From POC to production: the 5 blockers that kill your AI projects
Many AI POCs never make it to production. Here are the 5 structural blockers we keep seeing at our clients, and the decision grid to dodge them from the scoping phase.
We were told for three years: « you have to do POCs ». Today, executive boards are waking up to a brutal reality. The vast majority of AI proof of concepts launched since 2023 have never left demo status. Not for lack of technology. For lack of method.
1. The use case is picked out of enthusiasm, not ROI
First trap: the POC is launched because a member of the executive team saw a stunning demo at CES. The topic is technically seductive but operationally marginal. Six months later, the project delivers what it promised, except no one in the organisation actually needs it.
« A good AI use case is boring to talk about over dinner. It's a repetitive, high-volume task, with an ROI you can compute in a spreadsheet. »
2. No budget line for the run
The POC budget is sponsored by innovation. The move to production requires a recurring line on IT opex, which was never planned. The project dies between two budget committees.
3. The data isn't ready (and no one wants to say it)
We discover in month 3 that the data needed is scattered across 4 systems, badly qualified, with inconsistent naming conventions. The POC runs on a hand-picked sample. Production would require 6 months of upfront data engineering. No one had costed it.
4. No identified business owner
The POC was led by IT or an external firm. At delivery, no business director is responsible for adoption, impact measurement or evolution. The tool is put online, no one looks after it, it dies within six weeks.
5. No impact measurement protocol
We launch, we observe « it seems to work », but we never defined the before/after KPIs, the measurement scope, or the threshold above which we industrialise. Without a measurement grid, impossible to defend the next phase to the board.
The ProofPilot decision grid
- Does the use case represent > 200h/year of measured repetitive task?
- Is there a named business sponsor with a quantified objective in their annual scorecard?
- Is the 24-month run cost in the multi-year budget?
- Is the data accessible, clean and authorised? (data audit upfront, not during POC)
- Are the before/after KPIs defined, instrumented and signed by the sponsor?
If you have 5 yeses, launch. If you have less than 4, don't launch, or launch a 2-week scoping first to turn the noes into yeses. It's our internal rule, and the one we apply on every mission.
