What an AI employee cannot do

What can an AI employee not do?

It cannot know a rule you never wrote. It will guess. Guessing looks like policy when the sentence is clean. I have read those sentences. They are how a team ships a rule that nobody voted for.

It cannot hold a relationship. A clinic manager who trusts your team trusts a person who picked up when things broke. A worker can send the reminder. It cannot sit in the awkward call after an outage.

It cannot be the adult in the room. Adults decide what 'good enough' means when two policies conflict. If your company runs on conflict, you need people, not a tie-breaker with a temperature setting.

Can it invent or update policy?

No. It can show you that two documents disagree. A person updates the page. If you let the worker 'just handle it', you have given it a legislature.

This is why I want structured output and a policy id on every close. If the id is missing, the close is invalid. That sounds strict. It is how you stop a silent rewrite of the business.

When I did review work on compliance documents, the win was not a model that 'understood regulation'. The win was fields a person could check, and a pile that got smaller. Understanding is a story. Fields are a job.

Can it take hiring or credit decisions alone?

Not in any company I want to be associated with. Software that ranks people for jobs is high-risk work. New York already asks for bias audits on certain tools. Europe puts recruitment AI in a high-risk bucket. None of that is an exemption. It is a spotlight.

Credit is the same shape. At 10mg, a loan is not a chatbot opinion. It is a path with data, a decision, and a disbursement that has to land. A model can assist the path. A person and a written standard own the decision. If that slows you down, good. Speed that skips a standard is how you lose the book.

If a vendor says their agent is 'fair' and therefore fine, ask for the audit, the adverse-action story, and the name of the human who explains a no. If they stall, you have the answer.

Can it replace a person who holds context?

Context that never hits a system is invisible. The salesperson who knows the customer is about to churn because of a fight last year. The pharmacist who knows the generator failed on that street. A worker cannot apply what was never written.

You can force more context into tickets. That is a management job. Until you do, the AI employee will be cheerfully under-informed. Do not fire the person who held the context and then act surprised.

I sold Insurpass and Shopl. The due diligence questions were about whether the company still worked if a person left. If your plan is that the model holds the context, you will have a short diligence. Write the context down or keep the person.

Can it sit outside the law?

No. There is no special lane where a model can discriminate, invent consent, or skip a record you would owe if a person did the same task. If a person would need a log, the worker needs a log.

Teams sometimes hope the vendor contract moves the duty. It does not move it in the way you want. You still have the customer. You still have the regulator. You still have the screenshot.

I am not a lawyer. I am an engineer who has watched teams try to be clever. Clever is expensive. Boring compliance is cheaper than a clever miss.

• Hiring tools: audits and high-risk rules may already apply.
• Health and credit: your existing duties did not shrink.
• Customer messages: you said it, even if a model typed it.
• Records: if you cannot replay the case, you cannot defend it.

What should you still ask it to draft?

The first pass of the pile. The summary a reviewer can scan. The fields that always appear. The routing that sends medical questions to a clinician and billing questions to billing.

Drafts are a gift if a person still sends them. I will take that gift. I will not pretend the draft is a decision because the sentence is polished.

If you only remember one limit, remember this: an AI employee cannot take your name off the work. It can take hours off the easy part. Keep your name. That is the job.

Draft the reminder. Draft the field pull. Draft the internal note. Then look at the send button and ask who is named next to it. If the name is missing, you have crossed the limit even if the sentence is fine.

What happens if you ignore these limits?

You get a sentence that sounds like policy and is not. A customer keeps it. A regulator asks for the rule. You cannot produce the rule because it was never written. You produce a chat log. Chat logs are not policy.

You get a hiring or credit path that looks faster and is harder to defend. The law did not get softer because the ranking was a model. The person who used to sign the no still has a duty. Hiding them behind software makes the duty uglier, not smaller.

You get a team that stops writing things down, because the worker 'just knows'. Six months later nobody knows. I have inherited systems like that. They are expensive. The cost is not tokens. The cost is the week you spend reconstructing why a decision happened.

Ignore the limits and you will still have an AI employee on a slide. You will not have one in the sense I use. You will have a fluent intern with production keys. I will not help you hire that. I will tell you to take the keys back.