputs, the rubric for what correct looks like, and the regression gate that fires when the model alias updates. A PRD without the appendix means this deliverable has been cut.
Deliverable 2 — Curated eval set + grading harness
A CSV of 100–200 founder-curated inputs paired with a harness that grades outputs against the rubric. Curating the inputs is founder work; the harness is engineering work that makes the eval set rerunnable on every prompt change, model update, and regression check. This is the artifact that says “the build is defensible” instead of “the demo looked good.”
Deliverable 3 — Built capability + retrieval scaffold
A prompt-engineered workflow with light retrieval-augmented grounding, accessible through a minimal interface. One capability, one prompt chain, one retrieval source, one output format. The narrow scope lets the build stay defensible against the eval rubric inside six weeks.
Deliverable 4 — One integration surface
The capability is wired into one upstream surface — an API the founder’s customers already call, a document store the team feeds, or a thin web app for a closed user group. Adding a second integration is the most common cause of a starter shipping at $130K instead of $75K — the silent scope drift vendors most often absorb without repricing.
Deliverable 5 — Fallback path + observability
A written fallback for failure modes the eval set surfaces: what the system does on low-confidence output, an integration timeout, or out-of-distribution input. Observability ships alongside — structured logging of every inference call, the prompt used, the output returned, and the eval grade where in-distribution. This is what makes the build production-ready in the operational sense: not “the demo worked,” but “the failure modes are named and the operator can see them.”
Deliverable 6 — Handoff package
A repository the founder owns: prompt files, eval CSV, harness code, runbook for rerunning evals and redeploying the prompt, README naming every dependency and its cost band, and a 1-page architecture decision record. The handoff turns a vendor engagement into a founder-owned asset.
Deliverable 7 — 30-day post-handoff on-call
A fixed 30-day window during which the senior engineer triages bugs, fixes regressions surfaced by the eval set, and answers founder questions. Not a retainer. Not new features. Designed to absorb the first month of real-world usage and the first frontier-model alias update. If on-call is sold as “we’ll help when needed,” this deliverable has been cut.
What is NOT in the starter package
The exclusion list is what makes the inclusion list defensible. The starter does NOT cover:
- A second AI capability. Two capabilities is the next tier — $120K–$150K — because each capability needs its own eval set, prompt chain, and grading loop.
- A second integration surface. One upstream surface is funded. A second is +$15K–$25K depending on hairiness.
- Compliance evidence. HIPAA, SOC 2 Type II, GDPR Article 35 DPIA, FedRAMP — none are in scope. The starter can ship into a compliance-shaped environment, but the evidence trail is not in the deliverables.
- Multi-tenant authentication and billing. Basic auth at most. A multi-tenant SaaS shape — org-level isolation, role-based access, billing — is +$30K–$50K.
- Customer-facing production launch. The starter ships an MVP defensible against the eval set; selling it to customers with SLAs and uptime guarantees is a hardening project.
- A second model migration cycle. The first frontier-model alias update inside the 30-day window is absorbed; the second is a separate engagement.
- Dedicated designer or PM. The senior engineer wears the design and PM hat for the minimal UI. A dedicated designer or PM is +$15K–$30K.
A vendor quoting the starter package without naming exclusions has either absorbed them (which means cutting something else) or plans change orders in week 4. Both break the price.
Who the starter package is right for
A founder who can answer “yes” to all four:
- Single capability is enough to prove the idea. One AI-driven workflow step that, if it works reliably, makes the product credible to the next stakeholder — a design-partner customer, a seed investor, a hiring candidate.
- One integration is enough to validate the use case. The capability does not need to be wired into three places to be tested; one surface is enough.
- A graded eval set against 100–200 inputs is enough to demonstrate defensibility. The founder has — or can curate inside the first two weeks — a representative input set and a written notion of what correct looks like on each.
- The audience for the MVP tolerates rough edges. Internal users, paid pilot customers who signed up for early access, a design-partner cohort. Not unknown customers at a public launch.
If all four are yes, the starter package is the correct dollar discipline.
Who needs the next tier up
The next tier is $120K–$180K and funds one named addition: a second capability, a second integration, or a dedicated eval engineer with an LLM-as-judge harness. A founder needs the next tier if:
- Two AI capabilities are required to be credible — one that retrieves and one that generates; one that summarizes and one that classifies. If a one-capability demo does not move the next stakeholder, upgrade.
- The integration is genuinely multi-surface — the MVP must ship into the founder’s app AND a customer-facing integration AND a back-office dashboard.
- Eval discipline must be production-grade — LLM-as-judge with sampled human review, drift detection, regression gate wired into CI. The starter eval discipline is fit-for-MVP, not fit-for-scale.
