The budget you sign is the budget you’ll exceed unless 15% sits in a named contingency line. Most AI MVP budgets fail because contingency is implicit — a vague “we’ll figure it out” buffer that lives in everyone’s head and nowhere on the spreadsheet. On a $150K AI MVP the defensible reserve is $22.5K, decomposed into five specific unknowns: model deprecation mid-build, eval scope expansion, vendor outage workarounds, an unexpected fallback path, and late-stage regulatory friction. Each sub-line has a name, a defensible price, and a documented trigger. That decomposition is what converts the reserve from a vague buffer into a CFO-defensible line.
This article is part of the AI MVP economics playbook within the broader idea-to-product manifesto. It pairs with the model deprecation reserve playbook and the AI project TCO frame.
Why a named reserve, not an implicit buffer
The standard objection is that “the vendor already builds buffer into the quote.” Correct and irrelevant. The vendor’s buffer protects the vendor’s margin; the founder’s reserve protects the founder’s ability to ship at quality without re-opening the contract. Two different instruments, two different risk classes, both belong on the budget sheet.
PMI’s PMBOK guide separates contingency reserve (identified risks with quantified probabilities) from management reserve (unidentified risks). AACE International’s recommended practice sets the typical capital-project contingency at 10–30%, scaled by novelty and scope-definition maturity. A 2026 AI MVP is moderately novel and moderately well-defined; 15% sits at the lower edge of the software-project norm and lower-middle of the R&D norm. Defensible floor, not aggressive number.
Most founders skip the reserve because they confuse it with the vendor’s buffer and assume one of the two has it covered. Both have to exist. When the vendor’s buffer absorbs a 1-week delay from a broken integration API, the founder’s reserve is still available for the model deprecation event in week 9. Collapsing the two is how founders end up renegotiating in week 10 instead of launching.
The 5 unknowns an AI MVP budget has to absorb
Five risk categories recur frequently enough across 2024–2026 AI MVP builds to be named and reserved against. They are the ones a founder should expect to encounter at least once during a 6–12 week build.
Unknown 1. Model deprecation mid-build
The vendor you eval’d against ships a successor and announces a deprecation window for the variant your prompts and evals were tuned against. Anthropic, OpenAI, and Google each publish deprecation timelines; the cadence of frontier releases has run every 2–4 months per family across the past year. A 6–12 week build straddles at least one release event with high probability. When a model alias updates, the eval suite must be re-run, prompts may need re-tuning, and an agentic scaffold may need adjustment — none of which is in the original quote.
Magnitude: ~3.5% of total budget ($5.25K on a $150K MVP). Funds an eval re-run, a prompt port for any regressed category, and a 1-week calendar buffer. Deeper treatment in the model deprecation reserve playbook.
Example. A founder signs against the current frontier model in week 1. In week 7 the vendor announces deprecation in 6 months and ships a successor. The eval re-run surfaces two regression categories needing prompt adjustments — real work, real timeline pressure, no original budget.
Unknown 2. Eval scope expansion
During dogfooding or the first internal demo, a failure class surfaces the original rubric did not contemplate. The eval suite grows 20–40% in category count to cover the new failure mode. The eval set is the spec — a 2018 SaaS scope discovery cost a few engineer-hours; a 2026 AI scope discovery means representative inputs sampled, a rubric written, baselines hand-graded, the LLM-as-judge configuration updated, the CI gate re-anchored. Structurally larger work.
Magnitude: ~3.5% of total budget ($5.25K on a $150K MVP). Funds the suite expansion, baseline re-grading, and the prompt or scaffold work the new categories reveal.
Example. A customer-support assistant ships an internal demo in week 5 that surfaces escalation requests phrased as polite questions — a class the original rubric missed. Adding the category needs 30–40 sampled inputs, an anchored rubric, and re-grading. About a week of work, $4–5K of cost.
Unknown 3. Vendor outage and rate-limit workarounds
An outage, a rate-limit change, or a tier reshuffle forces the team to add a fallback provider, a queue layer, or retries-with-jitter not in the original architecture. SLAs on frontier inference APIs are weaker than SLAs on commodity cloud compute, and adding a second provider — abstraction layer, secondary credentials, cross-provider observability — must be done before launch traffic exposes the gap.
Magnitude: ~2.5% of total budget ($3.75K on a $150K MVP). Funds the abstraction layer, queue or retry mechanism, secondary credentials, and cross-provider observability.
Example. A real-time AI feature ships against a single provider. In week 8 a rate-limit adjustment pushes a hot prompt path past the new ceiling. The team adds a secondary provider, a queue with backpressure, and a routing layer — about two engineer-weeks.
