I love AI agents. I just don’t trust them with everything.

“Can we let the AI agent handle the whole process?”

Maybe.

“What process?”

“Customer refunds.”

Okay. Now we’re having an interesting conversation.

I like AI agents.

I like that they can understand instructions, work with information, and help complete tasks that previously needed several applications and a very patient human.

But there’s a difference between asking an AI agent to summarize a refund request and allowing it to issue a $10,000 refund.

One helps someone make a decision.

The other moves money.

And that difference matters.

Imagine a customer service team handling refund requests.

Today, an employee reads the complaint, checks the purchase, reviews company policy, and decides what happens next.

Some requests are simple. Others involve damaged products, missing documentation, suspected fraud, or customers who have already received a refund.

Now imagine adding an AI agent.

It reads the complaint, identifies the order, checks the policy, and recommends a response.

Excellent.

Should it also approve every refund?

I’d ask a few more questions first.

• What happens when the customer’s message contains misleading instructions?
• What if the AI misunderstands the policy?
• What if the order belongs to another customer?
• What if the refund exceeds the employee’s authority?
• And who is accountable when something goes wrong?

This is why I prefer deterministic-first systems.

That’s a technical phrase for a fairly simple idea.

If a business rule has a clear answer, use the rule.

If a refund must never exceed the original payment, enforce that in the transaction system.

If a payment above a certain amount requires approval, make approval mandatory.

If an employee isn’t authorized to view a customer’s information, an AI agent shouldn’t get around that restriction.

AI can help interpret information and recommend what happens next.

But the actual business transaction should still follow the same rules, permissions, and checks that apply to everyone else.

Here’s the part I find interesting.

This approach doesn’t necessarily make AI less useful.

It can make AI much easier to adopt.

A business is more likely to trust an agent when it knows what the agent can do, what it cannot do, and how mistakes will be handled.

And the controls don’t have to be the same for every action.

Summarizing a customer complaint might need no approval.

Sending a routine acknowledgment could be allowed automatically.

Issuing a large refund might require a supervisor.

Different risks. Different boundaries.

That’s human-in-the-lead.

It doesn’t mean a person must click Approve on every tiny action.

It means people design the boundaries, remain accountable, and can intervene when needed.

So yes, I love AI agents.

I just prefer them with a job description, limited permissions, and someone responsible for the outcome.

Pretty much how I’d manage a new employee.

Except this one can process a lot more requests before lunch. 😅