- Paying customers are expected inside week 8 — paying customers expect uptime, incident response, and a documented retention policy. Production customers shift the engagement toward the eval-first build playbook shape.
Buying the starter when the next tier is required ships an MVP that breaks at first contact with the use case that triggered the upgrade. Buying the next tier when the starter would have shipped overpays by $40K–$90K for capacity not yet needed. Picking the right tier is the budget discipline.
How to read a vendor proposal against this spec
Four checks:
- Priced between $60K and $90K? Below $60K, ask which structural line (eval engineering, observability, fallback) has been cut. Above $90K, ask which next-tier addition has been silently included.
- Seven deliverables named? PRD with eval contract, eval set + harness, built capability, integration surface, fallback + observability, handoff package, 30-day on-call. Missing deliverables surface as change orders in week 5.
- Exclusions written down? A starter without exclusions has not committed to a scope.
- Senior engineer and fractional eval partner named? A starter sold against “our team” without names is priced at senior rates and delivered against juniors.
The companion piece on pricing models walks the contracting mechanics; the anatomy of a $75K AI MVP shows a representative starter’s line-by-line decomposition; the companion piece on decoding “production-ready” in agency proposals walks the language traps around fallback and observability claims.
Frequently asked questions
What does an AI MVP starter package actually include?
Seven deliverables: a PRD with eval contract appendix; a curated eval set plus grading harness; one built AI capability with retrieval scaffolding; one integration surface; a fallback path with observability; a handoff package with runbook and architecture decision record; and a 30-day post-handoff on-call window. Six weeks elapsed. One senior AI engineer plus a fractional eval partner plus the founder as co-author.
How much does an entry-level AI MVP cost in 2026?
$60,000 to $90,000 for a defensible starter. Below $60K, structural lines (eval engineering, observability, fallback) get silently cut and the build stalls at the field 80–85% rate. Above $90K, the package has funded a second capability, a second integration, or a dedicated eval engineer — next-tier items.
Why is the floor $60K and not $25K or $40K?
At $25K–$40K the engagement pays for one engineer alone, against an informal eval discipline, with no observability and no fallback. $60K is where evals, observability, and fallback survive without being cut.
How long does a starter package take to deliver?
Six calendar weeks. Week 0 is discovery and scoping. Weeks 1–2 cover the PRD, eval set curation, and architecture decision. Weeks 3–5 are the build with eval iteration in parallel. Week 6 is handoff. Compressing below six weeks requires cutting eval curation or eval grading — both break the defensible property.
Is the starter package a fixed-price contract?
Labor lines are fixed-price; inference and observability are pass-through with a banded estimate (typically $4K–$6K across the window). A fully fixed-price quote including inference at this tier means the vendor either marked up inference 2x or plans to cut eval iterations if inference runs hot.
Can a starter package go to production with paying customers?
For a closed pilot with design-partner customers who signed up for early access, yes. For a multi-tenant SaaS launch, enterprise procurement, or a compliance-scoped deployment, no.
What happens after the 30-day on-call window ends?
The vendor relationship ends unless the founder books a separate hardening engagement (typically $30K–$60K for a 4–6 week pass with longer on-call, expanded eval coverage, and any compliance work). The starter is engineered as a clean exit, not a soft handoff into a retainer.
Who shouldn’t buy the starter package?
A founder needing two AI capabilities to be credible; a founder needing more than one integration; a founder with paying enterprise customers expecting SLAs from week 8; a founder needing HIPAA or SOC 2 evidence inside the engagement; a founder in a compliance-heavy vertical without an existing evidence trail. Each is a signal to skip to the next tier.
How do I tell an honest starter package from a vendor cutting corners?
Walk the proposal against four checks: priced between $60K and $90K; seven deliverables named; exclusions written down; senior engineer and fractional eval partner named with rates. A proposal failing two or more is a larger project mispriced as a starter — the gap surfaces as change orders in week 4 or de-scoped milestones in week 6.
What’s the difference between a starter package and a discovery week?
A discovery week is a 3–5 day paid scoping engagement (typically $8K–$15K) producing a PRD, eval contract draft, and architecture decision — no build. The starter package includes the discovery work as Deliverable 1 and ships the build over the following five weeks. Founders unsure whether their idea is build-ready should run a discovery week first.
Ready to scope yours?
Book a 30-minute idea review. Bring a one-page product description, the user persona, and the back-of-envelope budget. We say within thirty minutes whether the idea fits the starter package, needs the next tier up, or should run a paid discovery week first. Schedule the review.
Arthur Wandzel
