What went wrong
Right answers nobody used.
The agent's markdown recommendations held up well against what experienced buyers later chose. But it presented a number, not a reason, and the buyers carried the margin risk personally.
We had optimised for accuracy and skipped the harder question: would the people accountable for the decision act on it? By week eight, most of them were not opening the recommendations at all.
What we changed in how we work
Lesson 1
Explain before you recommend
Every agent now shows the records and reasoning behind a suggestion, not just the suggestion.
Lesson 2
Ask who carries the risk
If the user owns the consequence, they co-design the agent from week one.
Lesson 3
Measure use, not just accuracy
Adoption is a release criterion alongside the test-set score.
It was probably right. I just couldn't defend it in the Monday meeting.
