Scope creep in a 2026 AI MVP does not look like the scope creep your project-management textbook warned about. It does not arrive as a stakeholder asking for “one more screen”. It arrives as an eval suite that grows from 12 cases to 140 between weeks three and five, a second model added “for comparison”, a fallback chain quietly handling seven unscoped edge cases, and an observability dashboard now storing 90 days of full traces. Each move is individually reasonable. Together they overrun the budget by 30–60% and ship a system the eval contract no longer defends. This piece names the six AI-specific scope-creep patterns by mechanism, prices each in line items, and gives the refusal script that holds the original SOW.
A companion to the AI MVP economics playbook and the idea-to-product manifesto. The AI agency change-order playbook covers the procedural side. See also the AI MVP cost curve and stop scoping AI projects in features, scope them in evaluations.
Why AI MVP scope creep is mechanically new
Pre-2024 scope-creep textbooks were written against deterministic software — “another screen”, “another integration” — with change-control boards as the fix. That orthodoxy still covers the deterministic surfaces of a 2026 AI MVP. It does not cover the four new surfaces every AI MVP carries: the eval surface, the model surface, the trace surface, and the non-determinism surface.
| Surface | New scope-creep pattern | Why it lacks a 2018 analogue |
|---|---|---|
| Eval | Eval surface inflation | Eval suite did not exist as a budget line in 2018 |
| Model | Second-model temptation, agentic escalation | Single-vendor backends rarely tempted dual-vendor builds |
| Trace | Observability sprawl | LLM trace volume is 10–100x CRUD app logs |
| Non-determinism | Unbounded fallback, hallucination chase | Deterministic code has no “wrong-but-confident” failure mode |
McKinsey’s State of AI reporting across 2023–2024 shows 70–85% of enterprise AI pilots stalling before production. A large share is funded scope creep across the six patterns below. A 2026 MVP that finishes on the original budget is one whose vendor refused each pattern by name before it cost anything.
Pattern 1 — Eval surface inflation
How it starts. Week three of a 6-week build. The suite is at 12 cases — the PRD-bound test set the eval-first scoping piece prescribes. A stakeholder asks at demo: “What about Spanish prompts? Adversarial users? Empty inputs?” The engineering lead adds five cases per Friday. By week five the suite is at 140 cases. Runtime: 90 seconds to 28 minutes. Cost per LLM-as-judge run: $12 to $180.
What it costs. Eval-engineering line moves from 25% of the build crest to 40–50%. Each retroactive case forces a re-run across every prior model variant — quadratic, not linear. Trace storage adds 3–5x to the observability bill. Year-1 magnitude: $15–35K unbudgeted plus $400–800 per month of run-rate inflation.
How to refuse. “We add this case to the backlog of eval candidates, not the current contract. The contract was set against the PRD’s behaviour list. If that list changes, that is a PRD change — priced separately.” Eval changes require a documented PRD addendum.
Pattern 2 — The second-model temptation
How it starts. Week four. Anthropic, OpenAI, or Google ships a new variant — cadence on the Artificial Analysis LLM Leaderboard runs roughly one major update every 6–10 weeks per provider. The vendor’s engineer asks: “Should we A/B the new alias before we lock?” The framing is “free intelligence”. It is not free. A second model means a second routing layer, a second cost-per-query line, a second eval pass, a second prompt-tuning cycle, and a second incident surface.
What it costs. Roughly $8–15K of one-time engineering to integrate, evaluate, and route. Plus 30–60% inflation on the M5–M8 inference line if the routing stays. Plus an open-ended commitment: every future migration now requires two evals instead of one.
How to refuse. “We lock the model at PRD signature. Migration is a scheduled event, not a parallel A/B inside the build. The model-migration buffer in the M5 retainer fires when the alias drifts, not when a new alias ships.” Single-model builds finish. Dual-model A/Bs become dual-model maintenance contracts.
Pattern 3 — Unbounded fallback logic
How it starts. Week four or five. QA surfaces edge cases — malformed input, non-English string, 503 from the provider. The engineer adds a fallback: retry, then degrade to a cheaper model, then return a canned response. Two weeks later there are seven fallbacks, three nested. The fallback chain is now a system, with its own failure modes, observability needs, and eval cases.
