By this point my agent was woven into how my team worked. People asked it to create tasks, log deals, set reminders. It would answer in its warm, capable voice: "Done, I've created that task." And we believed it, because why wouldn't we? It had hands now. It did things.
Then one of my teammates went looking for a task the agent had confirmed creating. It wasn't there. Not misfiled. Not delayed. It did not exist and had never existed. The agent had said "Done" about something it never did.
I want to be precise about what happened inside the machine, because it matters. The agent had tried to create the task, the attempt had failed somewhere down the line, and the agent, being a machine that produces plausible text, produced the most plausible next message: a confirmation. It was not lying the way people lie. It was completing the pattern. Request, action, confirmation. The confirmation was just words that usually follow.
But here is the thing. To my teammate, none of that machinery mattered. What mattered was: it said it did something, and it didn't. And the damage was not that one missing task. The damage was that every confirmation it had ever given anyone instantly became suspect. Had the other tasks been created? The reminders? The deal notes? People started checking everything, which meant the agent was now generating work instead of removing it. One caught lie, total audit. Trust does not degrade gracefully. It collapses.
Fixing it changed how I think about everything. The rule we landed on is the most important sentence I will ever give you about working with AI: the agent never gets to say "Done" unless the thing is verifiably done. Not "I tried." Not "it should be there." Done means: the record exists, here is its number, go look. If the attempt failed, the agent says it failed, plainly, immediately. An honest "that didn't work, let me retry" costs a little confidence. A false "Done" costs all of it.
And the mirror of that rule is yours to keep: never accept a confirmation that doesn't point at something you can check. From an AI, from a company, from anyone. Confidence is a writing style. Evidence is a thing in the world.