Unknown 4. The unexpected fallback path
A class of inputs surfaces during dogfooding or the first external pilot that the LLM cannot resolve at the required quality bar. The class needs a non-LLM fallback: a deterministic rule layer, a search-based resolver, or a human-in-the-loop escalation. AI MVPs routinely have classes the model cannot resolve at the eval bar (numerical reasoning at the tail, domain-specific terminology, structured-output edge cases) — they only surface when representative traffic hits the system.
Magnitude: ~2.5% of total budget ($3.75K on a $150K MVP). Funds the fallback design, routing logic, and observability to track when it fires.
Example. An AI document classifier discovers in week 6 that short, high-stakes records produce unreliable model output. The fallback is a deterministic rule layer plus a “send to a human reviewer” path for ambiguous cases — about 1.5 engineer-weeks.
Unknown 5. Regulatory friction
Late-stage discovery of a regulatory requirement — GDPR data residency, EU AI Act logging or documentation, a sector rule (HIPAA, FINRA, FERPA) — adds legal review, redaction, audit logging, or a conformity-assessment step not in the original scope. The EU AI Act (Regulation 2024/1689) has rolling effective dates through 2026 with specific obligations for high-risk categories; US sector rules stack their own logging and consent requirements. Founders discover the obligation late — at launch readiness or during a customer security questionnaire — because the AI regulatory surface is moving fast and rarely covered in depth at week-1 scoping.
Magnitude: ~3% of total budget ($4.5K on a $150K MVP). Funds outside-counsel review, targeted redaction or logging work, and the documentation a customer security questionnaire requires.
Example. An EU-touching SaaS product discovers at launch readiness that its data routing places the AI feature under GDPR data-residency obligations. The fix is regional inference routing, a logging adjustment, and a $2.5K outside-counsel review. The reserve funds it; without one, launch slips.
The 15% reserve math
The defensible 2026 reserve is 15% of the signed budget, line-itemed into five named sub-lines:
| Reserve sub-line | % of total | $ on $150K MVP | Triggers when |
|---|---|---|---|
| Model deprecation mid-build | 3.5% | $5,250 | Vendor ships a successor or announces a deprecation window inside the build calendar |
| Eval scope expansion | 3.5% | $5,250 | A failure class surfaces that adds 2–3 eval categories |
| Vendor outage / rate-limit workaround | 2.5% | $3,750 | An outage or rate-limit event forces a fallback provider, queue, or retry mechanism |
| Unexpected fallback path | 2.5% | $3,750 | A class of inputs the LLM cannot resolve at the eval bar is discovered |
| Regulatory friction | 3.0% | $4,500 | A regulatory obligation surfaces late and adds review, redaction, logging, or documentation |
| Total reserve | 15.0% | $22,500 |
Each sub-line has a name, a magnitude, and a trigger. When a risk fires, the founder spends the matching sub-line; the others stay untouched. The governance rule — one sub-line at a time, only when the named trigger fires — keeps the reserve from being drained by ordinary scope creep.
The reserve does not fund routine scope changes the founder requests (those re-open the quote), warranty bugs in the vendor’s code, a feature the founder forgot to ask for in week 1 (change order), or vendor underestimation in any single line (vendor’s buffer). It funds the five named risks. Nothing else.
For the $150K baseline, see how much an AI MVP costs in 2026 and the AI MVP pricing model comparison.
When 15% isn’t enough
Four conditions push the defensible reserve from 15% to 20–25%.
- Regulated industry. Healthcare, finance, legal, or any sector with a mandatory compliance regime moves the regulatory-friction sub-line from 3% to 6–8%. Reserve → 18–20%.
- Novel model or vendor. If the team has not shipped against the model or provider before, deprecation and outage sub-lines each move up 1–2 points. Reserve → 18–19%.
- Multi-tenant from day 1. Tenant isolation, per-tenant rate limits, and per-tenant eval calibration are real engineering scope the standard 15% does not fund. Reserve → 18–20%.
- Mission-critical eval bar. Domains where a failure is not safely recoverable (medical guidance, financial transaction execution, legal advice) carry a structurally larger eval-expansion sub-line. Reserve → 20–25%.
Rule of thumb: zero conditions → 15%. One → 18%. Two → 20%. Three or four → 22–25% and a re-read of whether the project is correctly classified as an MVP at all. The anatomy of a runaway AI project covers the cost-side root causes that recur when these conditions are present.
How to defend the reserve to a CFO
Three framing moves carry most boardroom conversations.
Move 1. Name the unknowns. Show the five sub-lines explicitly. A CFO who has signed off on construction or traditional software has seen 10–30% contingency lines; the framing is familiar. The decomposition is the unfamiliar part. Showing model deprecation, eval expansion, vendor outage, fallback path, and regulatory friction as five named sub-lines converts “what is this 15% for?” into “we have a named line for each of these five risks.”
