I Gave AI Bots Real Jobs. Here's The Org Chart.

You pay for five AI tools. Maybe more.

Ask yourself who checks the work of each one. Not who set it up. Who looks at what it produced this week and says yes, that’s right.

If you run a service firm with somewhere between 5 and 50 people, I’d bet the answer is nobody. Or it’s you. Which, on a busy week, is the same thing.

Tell me if any of this sounds familiar:

• A tool tells you a job is done, and nobody opens the result.

• Every reply, every post, every proposal waits for your approval, so nothing moves while you’re in meetings.

• You keep adding tools because the last one didn’t fix it.

By the end of this issue you’ll have a one page org chart for your AI tools, and a 20 minute way to fill it in tonight.

I Run My Business On A Team Of Bots

Some context first.

I run Robossist, a small automation practice in Hong Kong. Behind it sits a team of AI agents. Each one has a seat and a job. One handles sales research. One runs my X account drafts. One runs this newsletter. One makes the illustrations. One is a red team whose only job is to find holes in the others’ work.

Today the chart has 17 seats, and 15 of them have a bot working in them.

That sounds impressive. It wasn’t, at first.

For the first few weeks I treated these bots the way most firms treat AI tools. I bought them, switched them on, and assumed the work was happening. Then the receipts came in.

Four Times It Broke

These are real. I’m putting them here because you will hit the same ones.

1. Eleven days of “done.” One bot reported its work as complete for 11 days straight. A good part of that work never showed up where it was supposed to land. I wrote the full story in Issue 2. The short version: “done” was a status message. Nobody had defined what done looked like from the outside.

2. The tool that ran past its cap. A paid scraping tool hit its monthly limit. Everything that depended on it stopped. Nothing told me it would happen, and nothing stopped it from happening again until I wrote a freeze rule. The tool now has a written spending limit, a named owner, and a rule that every other bot stays off it until I lift the freeze.

3. Routines that died quietly. Some of my scheduled jobs, like the daily X brief, failed when the account ran out of usage. Nobody was told until the next morning, when the work simply wasn’t there. This week my newsletter’s morning routine skipped a run, and nothing alerted me. It ran fine on Wednesday.

4. The bottleneck was me. This one stung. Every external post, every email, every reply waits for my yes. That’s the right rule. But with 15 bots producing drafts and one human approving them, the queue piled up behind me. More bots would have made it worse. Every new seat adds drafts to the same pile, and the pile has one reader.

“Adding a bot without a boss is hiring someone and never telling them who they report to.”

Look at those four again. None of them is a model problem. The AI wrote fine. What was missing each time was a person or a rule around it.

The Framework: Job, Check, Boss, Bench

After those misses I gave every seat four things. It’s the same thing you’d give a new hire, and it’s the thing almost nobody gives software.

• Job: What is this tool for? One sentence, tied to money. “Replies to every new enquiry within the hour.”

• Check: How will I see the work landed? A place you can look: the CRM record, the sent folder, the live page. Not the tool’s own status.

• Boss: Who approves and who gets told when it breaks? One named human (or one named bot that reports to a human).

• Bench: When does it get paused? A written rule. Mine: no accepted work in 7 days and the seat is benched.

A few notes on each.

Job has to be narrow. “Help with marketing” is not a job. “Draft two LinkedIn posts a week from last week’s client calls” is a job. If you can’t write it in one sentence, the tool is doing too many things, or nothing.

Check is where most firms fail. A check is something outside the tool that proves the work exists. In Issue 2 I called this output liveness: go look at where the work should be, not at the tool that claims it did it.

Boss fixes the quiet failures. When my routines died, the failure had nowhere to go. Nobody was named to hear about it.

Bench is the one people skip. Without a bench rule, tools never get switched off. They just keep billing you.

What The Chart Looks Like Filled In

Here are three of my real seats, written the way I’d want you to write yours.

The newsletter seat.

• Job: ship one to two issues a week and daily Notes that bring in readers who might buy.

• Check: the scheduled post on Substack itself, with a screenshot. Not the bot saying “scheduled.”

• Boss: me. Nothing publishes without my yes.

• Bench: the standard rule. No accepted work in 7 days and the seat is reviewed for a pause.

The sales research seat.

• Job: build a short list of firms that fit my buyer profile, with a reason for each.

• Check: the list file exists, and each row links to the page it came from.

• Boss: me. The red team seat can challenge any row before it reaches me.

• Bench: same 7 day rule. A list nobody uses is not accepted work.

The scraping tool.

• Job: pull public pages for research, inside a monthly budget.

• Check: every run is logged with the credits it used.

• Boss: me, for any spend. Frozen right now until I lift it.

• Bench: already benched once. That’s what the bench is for.

None of this is clever. It’s the paperwork you’d do for a new hire, done for software.

What This Means For A 5 To 50 Person Firm

You probably don’t have 15 bots. You might have a chatbot on the site, a meeting notetaker, an email assistant, a CRM with “AI features,” and a couple of automations someone built two years ago.

That’s still an org chart. It’s just one nobody drew.

The place it costs you most is usually the front door: enquiries. A lead fills in a form, the automation fires, it says sent. Did anyone call them? I’ve sat in client calls where warm leads landed in a chat channel and never reached the CRM at all. Nobody noticed, because every tool in the chain reported success. That team ran a small sales crew in Australia. Their fix had two parts. First, a rule: every lead that shows interest gets a call within six hours. Second, a check: a list of everyone who had shown interest and never got that call. The list was the uncomfortable part, because every name on it was a person who had raised a hand.

The same thing shows up in recruiting. On a recent call, a small search firm described building a shortlist by hand. They search roughly 200 profiles, download them, paste them into an AI tool to screen, then build a list of about 20 by hand. The AI step works. Everything around it is a person copying files. There’s a job in there, but no check and no boss, so it all lands on whoever has time.

That’s the pattern. Each tool does its piece. Nobody owns the handoff.

Do This Tonight: 20 Minutes, One Page

Grab a notebook. Then:

1. List every AI tool or automation you pay for. All of them, including the ones you forgot about.

2. Write its job in one sentence. If you can’t, put a question mark.

3. Name its check. Where would you look to prove the work happened this week?

4. Name its boss. One person. Not “the team.”

5. Circle the one with no check. Fix that one first. Just that one.

Give it 20 minutes, not an afternoon. You’re looking for gaps tonight. Fixing comes later.

Most people find at least one tool with no check and no boss. Very often it’s the one touching leads.

What I’d Do Differently

If I were starting the team again:

• Fewer seats, earlier. I’d run five bots for a month before adding the sixth.

• Red team every new seat before it touches anything external.

• Count accepted work, not activity. A bot that produced 40 drafts and got 2 approved did 2 things.

• Plan for the human queue. I now batch approvals into two sittings a day. That one change moved more work than any new bot.

Some of you will do the notebook exercise and be done. Good.

Next Week

Tuesday’s issue picks up the enquiry thread: the lead that landed in a chat channel and never got a call, and the simple rule that fixed it.

Talk soon,

Dhruv