The numbers
| Finding | Figure |
|---|---|
| Companies that abandoned most of their AI initiatives in 2025 | 42% (up from 17% in 2024) |
| Proofs of concept scrapped before production, average organisation | 46% |
Source: S&P Global Market Intelligence, Voice of the Enterprise: AI & Machine Learning, 2025 — a survey of 1,000+ organisations in North America and Europe (S&P Global).
Three causes, none of them the model
| Where it stalls | What actually happened |
|---|---|
| Integration | The demo ran on a copy. Production needs the systems of record — and those have owners, rate limits and their own sign-in |
| Data readiness | It answered well on curated documents. The real corpus is duplicated, stale in places, and permissioned per user |
| Governance | Nobody could say what “working” meant, who is accountable when it is wrong, or what gets logged |
All three are delivery problems. Swapping the model fixes none of them.
What to fix first: a gap list before a build quote
Before anyone quotes a build, write down what is actually in the way. A useful gap list covers four things:
- Data readiness — which sources disagree, and who owns the conflict
- Integration points — which reads and writes exist, and which approvals do not yet
- Human in the loop — where a person reviews, and what they review with
- Success metrics — for a support assistant, for example: deflection rate, escalation accuracy, time to resolution
Most stalled pilots never had success metrics. That line is usually the one that matters.
Define “working” before you build
A pilot that impressed everyone in a demo has not been measured. Agree the test set, the metric and the threshold for go-live first. Then “is it working?” has an answer that is not an opinion.
A fixed-scope path out
The way out is narrow on purpose: one workflow, at a fixed price agreed at scoping. Not a platform migration, not a strategy deck — the one thing that was supposed to work.
Pricient quotes IOanyT’s Proof Sprint for this: in three to four weeks, one workflow is compiled — LLMs write and test the logic at build time, and deterministic code runs it, so the same input gives the same answer. It ends with a consistency and accuracy report on your own data and a written go / no-go, and you keep the code either way (what Compiled AI means). A sprint worth paying for also leaves you with:
- Evaluation against a fixed test set
- Permission hygiene — the assistant sees only what each user may see
- An incident log
- A named, accountable engineer
- A quote for running it afterwards, priced as an outcome rather than as a markup on your licences
Who signs off
The team that builds the agent should not be the one that certifies it. Independent review exists for exactly that reason; a vendor that does both has given you an opinion about itself.
How Pricient handles a stalled pilot
Describe what stalled — a Copilot rollout, an Agentforce build, a Bedrock agent, a RAG prototype. A person reads it; this lane’s own intake is not live yet. The specification comes with the gap list, before anyone quotes the Proof Sprint.