Why AI projects fail at the process layer
A useful AI initiative still collapses if the underlying process is unclear, expensive or fragmented. The model gets blamed — but the process was the failure point all along.
The core issue
Most AI projects do not fail because the model is weak. They fail because the workflow around the model is not ready.
If the process is inconsistent, the AI layer only helps the inconsistency travel faster. Automation without a clear workflow ends up automating the exceptions, not the rule — and the exceptions are almost always what was breaking the business in the first place.
Three signals the process is not ready
You see a different version of the process on every team member's desk. Nobody is wrong — everyone is running their own patched variant. The workflow diagram the business thinks it has does not exist anywhere except in the principal's head.
Exceptions outnumber the rule. If 'the standard way' happens 40% of the time and six different 'edge cases' make up the rest, there is no process for AI to augment. There is a bundle of conventions held together by individual judgement.
Nobody owns the end-to-end. Operations owns steps 1–3, sales owns step 4, finance owns step 5, and the handoffs live in email. This is the most common and the most expensive pattern — it makes every AI improvement load-bearing on a human who is not in the room.
The diagnostic we run at the start
Before any model selection or data-pipeline conversation, we run a two-hour workflow walkthrough with the three people who actually do the work. Not the team lead. Not the department head. The operators.
We draw the workflow on one page in front of them. We ask the same three questions at every step: what is supposed to happen, what actually happens, and what happens when it breaks. The gap between the three answers is the map of where AI can and cannot help.
When the fix is process-first, not model-first
If the diagnostic shows that the process is ambiguous, inconsistent or unowned, the first week of the engagement is not an AI project. It is a process-redesign project with an AI component scoped to phase two.
This is the conversation clients dread and respect in equal measure. Nobody wants to hear that the sexy AI initiative has to wait for a workflow rewrite. But the ones who act on it are the ones whose pilots ship — and stay shipped.
More from the field notes
The AI pilot blueprint
The shape of an AI pilot decides its fate before anyone writes a prompt. Here is the engagement we run — phase by phase — and why every slot is engineered to close on a handover, not to keep the advisor in the room.
Read morePrioritizing AI use cases without the hype
The right use case is the one that fits the operating model and creates leverage the business can own — not the one that makes the best demo. A field-tested filter for the shortlist.
Read moreThe cost discipline framework
Cost work becomes effective when it is tied to the operating model instead of one-off budget cuts. A short framework for cost programs that survive the quarter.
Read more