Autopsy of a pilot: why AI PoCs die in month 14

Autopsyof a pilot.

Why AI proofs of concept die in month fourteen — five findings from the table, and the boring fixes that keep AI alive in production.

Build by The Voidgap team 18 Sep 2026 5 min read draft

voidgap / autopsy reportcase no. 014 · composite

deceased: pilot_claims_copilot
age: 14 months
last seen: steering committee, v23
owner: none
evals: 0 / 0
security review: pending since month 5
risk class: unknown
runbook: —

cause of deathNobody decided.

Every dead pilot gets the same obituary. A demo that made the board lean forward. A steering committee that met fourteen times. A slide titled “Next steps”. Then, somewhere around month fourteen, silence. The deck survives. The system doesn’t.

This isn’t bad luck, and it’s rarely a model problem. Gartner expected at least 30 % of generative AI projects to be abandoned after proof of concept by the end of 2025 — because of poor data quality, inadequate risk controls, escalating costs or unclear business value.¹ For agentic AI it expects more than 40 % of projects to be cancelled by the end of 2027.² MIT NANDA’s 2025 study was blunter: 95 % of the generative AI projects it looked at showed no measurable business return, and only 5 % of custom enterprise AI tools reached production.³

30%of GenAI projects abandoned after PoC by end of 2025, per Gartner’s forecast [1]

40%+of agentic AI projects cancelled by end of 2027, per Gartner’s forecast [2]

5%of custom enterprise AI tools reach production, per MIT NANDA [3]

We have opened a lot of these bodies since 2011. The causes of death repeat. Here is the report.

The timeline

A composite, not one client. If you’ve sat in one of these projects, you’ll recognise the months.

month 0Kick-off. The demo works on stage. The budget comes from the innovation lab.
month 2The PoC works. On 40 hand-picked examples.
month 5Security asks what the agent is allowed to write to. Nobody knows.
month 8Legal asks for the risk class under the EU AI Act. The pilot waits.
month 11The champion changes jobs. The Slack channel goes quiet.
month 14Status: parked

Cause of death: five findings

finding 01 · build“unclear business value”

No owner.

The pilot belonged to a lab, not to a P&L. When nobody’s number moves, nobody fights for the budget in the next planning round. Gartner calls this unclear business value. We call it an orphan.

FixBefore line one of code, name the person whose KPI moves — and write their name into the README.

finding 02 · secure“it worked in the demo”

No evals.

A demo is not a test. Without a regression set, every prompt tweak and every model update is a gamble — and nobody with a signature signs off a gamble.

FixCollect a few hundred real cases, score them automatically, run them in CI from week one. The pass rate becomes the go-live criterion.

finding 03 · secure“inadequate risk controls”

No threat model.

Agents with write access to production systems. Customer data pasted raw into prompts. API keys in the repo history. Security finds it late and stops the rollout — correctly.

FixScope every tool, log every action, review before go-live — not after the incident.

finding 04 · comply“we’re not sure”

No compliance path.

Nobody classified the use case under the EU AI Act, so legal can’t say yes. In a large company, “we’re not sure” means no.

FixRisk class and documentation plan in month one. The AI Act reads like a spec — treat it like one.

finding 05 · enable“escalating costs”

No runbook.

Costs creep, nobody is on call, the model provider ships a change, and the pilot breaks on a Friday evening. Nobody knows who to wake up.

FixOwner, on-call rota, cost budget, runbook. The boring parts are what turn a demo into a product.

The PoC didn’t ~~fail~~. It was never built to live.
— every autopsy we’ve done, eventually

What survivors have in common

The pilots that make it past month fourteen aren’t smarter. They are planned as products from day one — across four moves that usually sit in four different departments:

BuildOne use case, one owner, one number that has to move.
SecureEvals in CI, scoped tools, an audit trail per decision.
ComplyRisk class, technical documentation, human oversight where it’s required.
EnableRunbook, on-call and a team that can run it without us.

Miss one and the pilot stalls at the hand-over. That’s where the gap opens — not inside the disciplines, between them.

The month-one checklist

Tick what your pilot already has. It stays on this device, nobody else sees it.

A named owner whose KPI moves
One use case, written in one sentence
200+ real cases for the eval set
Evals running in CI
Tool scopes reviewed by security
Risk class under the EU AI Act documented
A monthly cost budget, with an alert
Runbook and on-call before go-live

0 / 8 — month fourteen is coming

Where it still hurts

Honest part

Evals are expensive to build. The first version takes real hours from the people who know the process — the ones you can least spare.

Not every process deserves an agent. Some need a form and three rules. Shipping the boring version is still shipping.

Some pilots should die. Killing one in week six, with data, is a result. Dying in month fourteen without anyone deciding is not.

If your pilot is in month nine, you still have time. Call us before month fourteen.

Sources

Gartner, “Gartner Predicts 30% of Generative AI Projects Will Be Abandoned After Proof of Concept By End of 2025”, press release, 29 July 2024. gartner.com ↗
Gartner, “Gartner Predicts Over 40% of Agentic AI Projects Will Be Canceled by End of 2027”, press release, 25 June 2025. gartner.com ↗
MIT NANDA, “The GenAI Divide: State of AI in Business 2025”, as reported by Virtualization Review, 19 August 2025. virtualizationreview.com ↗