News
89% of AI agent pilots never reach production. What the other 11% did differently.
Gartner puts the number at 89%. The survivors share four traits, and none of them is the model.
Production notes
Frontal Designs
It is an engineering problem in an AI costume
Most agent pilots start as a one-week demo that reads an invoice and drafts the entry. Then it meets four thousand real invoices a month in eleven formats and falls over. The demo answered whether a model can do the task. Nobody asked whether it can run unattended on a Tuesday at 2am when the ERP is slow.
Scope creep and data quality together cause 61% of failures.
Agents with full automated evaluation had a 9% rollback rate. Agents without had 47%.
Unclear ownership appears in every failure analysis we have read.
Own a single process
Not an AI assistant for finance. One process: invoice in, matched to a purchase order, posted, anything odd to an exception queue. If you can draw it on a whiteboard in two minutes you can build an agent that finishes it. If you cannot, you are building a chatbot and calling it an agent.
A human in the loop by design, not by apology
The agent takes the 80 or 90 percent it is confident about and hands the rest to a person with its reasoning attached. Teams that push for full autonomy on day one usually reach zero adoption by month three, because the first bad output kills trust and nobody switches it back on.
Somebody's name is on it
An agent that belongs to the AI team belongs to nobody. The agent that reconciles supplier data belongs to the head of procurement, and their ops lead checks the queue every morning. That is what keeps the thing switched on.
Do not skip it. That is the whole secret of the 11%.
Blog & Insight
Read More Notes
89% of AI agent pilots never reach production. What the other 11% did differently.
