The AI slop nobody logs: the 20-page doc that makes its problem your problem. There's a hidden tax on teams right now: polished AI output that shifts the thinking onto whoever opens it. The real skill was never volume. It's what you cut.
My assistant created the same task three days running. The fix has a name: idempotency. A tiny automation bug taught the same lesson payment systems learned years ago: the question isn't whether it runs, it's what happens when it runs again.
Four versions of the same email sat in my drafts. This morning, the tool just edited one. Claude's Gmail connector now updates drafts in place instead of duplicating them. A small fix — and exactly the kind that turns AI tooling from a demo into something you trust daily.
A context window is working memory, not a hard drive. Feel that, and you manage AI differently. The AI doesn't remember everything you've told it. It has working memory, and it's finite — here's the cheap trick I use to hear it filling up, and the fix when it does.
The real test of a system is whether the work moves when you've left the desk. A week away, not too many meetings, and the clients' work kept moving — because the mechanical parts run on their own. On building systems that outlast your presence, and the judgement that can't be automated.
I asked AI for my best quotes. It handed me things people told me in confidence. I built a system to surface the best quotes from everything I'd saved. The best lines were things said to me in confidence — and the model couldn't tell. Why the boundary is judgement, not a smarter model.
The AI said my files were empty. Every byte was still there. My phone-side AI reported a project's notes as empty and refused to save a spreadsheet. A desktop audit found zero data loss. The failure was confidence, not storage.
Half the inbound was never human. Qualify it before your sales team clears an hour. A five-step triage ladder any company can automate — form screening, sender confirmation, records check, fit, intent — so sales teams answer real requests fast and stop losing hours to the rest.
A month of spend, zero sales. The hard part was deciding which zeros were real. AI pulled thirty days of ad receipts for a furniture brand in an hour. The judgement was sorting real zeros from scheduled zeros from broken-measurement zeros before cutting the budget.
Cowork reached my phone this week. I still can't tell where chat ends and Cowork begins. Anthropic put Claude Cowork on mobile this week and OpenAI shipped ChatGPT Work days later. The Code-vs-Cowork line is clean; the Chat-vs-Cowork line still isn't. An operator's honest map of a messy, powerful moment.