career bridge

When it goes wrong

Cost-rank every action, design escalation as a real surface, and write the one sentence naming who answers when the agent is wrong.

You can scope an agent and say what good means. Now the part that separates this role from product management generally: the thing acts. When it acts wrongly, somebody is out of pocket, or waiting, or angry, and a question arrives about who decided this was allowed.

What you do

Design the failure path with the same care as the happy one. Start with cost. For every action the agent can take, write what a wrong instance costs and who bears it: the customer, the company, one named team. Then order them. The expensive ones are where your controls go, and the cheap ones are where you can afford to let it run. Next, design escalation as a product surface rather than a fallback: what triggers a handover, what the human sees at that moment, how much context comes with it, and how long they have to act. A handover that dumps a transcript on somebody is an escalation in name only. Write the reversal path for every action that can be undone, and mark plainly the ones that cannot, because those need approval before rather than apology after. Then write the accountability line in one sentence: when this agent does something wrong, who answers for it. If you cannot write that sentence, the design is not finished. Finish with a pre-mortem run with two other people. It is six months from now and this was pulled. Write down why.

Done when

  • Every agent action has a written cost of being wrong and a named party who bears it.
  • Escalation is specified as a surface: trigger, what the human sees, context handed over, time budget.
  • Reversible and irreversible actions are separated, with approval required before the irreversible ones.
  • One written sentence names who answers when the agent is wrong.
  • A pre-mortem with at least two other people is written up, with the top three reasons it failed.

What you end up with

A failure design: cost-ranked actions, an escalation surface specified in detail, a named accountable owner, and a written pre-mortem.

If you get stuck

The common failure is treating escalation as a safety net rather than a feature, which produces a handover so poor that the humans receiving it stop trusting the queue and start ignoring it. Design what the person on the other end actually sees. The other failure is refusing to write the accountability sentence because it is uncomfortable and political. That discomfort is the job. Someone will answer for it, and if it is not written down it will be whoever happens to be nearest.