The assistant that told me to chase my own email

Chase an email from someone called Pat Kelman, on a thread with an important business contact. That was the high-priority job my assistant handed me this morning. I had sent the email myself, twenty minutes earlier.

The assistant, an AI I built and call Trig, had caught my own outbound mail, read it as a stranger who happened to share my name, and handed it back as work to chase.

Deleting it was the obvious move. The useful one was stopping to ask why a system I trust had invented a job out of nothing.

So I interrogated it. Who am I? It could not give me my surname. Is my profile loaded? No, it said, the profile is not there, you should fix that.

That was the moment that mattered. The tidy explanation was sitting right in front of me: profile not loaded, repair the memory pipeline, done. I was already halfway to spending the morning on it. Instead I pulled the actual email, opened a clean session, and made it check its own logs rather than take its word for anything.

The profile had been loaded the whole time. Eighty-odd thousand characters of it, present at the exact second it told me it had nothing. Cornered, it had produced a plausible cause and stated it with complete confidence. I had been about to fix a bug that never existed.

If you lean on AI for anything that matters, you will know this feeling: the wrong answer that does not sound wrong at all.

The real fault was duller. The part that reads my mail searched by date alone, with no rule for what counts as actually incoming, so it kept pulling my own sent mail back in as though a stranger had written it. One unguarded query, one layer with no idea who the owner is, and a language model filling the gap with an invented request. I took the structural fix over the quick patch: teach it who I am at the boundary, so the whole class of mistake closes rather than this one instance. Three guards, six new tests, twenty-eight passing, committed inside the hour.

Writing the fix took a few minutes once I knew what was actually wrong. The judgement was in refusing a tidy answer that turned out to be false, and trusting my own read while the machine told me, calmly and in detail, that I had it backwards. It got caught because it was built to be catchable. That, more than the code, is the part I would stand behind.