When not to hire an AI employee

When should you not hire an AI employee?

When you cannot describe the job to a new person in an hour. If a smart intern would get lost, a model will get lost with better grammar. Write the job. Then talk about hiring.

When the close action can hurt someone and you have not named the human who must approve it. Hurt means money they cannot easily get back, care they should not take, or a legal statement you will have to defend.

When the real request is 'make us look current'. I will not help you hire a costume. I will help you hire a process. If the board wants a costume, that is a communications problem. It is not an engineering one.

What if the process lives in one person's head?

Then you do not have a process. You have a specialist. Specialists are valuable. They are also a single point of failure. Automating them without writing the rules copies the failure into software.

I have been that specialist. It feels like speed. It is actually a queue that only you can drain. The day you are sick, the company learns what was never written.

The work before the hire is interviews and a one-page SOP. Inputs. Exceptions. The three cases that always go sideways. If the specialist will not sit for that, they are not ready to be cloned. Respect that. Pay them. Do not sneak a model around them.

What if the decision is medical, legal, or a refund?

Then the AI employee may draft and sort. It may not close. I treat this as a hard gate, not a confidence score. Scores drift. Names on a rota do not.

At 10mg the path that matters ends in care and credit. I will not let a model tell a provider something that sounds like a clinical call. I will not let it move money past the point a person must see. That is not caution for a blog. That is how you still have partners.

Refunds look smaller and they train bad habits. If the system can issue money, someone will raise the limit because the queue is long. Write the limit. Write the person. Write the after-hours rule. If you cannot, do not connect the wallet.

What if you want a platform, not a queue?

Say that. Then go through procurement and be honest about the year it will take. Do not hire 'one AI employee' as the thin end of a company-wide rollout. The thin end will be a mess, and the rollout will inherit it.

Platforms need owners in every function, a data map, and a governance story. Half the people in a large 2026 survey still say they have no clear governance for agents. I believe them. I have sat in those meetings. The slide exists. The rota does not.

One process can be done in a week if the policy exists. A platform cannot. If your calendar says platform and your patience says week, you will get a demo environment. Demo environments do not drain piles.

What if nobody will sit the misses?

Then you will get silent rework. People will redo every case and tell you the tool 'is not ready'. They are right. A tool with no reviewer is not ready, even if the model is fine.

The reject queue is a shift. It needs a name, a backup, and a time box. If you cannot staff that, keep the original team on the full pile. You have not failed. You have avoided a fake win.

I would rather a company stay manual for a quarter than ship an unsupervised worker on customer money. Manual is honest. Unsupervised is a story you tell until the first screenshot.

• No SOP: wait.
• No owner: wait.
• No gate on harm: wait.
• No time for rejects: wait.
• Hiring or credit decision with no audit path: wait, and read the law.

What should you do instead this quarter?

Write one process. Count the cases. Count the minutes. Write the stop list. Name the reviewer. That document is the hire. The model comes after.

If leadership still wants motion, pick a pile that is boring and low harm. Appointment reminders. Document field pull with a person on every file. Internal routing. Learn your miss rate where a miss is an apology, not a lawsuit.

If you want a second opinion on whether to wait, that is a 30 minute call. I will tell you no if the process is not real. That is the useful answer. A yes on a slogan is how you waste the next two months.

Write the no in the same document you would have used for the yes. Process name. Why it fails the test. What you will write down this month. When you will look again. A dated no is a management artifact. A vague delay is how the same idea returns in six weeks with a new vendor.

What does an honest no look like in writing?

It is short. 'We will not hire an AI employee on refunds until Amina owns a nightly reject queue and the limit is written. We will write the SOP for intake this month. We will look again on the first Monday of next quarter.' A board can live with that. A team can work against it.

It names the missing piece. Process, owner, gate, or law. If you hide the missing piece, someone will fill it with a pilot that has none of them. Pilots without a missing-piece list are how you get a fluent miss and a retrospective that blames the model.

It does not insult the people who wanted the hire. Wanting a smaller pile is rational. The no is about unreadiness, not about fashion. I want the pile smaller too. I also want to sleep. Those are the same goal if you wait until the page exists.

If leadership still wants motion, give them the low-harm pile I already named. Motion that cannot hurt someone is still motion. It is also practice for the rota. Practice is allowed. Pretend production on refunds is not.