What it costs. Each fallback adds a code path the eval suite must cover and the trace surface must log. A 7-deep chain inflates eval-engineering by 25–40% and the M6–M8 plateau retainer by 15–25%. The hidden cost is opacity — the chain is undocumented, and nobody who built it is on-call by M9.
How to refuse. “Fallback logic is a separately priced workstream. Each fallback requires three artifacts: an eval case proving it improves the rubric score, a trace event proving it fires, and an on-call SOP. Without all three, the fallback does not ship.” The discipline the change-order playbook calls “structured refusal” — anchored in contract, not in engineer reluctance.
Pattern 4 — Observability sprawl
How it starts. Week five. Hardening surfaces a real concern — “how do we debug an incident at 2am?”. The engineer adds full-trace logging, then 30-day retention, then 90-day because “audit”, then a per-workflow dashboard, then percentile alerting, then a weekly trace digest. Each move is reasonable for production. None were in the SOW.
What it costs. Observability bills run 5–15% of total inference when bounded, 30–50% when unbounded. Public pricing on the major trace platforms (Langfuse, LangSmith, Arize, Helicone) puts a 10M-event-per-month workspace at $400–1200 per month depending on retention. Unbounded 90-day retention can land at $3–6K per month for a single-workflow MVP.
How to refuse. “Observability scope is named in the SOW with three numbers: retention window, volume cap, dashboard count. Anything beyond is a change order.” The hardening sprint is where this conversation lands — before the launch invoice, not after the M6 bill arrives.
Pattern 5 — The hallucination chase
How it starts. Week four through six. Demos surface hallucinations — wrong-but-confident outputs. The vendor chases each one: a new prompt instruction, a new retrieval source, a new validation pass, a new system message. The chase is real engineering, and unbounded — hallucinations on long-tail inputs cannot be eliminated, only ranked by frequency and addressed in order. Without an eval-anchored stopping rule, the chase consumes the hardening sprint and the on-call retainer.
What it costs. Chase cycles consume $10–25K unbudgeted engineering in hardening and inflate the M6–M9 plateau by 20–40% as new validation passes ship. Hours chasing tail hallucinations are hours not spent on the launch crest and observability stand-up.
How to refuse. “Hallucination defects are addressed in priority order against the eval rubric, with a stopping rule named in the SOW: we stop chasing when the rubric crosses [target score] across [target case set]. Long-tail hallucinations below threshold are tracked, not chased.” The discipline of scoping in evaluations, not features — the rubric, not the demo, decides what ships.
Pattern 6 — Agentic escalation
How it starts. Week three or four. The PRD specified a static-prompt MVP. The engineer proposes “a small agentic loop, two tools, one path”. The second tool needs a third. The loop needs a retry, then trajectory eval, then a planner. By week six the system is a 5-tool agent with a new failure surface the original eval contract does not cover.
What it costs. The most expensive of the six. A static-prompt MVP that becomes a 5-tool agent typically adds $20–40K to the build crest, 30–60% to the plateau retainer, and near-doubles the eval line. Agentic systems are also more brittle to migration — the crest in the cost curve piece lands harder. Public benchmarks (SWE-bench, GAIA) show agentic systems trail single-call systems on consistency.
How to refuse. “Static-prompt MVPs ship; agentic MVPs become research projects. If the workflow genuinely requires agentic behaviour, we re-scope in writing — new PRD, new eval contract, new budget envelope. We do not slip from static to agentic inside a fixed-price build.” Same logic as the smallest-possible-feature spoke: ship static first, escalate later under a new contract.
How the six patterns interact
The patterns are not independent. Interaction effects are where the unmanaged budget actually breaks.
| Pattern | Pulls in |
|---|---|
| Eval inflation | Observability sprawl (more cases need more trace storage) |
| Second model | Eval inflation, hallucination chase (two models, two eval passes, two failure surfaces) |
| Fallback logic | Observability sprawl, eval inflation (each fallback needs a trace event and eval case) |
| Observability sprawl | (terminal) sprawl funds itself once retention expands |
| Hallucination chase | Eval inflation, fallback logic (each chased defect becomes a case and a fallback) |
| Agentic escalation | All five others |
The two highest-impact patterns to refuse first are agentic escalation (pulls in all five others) and eval inflation (pulls in observability and fallback logic). A vendor that holds the static-prompt boundary and the PRD-bound eval contract through week five typically finishes within 5–10% of the original M5 invoice. A vendor that lets either drift overruns by 30–60%.
