I Installed 3 AI Colleagues. Two of Them Stopped, and Nothing Told Me.

Three weeks ago I published my own AI Brain diagnostic score, 10 out of 18, along with the fix I’d installed the same day. Since then, two of the three workflows I put in have stopped running. Nothing told me. I found out because I went and looked.

The part worth writing about isn’t that they stopped. Everything stops eventually. It’s that stopping made no noise at all, and I’m the person who sells this for a living.

What went in

On 17 July I scored 0 out of 3 on Archetypes, the category that asks whether any named AI role does its job without being asked. The score was zero, so I installed three roles, each one with a written job description and exactly one weekly job it owns.

An operations manager runs a weekly operating review every Monday morning, unattended, and hands me one brief on cash, pipeline and the week ahead. A content creator drafts one blog post and one LinkedIn piece every Wednesday. A founder challenger argues with me about one live decision a week, and that one is interactive by design, so it needs me in the room.

Each pair got a done-check written before the install went in: the thing that has to be true before I’m allowed to call it working. The operating review needs 2 consecutive unattended Mondays. The content run needs 2 posts shipped with edit passes under 30 minutes each. The decision session needs 2 sessions with at least one decision materially changed.

19 days on, 0 of 3 have passed.

The first failure was an absence

On Monday 27 July there was no operating brief. On Wednesday 29 July there was no draft.

Nothing errored. Nothing went red. Both workflows write a dated file into a runs/ folder when they run, which means the entire evidence of a miss is a file that isn’t there. A file that isn’t there doesn’t notify you, doesn’t appear in a log, and doesn’t interrupt your Monday.

I didn’t notice on the Monday. I noticed on the Thursday, and only because I’d sat down to do something else with the repo.

The second failure was worse, because it kept running

When I dispatched the catch-up run by hand on 30 July, it found something I’d have preferred not to know. The operating review had been reading 1 of my 13 calendars since the day it was installed.

Every week it had reported a diary of gym slots and daily practice with almost no business on it, and I’d read that as an accurate picture of a quiet week. The client meetings, the cal.com bookings and the fractional CTO sessions all sit on a separate work calendar the job had never been pointed at. It ran perfectly. It was confidently wrong every single time, in exactly the same way.

That’s the more dangerous of the two failures. A stopped job gives you nothing, and nothing is at least honest. A job reading a partial source gives you a clean, well-formatted, plausible brief, and you make decisions off the back of it.

Why nothing told me

The likely cause is dull: these jobs are scheduled tasks that fire while the app is open, otherwise on next launch. If the laptop is shut across the slot, the slot passes in silence.

I can’t prove that’s what happened on 27 July. I can prove it for the one after. The 30 July catch-up brief predicted in writing that the Monday 3 August run would miss too, because I’d be away from Friday evening until Monday evening with the machine closed. It missed. That week’s brief eventually ran on the Tuesday, dispatched by hand, and opened by telling me Monday was already behind me.

Here’s the part that stung. My day-0 diagnosis, in my own words, was that I’m the dispatcher: nothing runs unless I press the button, so my AI capacity is capped by my attention rather than my subscription. The whole install existed to fix that one thing. A schedule that only fires when my laptop is open still has me pressing the button, later and with more steps in between. I moved the dependency. I didn’t remove it.

Built, running and proven are three different states

I’d been treating the install as the finish line. It’s the start of the only part that matters.

Built means the configuration exists. The role has a job description, the workflow has a brief, the folders are there. This is the state most AI setups reach and stay in, and it’s the state that demos well.

Running means it produces output on its own schedule, without you starting it.

Proven means it has run unattended often enough that you’d notice if it stopped.

Most of what gets called an AI system is sitting at the first state. Mine reached the second for about a week and then slid back. The only reason I can tell you that precisely is the one thing I got right on day one: every workflow writes a dated file every time it runs, so the run history is a folder I can count. Without that folder I’d have a vague sense that things had gone a bit quiet, and no way to check.

4 things I’d do differently on day one

Make every run leave an artefact. A dated file per run, written even when the run finds nothing worth saying. “Ran, nothing to report” is data. Silence isn’t.

Write the done-check before the install, with a number and a date in it. “Runs unattended” is a hope. “2 consecutive Mondays with a file in the runs folder” is a check, and it either passed or it didn’t. Mine were written down in advance, which is the only reason I’m able to tell you I’ve passed none of them rather than telling you it’s going well.

Monitor the absence, not just the output. Something has to go looking for last week’s file and complain when it isn’t there. I have alerts watching competitors, inbound leads and overdue invoices. Every one of them watches the outside world. Not one of them was watching whether my own workflows had run.

Refuse any trigger that depends on you being present. If it only fires when your laptop is open, you’re still the trigger, and you’ll find out it stopped at the exact moment you most needed the output.

What I’m not going to claim

The running ROI table in my repo has one row per workflow and an empty cell in every hours column. Not roughly estimated. Empty. The 20 hours a week I say AI gives me back today is my own aggregate estimate across everything I do, mostly Claude Code on engineering and site work, and explicitly not these three workflows. The 40 I’ve bet on after a full install is a bet, and I’ve labelled it as one.

I’ve also never once timed my edit pass on a drafted post, which is the single number this content pipeline exists to move. Two posts in, zero timings. That makes it the weakest evidence in the whole install, and the cheapest thing on the list to fix, because it needs a timer and nothing else.

I could write the case study today, and it would be dishonest. It would describe a system that has run properly for 1 week in 3, with blanks where the returns go.

The honest scoreboard

This post came out of the Wednesday run, on time, 1 week after that same run missed its slot. That’s 1 Wednesday. The check needs 2, and until it gets 2 I don’t get to claim anything.

If you’d rather install this on your own business than read about mine going sideways, that’s what the Build Your Own Brain course does over 6 weeks: your baseline scored, your workflows tuned, your roles written down, all of it in a repo you own and can rebuild from when the tools change underneath you. You’ll also get my done-checks, including the ones I’m currently failing. Join the waitlist.