Insights
Everything here comes out of running the work rather than watching it. Support operations, agentic AI in customer-facing workflows, evaluation loops, guardrails, and the parts of a rollout that only appear once real tickets hit the system. No forecasts about where the industry is heading. No vendor scorecards. Written for people who have a working pilot and no path into production, by someone who ran the escalations in that support organization before he owned any of the AI.
Featured
Containment rate is the easiest number in support automation to move and the easiest to fake. Optimize it before resolution quality and you will build an agent that protects the queue from the customer. Here is what to measure instead, and in what order.
Enterprise AI does not usually die in the model. It dies in review, waiting on a document nobody was assigned to write.
Aviation did not get safe by hiring better pilots. It got safe by assuming good ones would fail, and writing down what happens next.
The job description keeps changing. The job has not: give the person on the other end something they can act on.
Rewrite every prompt, move platforms entirely, and a good evaluation set still tells you whether the new thing is better.
Escalation is not an edge case. It is a permanent, high-traffic route, and most agent programs design everything except the exit.
Published elsewhere
Publishing here first is deliberate — the writing stays close to the work it comes from.
Open to commissions and guest pieces on shipping agentic AI into live support operations. Contact