A week-by-week scope-creep self-test
Run at every Friday review, weeks two through five. Each question maps to one or two patterns. Any “yes” without a documented change order is the scope-creep signature.
- Did the eval suite grow this week without a PRD addendum? (Pattern 1)
- Was a second model integrated, A/B’d, or “compared” this week? (Pattern 2)
- Was a fallback added without a corresponding eval case? (Pattern 3)
- Did trace retention or dashboard count change this week? (Pattern 4)
- Was an agentic loop, planner, or tool added that was not in the original architecture? (Pattern 6)
The hallucination chase (Pattern 5) is the residual — if questions 1, 3, and 5 all return “yes” in the same week, the team is in active chase and the hardening sprint is at risk. A vendor that runs this test every Friday is the vendor whose M5 invoice matches the SOW.
Frequently asked questions
What percentage of AI MVPs run over budget because of scope creep?
McKinsey’s 2023–2024 State of AI shows 70–85% of enterprise AI pilots stall before production. A large share traces to the six patterns. The 30–60% overrun range is consistent with what SFAI Labs and peer agencies see across 2026 fixed-price MVP engagements when discipline is absent.
Is scope creep mostly a founder problem or a vendor problem?
Both — but the vendor owns the refusal script. Founders ask reasonable questions in week three. Vendors who say “yes, free” rather than “yes, priced separately” are the proximate cause of the overrun.
How is AI scope creep different from SaaS scope creep?
Pre-LLM SaaS scope creep was mostly feature additions. AI scope creep operates on four new surfaces (eval, model, trace, non-determinism) that did not exist as budget lines in 2018. Generic change-control templates do not catch the six patterns.
Which of the six patterns is most common in a 6-week MVP?
Eval inflation (Pattern 1) and hallucination chase (Pattern 5) by frequency. Agentic escalation (Pattern 6) and observability sprawl (Pattern 4) by cost when they fire. Refusing Patterns 1 and 6 first gives the highest budget protection per refusal.
What is the cheapest single change that prevents scope creep?
A weekly Friday review with the five-question self-test, anchored to the eval contract and PRD addendum process. The change-order playbook documents the procedural shell.
Should the SOW explicitly name all six patterns?
Yes — and the refusal scripts. Naming patterns at week zero is the cheapest insurance against funding them at week five. A proposal that does not name them has not priced them.
Can scope creep be eliminated entirely?
No. Some expansion is legitimate — PRD changes, new edge cases. The goal is priced expansion: every change moves through a documented PRD addendum and a re-baselined budget. Unpriced expansion is the failure mode.
What is the role of evals in scope-creep discipline?
The eval rubric is the contract. The six patterns all break it in different ways — inflation expands it, hallucination chase tries to satisfy it asymptotically, agentic escalation replaces it. Holding the rubric is the operational definition of holding scope.
How do the six patterns interact with model migration in year one?
The migration crest in the cost curve piece lands harder on systems built under any of the six patterns. Tight build-time discipline keeps the M7–M9 migration a small spike, not a budget event.
What if the vendor will not name the six patterns?
A vendor that cannot articulate the six patterns and the refusal scripts will overrun on at least two of them. The five-question self-test doubles as a vendor-evaluation test — run it alongside the field guide to evaluating an AI agency in under 90 minutes.
Where to go next
- Pre-signature: walk the proposal against the five-question self-test. If any pattern is unnamed, ask for the refusal script in writing.
- Mid-build, suspecting drift: run the Friday self-test for two reviews. Any “yes” without a PRD addendum is the trigger to escalate before the M5 invoice.
- Post-overrun: anatomy of a runaway AI project names the five cost-side root causes — map the six patterns onto that frame to renegotiate the M6 retainer.
- Procurement-side: the fixed-price AI MVP contract: 7 clauses worth negotiating and the change-order playbook provide the contract language behind the refusal scripts.
For the six patterns translated into a scoping worksheet — line items, refusal scripts, Friday-review checklist — download the SFAI Labs AI MVP Scoping Worksheet.
Arthur Wandzel
