How do you measure AI support success beyond deflection rate?

Source quality beat model size for us. Clean knowledge + tool scopes fixed more hallucinations than switching models.

Start with one bounded workflow that has a clear success metric. We tried to automate three use cases at once and none of them got good enough to ship.

The vendor demo is not the product. Ask to see the same workflow run on your data, not their sample data. That is where connector gaps and permission issues show up.

Measure rework, not just throughput. An agent that resolves 80% of cases but creates 30% more manual cleanup is not saving time.

We learned the hard way that 'human in the loop' is not a checkbox. If the approval UI is buried or slow, reviewers will batch-approve without reading.

Security questions should be part of the first demo, not a procurement afterthought. Ask about retention, sub processors, prompt-injection testing, and audit logs before you waste time on a trial.