Digital Employees

What happens when an AI employee gets something wrong?

By Carol Roderick, PhD  ·  November 18, 2026 Digital Employees

An AI employee writing a law practice's posts will publish a date that's wrong. What matters is who catches it, and how fast. The ones I build stop and tell the person they report to when they aren't sure. A person in the practice owns what goes out.

I read every flag my own AI employees send me

When AI gets it wrong in my own business, I'm usually the first to know. Not always. One of mine made up the news for two days before I noticed, so I won't start by telling you yours never will.

People ask me how often it happens. I don't have a number, and I won't make one up. Ask anyone who hands you a figure where they got it, because there isn't an honest one for my work.

The pages that answer this question are mostly vendor pages. ClawHire's answer says yes, an AI employee can get something wrong, and then lists the safeguards. What none of them shows you is the day it happens in your own office, and that part I can tell you about, because I run one every day.

My rule is simple to say. When one of my AI employees isn't sure about something, it doesn't guess and it doesn't go quiet. It stops, writes to me, and tells me what it's stuck on. I'm the one who reads that message. And when one of them gets something wrong anyway, I own it. The AI doesn't.

Susie is the AI employee writing the practice's posts, and she reports to you

Picture one AI employee in your office. Call her Susie. She has one job: the practice's social posts. The information night you run twice a year. The question clients ask every January about their will. The new lawyer's introduction. The post about what to bring to a first meeting, because you've answered that one the same way for years.

Susie writes those in your practice's voice, from what you hand her, and she reports to you. She works in writing and she doesn't make or take calls.

You already have the number she works on. It's how many first consultations come off the back of those posts: the ones that arrive saying they saw it on your page. When Susie is doing her job, that count goes up, and you can see it without asking anyone.

Susie isn't taking work off anyone's desk. Right now the posts get written when somebody has a quiet afternoon, which is to say rarely, and the last one is from March.

An AI employee stops and asks you the moment it can't tell what's right

If Susie isn't sure, she stops and tells you right away, and she asks rather than guessing at an answer.

Flags come as they happen. Say the fee in her draft and the fee on your own fee sheet don't agree. You hear about that the same morning, not on Friday. Her weekly report sums up what went out, what came back and what she flagged, but it's never the first place you learn about a problem.

I can't promise a client will never see a mistake. I can promise that when Susie doesn't know something, you'll know too.

When a flag lands, you do one of three things:

Answer it. The post says your information night starts at 6:30, and the calendar says 7. You write back, and Susie carries on.

Fix the draft and send it. Susie has a post ready about the new parking rules, but the landlord changed them again this morning. You change one sentence and let it go.

Take it over. A client has written in about their own matter, and that isn't a post problem. You answer it yourself.

The AI employee publishes the wrong date, and you catch it the next morning

On Monday, one of the lawyers moves the practice's information night from the 21st to the 14th. She tells a client in a meeting and adds a note to the file. The date in the events calendar still says the 21st.

On Tuesday, Susie writes the post for the night, and the date in it is the 21st. She had no reason to flag it. The calendar said the 21st and nothing looked off to her. A flag won't catch this kind of mistake. You read the post on your way out the door and let it go, the way you'd skim a new hire's email in their first month. It's live by the evening.

On Wednesday morning you're reading what went out the day before. You see the 21st and you remember the meeting. By noon the post is down, the right date is in it, and the lawyer has read your note before it goes back up. The wrong date was public for about a day.

On Thursday the post is back up with the 14th on it, and two people have signed up for the night.

You correct an AI employee once, and the correction sticks

On Wednesday, once the post is fixed, you tell Susie what went wrong. The calendar entry and the note in the file didnt match. From now on, when those two don't agree, she asks you which one is right before she writes anything.

That's how you stop the same mistake coming back. You write the correction down, it becomes part of how she works, and you watch her next few posts to make sure it holds.

How much Susie does on her own is your call, and it changes as she earns it. At the start, everything she writes comes to you for a yes before it goes out.

Once you stop finding things to fix in the plain ones, you let her post those herself. A reminder that the office is closed for the long weekend. A post about the fundraiser the practice backs. Anything with a date or a fee in it stays with you longer.

Don't pull her back after the first mistake. That week is when she learned the most.

The business owner is accountable for what an AI employee sends in the practice's name

Susie reports to you, and you read her work. But anything that reaches a client carries the practice's name, and the business owner is accountable for that. Not Susie, and not me.

This is how I run it, and it isn't a statement of what the law says about any of this. One person reads the work and one person owns the outcome, and in a small office that might be the same person.

You don't have to wait for the report to see Susie's work. Her posts, the replies they get and her flags are there to read any day you want.

Book 30 minutes about the AI employee, and who it would report to in your office

To start with one AI employee writing the practice's posts, book Let's Chat AI. It's a 30-minute call. Bring the posts you wish you'd written and the questions clients ask you most, and we'll work out who Susie would report to in your office, and which posts she'd write first.

AI employee questions, answered

How do I find out when an AI employee gets something wrong?

There are two ways. When an AI employee doesn't know what to do, it flags the person it reports to right away, so you hear about it as it happens. For the mistakes it can't see, you read its work. Its posts and the replies they get are there any day, not only on report day. Now and then a client will spot a mistake first, and you'll see that in their reply.

Who is responsible when an AI employee makes a mistake?

A person in the practice, never the AI. The AI employee reports to one person on staff who reads its work, and the business owner is accountable for anything sent in the practice's name. That's how I run my own. I'm not stating what the law says about it.

What does an AI employee do when it is not sure about something?

It stops and asks. It doesn't guess, and it doesn't leave the work sitting. Say a post has a date or a fee in it and she can't tell which one is right. Susie, the AI employee writing the practice's posts, writes to you and asks before anything goes out.

How do I stop an AI employee making the same mistake twice?

Correct it in writing and make the correction part of how it works. Say Susie publishes a post with an old date because the events calendar and the note in the file don't agree. The fix is a new rule: when those two don't match, she asks you first. Then watch her next few posts to make sure the rule holds.

Can an AI employee be trusted with work a client sees?

Yes, once it has earned it, and you decide how much. Start with drafts you approve before they go out. As your corrections drop off, let it schedule the plain ones on its own and keep the ones with a date or a fee with you. Keep reading its posts even after you've stopped approving each one.

Carol Roderick, PhD, is an AI adoption consultant who builds digital employees for expertise-led businesses. Each AI employee holds one job and reports to someone on staff. More at https://carolroderick.ca.