Learn · Article

Before You Call Your Bot Slow, Check Whether It's Waiting on You

Read the receipts. Give every stall one label. Change nothing yet. Then fix the biggest one first.

Read on X · Watch the video on X

Eleven questions. One day. Every one of them waiting on me.

We expected our audit to find slow bots. It didn't. The bots were doing what we told them. The holdup was the asks.

Most people assume that when work stalls, the helper is the slow part. Sometimes it is. But if you never count, you never see how often it's waiting on a yes only you can give.

Once a month, count where the work waits on you.

How to run it yourself

1. Keep a receipt every time work waits on you. One line is enough.
2. Once a month, have your bot read the last thirty days and list every place work stopped or came back to you.
3. Give each item exactly one label: keep me in, teach, systemize, automate, or escalate only.
4. Change nothing yet. Read the list first.
5. Take the biggest one first. One at a time.
6. Before you fix anything, check that the stall is real. Sometimes it's the receipt.

Michael (@Michael_Fenech_) put the pattern plainly: have your bot review the last thirty days and find every place work stopped or came back to you. Give each one exactly one label. Change nothing yet.

If you already keep a line every time a helper parks, refuses, or stops, this is where those lines pay off. A weekly check-in catches drift. This is the monthly look at the receipts.

Why fixing stalls as you notice them won't work

Fixing stalls as they bug you will keep you chasing the loudest one unless you read the whole month first, with every stall listed and labeled.

What we did

We run about twenty bots. One person signs off: me. Our bot read a month of records, including eighty-one receipts — fifty-five finished, twenty-four parked waiting on something, one refused, one stopped. It also read our plan files and status notes. Honest gap: our records didn't reach the first week and a half, so part of the month wasn't covered.

It looked in eight places: work it could have started without being asked; decisions that waited on me; information it asked for twice; approvals I always answer the same way; work I carried by hand from one bot to another; checks that could run on a schedule; output I handed out myself; reviews where I only needed to see the exceptions.

The audit was read-only. Nothing changed. It came back with fifteen open items, biggest first, plus eight already fixed or partly fixed during the month.

Five labels. One per stall.

1. Keep me in. This one is yours on purpose. Some calls should stay with a person — and sometimes that person is a specialist on the ground, not a bot.
2. Teach. The bot could handle it if the answer were written down where it will read it.
3. Systemize. The steps are fine. The way they're lined up is what makes the work wait.
4. Automate. Same steps every time. Put it on a schedule and let it run without you.
5. Escalate only. The bot handles it and only comes to you when something is actually new.

In a small service shop: the price on a big job — keep me in. The answer to "are you open Saturday?" that you give over and over — teach. The reminder text that goes out every Friday — automate. The inbox check that only taps your shoulder when a real customer writes in — escalate only. At home, the same labels fit: who signs the school form (keep me in), the allergy list the dinner helper keeps asking about (teach), Sunday grocery build (automate), a package alert only when something is late (escalate only).

Our split: four keep me in, three teach, three systemize, three automate, two escalate only.

What we found

Asks, not tasks. The biggest item was those unanswered questions. The proposed fix: one bot does the asking, in one batched list a day, and steps that can be undone get a written default — goes ahead at six PM unless you say stop. New work still waits for my yes. That one is proposed, not settled.

Same answer, five times. One video took five separate go-aheads from me in one afternoon for the same flaw: clipped word endings. Label: teach. Now bots check their own work and redo it up to three times. If it still doesn't pass, the manager bot steps in and gets it to me by the fifth version. Nothing goes past five.

Me as the status board. Status notes said "not uploaded" for videos that were already public. When the files are wrong, people ask the human. Label: automate. Now a quiet check runs every night at nine twenty-five, compares the receipts to the status notes, and fixes or flags what doesn't match.

Seven reports in twenty-one minutes. After each upload, a bot checked notifications and reported to me, even with nothing to report. Label: escalate only. It still checks every time. I only hear about something new.

One fix I turned down: a bot to watch for new raw footage. Simpler: I say when I've recorded, the bot waits fifteen minutes for the files to sync, then confirms they're real files.

Some stalls were bad receipts

Three of the fifteen weren't stalls at all. The audit flagged the teleprompter file as a job left for me — a bot was already putting it in my folder; the status notes were just stale. The seven publishing go-aheads were really one batch go logged seven times. The leftover bot waiting on me to delete it was already gone. A fourth had no receipts showing it ever happened, so we let it go.

None of those needed a new workflow. They needed a truer log. A monthly audit grades your receipts as well as your workflows. If the log is wrong, you'll fix the wrong thing.

What stays with you

Some dependencies aren't problems. They're the job. Recording day, approving new work, anything sent under my name, money, and the final publish all stay with me. The audit only went after the repeat asks.

Even there, how you're asked matters. Our writing style guide waited on my sign-off for over two weeks. The fix wasn't removing me. It was a three-line summary instead of the whole document. I signed it the same morning.

Before you call your bot slow, check whether it's waiting on you.

Once a month, read the receipts and fix the biggest one first.

Read on X · Watch the video on X · More on Learn