What happens when AI makes a mistake?

AI will make mistakes. Here's what they look like.

I've built with AI every day since ChatGPT launched in November 2022. It has made mistakes every week of that time. Not because the tools are bad, but because of how they work: a language model predicts what text should come next. Most of the time that prediction is right. Sometimes it is fluent, confident and wrong.

In a small business, the mistakes fall into a few predictable buckets:

• Invented facts (hallucinations). Ask for a summary of your return policy and it may fill a gap with a policy you don't have. Ask for a source and it may cite one that doesn't exist.
• Stale information. The agent reads an old price sheet, last year's holiday hours, or an email thread from before the job was rescheduled. The info was true once. It isn't now.
• Misread context. It mixes up two customers with the same last name, or treats "maybe Thursday?" in an email as a confirmed booking.
• Confident wrong dates and numbers. Day-of-week math, totals across a spreadsheet, a tax rate. These hurt most, because they look exactly as believable as the right answer.
• Following instructions too literally. You said "reply to every review," so it politely replied to the one that was obviously spam.

None of this is rare or exotic. If you plan for AI to be right every time, you've planned for something that won't happen. If you plan for it to be wrong sometimes, you can build around it.

Why AI mistakes are harder to spot than human ones

When a new employee is unsure, you can usually tell. They hedge, they ask, the email reads a little off. AI doesn't do that by default. A wrong date and a right date come out in the same calm, polished sentence.

That's the real risk. Not that AI is wrong more often than people (on routine tasks it often isn't), but that its errors don't look like errors. A skimmed draft passes. A spreadsheet total that's off by one row looks fine at a glance.

So "looks good to me" is not a check. A real check compares the output against something: the calendar, the invoice, the CRM record, the original request. If nobody compares, nobody's checking. If you want plain-English definitions of terms like hallucination, the AI glossary has them.

Watch an agent make a mistake, and watch it get caughtA 30-second demo of the human check that keeps errors away from customers.
Watch the demo

How to design for mistakes

You don't fix AI mistakes by finding a tool that never makes them. You fix them the way you'd run any operation where errors are costly: with process. Here's what that looks like in practice.

1. Keep a human in the loop for anything customer-facing. Emails to clients, quotes, invoices, social posts, review replies: a person reads it before it goes out. Internal drafts can move faster. Anything with your name on it gets eyes.

2. Write source-of-truth rules. Every important fact needs one home, and the agent should be told where it is. "Appointment dates come from the job calendar, never from the email thread." "Prices come from the current price sheet in the shared drive." "Names and addresses come from the CRM." This one habit prevents more errors than anything else on this list.

3. Check against the source, then spot-check the rest. Things that matter (dates, dollar amounts, names, anything legal) get checked every time. Lower-stakes work, like tagging leads or renaming files, gets spot-checked: pull five at random each week and look closely.

4. Start with internal work. Your first automations should be things only your team sees: meeting notes, CRM cleanup, a weekly report draft, SOP drafts. You learn how the agent fails while the stakes are low. Customer-facing work comes after you trust the process.

5. Fix the instruction, not just the output. When a mistake gets caught, correcting that one email is half the job. The other half is changing the agent's instructions so it can't make that mistake again. Over a few months, this is what makes an agent reliable: a growing list of rules written from real errors.

6. Keep logs. You should be able to answer "what did the agent do, and when?" Most tools already keep history (sent mail, CRM activity, file versions), and workflow tools like n8n log every run. When something goes wrong, the log tells you whether it was one bad output or a pattern.

Who is responsible when AI gets it wrong

You are. Your business is. If an agent sends a customer the wrong price, the customer doesn't care that software wrote it. They care that your company said it.

That isn't a reason to avoid AI. It's a reason someone has to own it. Every agent needs a person whose job includes reviewing its work, maintaining its instructions, and noticing when something drifts. In a big company that's a team. In a small business it's usually the owner, an office manager with some spare capacity, or an outside fractional AI manager.

This is the part most people skip. They set up a tool, it works great for two weeks, and then nobody's watching when the price sheet changes. It's also why my own service is built around checking: my agents do the work, and I review every piece before a client sees it. The details are on how I work.

What this looks like in a real small business

Example: picture a six-person home services company. An agent drafts install confirmation emails from each customer's email thread, and the office manager reviews them before they go out.

One morning the agent drafts: "Hi Karen, you're all set! Your install is Tuesday, November 3 at 8 a.m." It reads perfectly. It's also wrong. Karen's install was rescheduled to Thursday, November 5, and the agent pulled the old date from earlier in the thread.

The office manager doesn't approve on vibes. She checks the date against the job calendar, sees the mismatch, and holds the draft. She fixes the date and sends it. Karen never sees the error.

Then she does the step that matters most: she adds a line to the agent's instructions. "Always take appointment dates from the job calendar, never from the email thread." The next hundred confirmations pull from the calendar, and the same mistake doesn't come back.

Total cost of the error: about two minutes. Without the check, it's a technician at an empty house on Tuesday and an annoyed customer. You can watch this exact scenario play out as "The catch" on my demo page.

The honest part

Review doesn't make AI perfect. People skim, especially on the fortieth email of the day. Checks catch most mistakes, not all of them, which is why source-of-truth rules matter so much: they stop errors before anyone has to catch them.

Some work shouldn't be handed to AI at all, or only with a licensed professional signing off: legal advice, medical guidance, tax filings, anything where a confident wrong answer causes real harm. AI can draft. A qualified person decides.

And review takes time. If a task takes five minutes to do and four minutes to check, automating it barely helps. The best candidates take a while to do but are quick to verify against a source. If you can't describe how you'd check the output, the task isn't ready for an agent yet.