Friend Anan panicked for half an hour — his AI sent 12 emails for him

Anan runs a study-abroad consulting service. He usually has his AI assistant draft replies for him. The other day he accidentally triggered the AI to send all the drafts — one of them had the client's name wrong. The client came back asking "what exactly did you promise me?", and Anan tried to find the records. The chat interface only shows the most recent message; the historical actions were scattered across his email, the AI backend, and IM. He spent half an hour digging through email before sorting it out.

I've been stuck in the same place too — once I had AI sync my Notion and email, and it changed a bunch of stuff. I never saw the process, so I ended up guessing line by line.

OpenAI hit this same wall: AI acts, evidence doesn't stick

A recent OpenAI incident pushed this into the spotlight: when their AI agents operate external services on someone's behalf, pulling up evidence afterwards is painful — data may be fragmented, may be private, or may be wiped from temporary caches.

If even a company like OpenAI is struggling with this, those of us "using AI to do our own work" face an even higher chance of hitting it. Who's already stepping in it? A friend doing e-commerce has AI auto-reply to bad reviews; a content creator friend has AI auto-publish. The moment AI makes its own decision and messes up, it's nearly impossible to reconstruct.

What you can do today — minimum cost: $0

  • Cost: $0
  • Time: 30 minutes today
  • Technical barrier: copy-paste skills
  • First step: open your notes app, create a new document

Do one specific thing: write three columns — "What AI did / Where to check / Who's responsible." Every time AI auto-sends something for you (email, message, comment, contract), copy the original + timestamp into this doc. Scan it 5 minutes a week.

Not everyone needs this tool — if you only use AI to draft and you manually send it, skip this for now.

Advice by stage

Just starting out: don't bother with systems yet. Just screenshot what AI did and save it to an album. Once you're using AI 10+ times a week, come back and set this up.

Got 1-2 clients: spend those 30 minutes now building the three-column log. It'll save you a lot more than "Anan's half-hour email dig." One accountability call from a client is enough to break you.

Scaling up: if AI is already talking to 5+ clients for you, try Zapier or Make for the simplest auto-log, or block 15 minutes weekly for a "review meeting" — go through the evidence list together.