Move 2. Anchor to PMI and construction norms. PMI’s PMBOK distinguishes contingency from management reserve and recommends both. AACE International sets capital-project contingency at 10–30%, scaled by novelty. A 2026 AI MVP is at least as novel as a construction project on a familiar site, and 15% sits at the lower edge of the range. This is less conservative than the construction median, not more.
Move 3. Propose an unspent-reserve-returns-to-budget clause. Write into the contract that any unspent portion at project closeout returns to the company. The reserve is not a pre-spend; it releases only when one of the five named triggers fires, with documented evidence. This converts the reserve from “an extra $22.5K the vendor might spend” into “an insurance line that returns if the named risks don’t materialize.” Most CFOs accept the framing immediately — they have seen the equivalent clause in capital-project and IT-services contracts.
The TCO decoder for AI projects walks through the seven cost lines CFOs typically miss; the contingency reserve is one of them.
Frequently asked questions
Is 15% a hard rule or a starting point?
Starting point. It is the defensible floor for a non-regulated, single-tenant, single-vendor, low-stakes AI MVP at $100K–$200K. Move up when the four conditions in the section above apply: one condition → 18%, two → 20%, three or four → 22–25%.
What’s the difference between contingency and overrun?
Contingency is funded in advance against named risks. Overrun is unfunded spend after the budget is exhausted. The point of the 15% reserve is to convert what would otherwise be overrun into planned contingency spend with documented triggers. Hitting the reserve is not an overrun — it is using the line as designed.
Can I cut the reserve if I’m using a fixed-price vendor?
No. The fixed-price quote protects the vendor’s exposure on known scope. The five unknowns are founder-side risks outside the vendor’s scope sheet. The fixed-price contract and the 15% reserve are complementary; both belong on the budget. The pricing model comparison walks through which model fits which context.
How is AI MVP contingency different from web MVP contingency?
Web MVPs were priced on deterministic engineering scope; a 10% buffer was usually enough. AI MVPs face a different surface: model behavior drifts on a 90-day cycle, the eval set grows as understanding deepens, vendor SLAs are weaker than commodity cloud, fallback paths surface only at dogfooding, and the regulatory surface is moving fast. 15% is the AI-specific floor.
Who owns the reserve — the founder or the vendor?
The founder. The vendor may invoice against the reserve when a trigger fires (with documentation), but the reserve never converts to vendor-controlled spend. The release decision sits with the founder.
What happens to unspent reserve?
Returns to general budget at project closeout. Name this clause explicitly in the contract — without it, the reserve risks becoming a slush fund the vendor can redirect into scope expansion.
Do I disclose the reserve to my board?
Yes, and as a separate budget line. A board reviewing the AI MVP budget should see a $150K build line and a $22.5K contingency reserve, with the five sub-lines in an appendix. Folding the reserve into the build line hides the most important budget discipline on the page.
How does the reserve change at the seed vs Series A stage?
At seed, capital pressure tempts founders to cut the reserve. Don’t — seed is exactly the build where the five unknowns are most likely to fire and where re-opening the contract is hardest. At Series A, the conditions for going to 20% are more likely (multi-tenant, rigorous compliance). The percentage adjusts upward; the five-sub-line discipline stays constant.
What if my CFO insists on 10% instead of 15%?
Two options. Walk through the five sub-lines and show 10% — each drops to roughly 2%, which covers none of the five typical risks. Or accept 10% but document that one risk firing will require a contract amendment. Make the trade-off visible. Most CFOs, shown the explicit choice, move to 15%.
Should the reserve grow if the project extends beyond 12 weeks?
Yes, proportionally. 15% is calibrated to 6–12 weeks. At 16–20 weeks, the model-deprecation and vendor-outage sub-lines face higher cumulative probability of firing; the reserve should grow roughly 1 point per additional 4 weeks. A 20-week engagement should carry an 18–19% reserve.
Closing
The 15% reserve is not generous and not aggressive — it is the defensible floor for a 2026 AI MVP, anchored to PMI and AACE norms, decomposed into five named unknowns, governed by a trigger discipline, and returnable to budget if the unknowns don’t fire. Most AI MVP budgets fail not because the vendor priced wrong but because the founder skipped this line and met the cost in week 9.
A founder walking into the contract conversation with a $150K build line, a $22.5K reserve broken into five sub-lines, an unspent-reserve-returns-to-budget clause, and a one-page CFO defense has done budget work most of their peers will not. That work is the difference between shipping in week 10 and re-opening the contract in week 10.
Next step. Before signing the next AI MVP quote, run the budget through the AI MVP scoping worksheet — it includes the five-sub-line contingency template, the four-condition reserve-sizing rule, and the unspent-reserve clause language.
Arthur Wandzel
