Home / Guides / Operations on AI Employees

Operations on AI employees.

Three seats. Fixes and ships. Spins up offers. Keeps the fleet up. Operations is the department most founders accidentally outsourced to themselves on nights and weekends. These three take it back.

In a $5-50M business, operations is where every small thing lands. A redirect that stopped working. A customer who cannot log in. A webhook that fires twice. A page typo on a lead magnet. None of these are big; all of them have a cost of carry; the person solving them is almost always you. Three seats take the load.

Seat 1: fixes and ships

A support ticket becomes a verified fix, a pull request, a merge, a release and an email to the person who reported it. The loop runs on one role. Nobody waits for a developer.

Read the ticket.

Open the inbox, pull the ticket, check whether the issue is already in the backlog, check whether a fix is already in flight. If yes, link them. If no, open an investigation.

Reproduce it.

Run the sequence the customer described. Capture logs. Confirm the bug or label it “cannot reproduce” with the exact steps tried. “Cannot reproduce” is a legitimate answer and goes back to the customer with the ask “please tell me which step on this list looked different.”

Fix it, or escalate it.

The rules say which issues the agent is allowed to fix without a human. Copy, redirects, config, static assets, and anything with a clear existing pattern. Issues outside that envelope get a decision brief, not a commit.

Ship it.

Open a pull request. Write a plain-English what-changed. Run the tests the project has. If green, merge. The merge triggers the deploy.

Close the loop.

Email the customer who reported the issue with the plain-English what-changed and a thank you. The ticket closes.

Seat 2: spins up offers

This whole site is an example of the output. Five domains bought, five sites live, redirects verified, legal pages updated, inside one afternoon.

The seat runs through a checklist: register the domain, scaffold the site in the monorepo, write the copy from one belief, configure the subscribe endpoint, create the Resend audience and the drip, enroll a test subscriber, verify the end-to-end delivery, attach the apex and the www redirect in Vercel, add the entry to the legal pages, update the domain ledger and open the pull request. Every step is a rule; the rule says what counts as done.

Seat 3: keeps the fleet up

Sixty-odd customers each run their own copy of the same employee on their own server. Each copy is a Railway service. Each service has logs, memory and a CPU curve. The seat watches all of them.

What watching looks like: every six hours, read the logs for errors, read the CPU for runaway processes, read the memory for leaks. If anything crosses the thresholds the owner set, surface a decision brief with the service name, the metric, the number and the suggested fix. If the fix is in the envelope of what the agent is allowed to do (restart, redeploy, roll back), do it and log the action.

What stays human

• Architecture decisions. No agent redesigns a system.
• Hiring tool choices. No agent adds a dependency or switches a service.
• Permissions and secrets. No agent provisions access it was not granted.
• Anything that touches money with new state. New Stripe products, new price points, new refund rules: owner only.

The compounding effect

Marketing hires are visible. Operations hires are not. Nobody writes a case study about the day the webhook stopped crashing. The payoff is quieter and larger: the founder stops being on call, the customers stop writing in about the same five things, and the time that was spent fixing is spent building. Over a quarter, operations on AI employees is usually the department that returns the most time to the person at the top.

“Operations is the department you cannot see until you have one, and once you have one you cannot imagine running without it.”

Where to start

Fixes and ships first. It is the seat that has the clearest loop (ticket in, fix out) and the easiest ROI (how many tickets closed in a week, how many did the owner touch). Spin-up and fleet watching come after, because they lean on the discipline the fix loop forces you to install.

Frequently asked questions

What if the fix is wrong?

The pull request is the review surface. For the issues inside the agent’s envelope, the test suite is the first guard; for anything else, the agent writes a decision brief and the owner reviews before the merge.

Can an agent really deploy to production?

Yes, when the deploy pipeline is a push to a branch, which is the industry default. The agent commits, the pipeline runs, the agent reads the pipeline result. If the pipeline fails, the agent surfaces the failure and does not merge.

What happens if everything breaks at once?

The fleet seat raises a priority flag, pauses its own actions, and asks for a human decision. Automation that cannot stop itself is not automation.

Do I need to be the developer to run this?

No. You need to know what counts as done in your business. The agents run the deploy; you review the plain-English what-changed.