Across the assessments we run, the same failures recur with enough regularity to be predictable. None of them are technical.
Tooling before policy
A team buys a platform, produces output at volume, and discovers later that nobody decided what quality bar applied or who signs off. The cost is not the licence — it is the rework, and the credibility lost with the stakeholders whose support the next project needs.
Piloting on data that does not exist in production
The pilot uses a clean sample somebody prepared. Production data has missing fields, inconsistent formats and duplicates nobody audited. The model was never the problem; the data was, and it was invisible until go-live.
No baseline
Work starts, results are claimed, and nobody can prove causation because the before-state was never recorded. This makes every subsequent funding conversation harder than it needed to be, and it is entirely avoidable with two weeks of measurement.
No named owner
AI systems drift. Models change, data shifts, edge cases accumulate. Without someone accountable for monitoring and revalidation, a working deployment degrades quietly until a customer notices — which is the most expensive way to discover it.
Scaling before the quality bar holds
A pilot produces acceptable output at low volume with careful review. It is scaled, review becomes a bottleneck, review is relaxed, and quality falls below the bar that justified the project. The pilot succeeded; the rollout removed the thing that made it work.
Why these persist
Every one of these is a governance failure wearing a technology costume, so technology teams cannot fix them alone and business teams assume they are technical. That gap is where programmes stall — and it is why our assessment work starts with decisions rather than with tooling.
What this means for your business
Before the next AI project, answer four questions in writing: who reviews the output, what happens when it is wrong, what the current process costs today, and who owns it in six months. If any answer is missing, that is the actual first task. Our method for identifying AI opportunities covers how to choose the project itself.