# How Do You Train Staff to Catch AI Mistakes at Work?

Canonical URL: https://business.theaieducator.io/posts/train-staff-to-check-ai-output
Publication: Dan Fitzpatrick Insights
Author: Dan Fitzpatrick
Topic: AI Training and Skills
Published: 2026-10-30T16:59:00.000Z
Modified: 2026-10-30T16:59:19.489Z

Telling staff to be careful with AI does not work. A twenty-minute Seeded Error Drill, with three planted mistakes and a sheet you can copy, trains the checking skill most AI training skips.

## In brief

Train staff to catch AI mistakes by practicing on them: give the team an AI-drafted piece of their own work with three planted errors (a wrong fact, invented support, a wrong fit), have them find the errors, and discuss why each looked right. Repeat monthly.

## Key takeaways

- Telling people to be careful with AI does not train them, because a warning does not show what a mistake looks like; practice does.
- A Microsoft Research survey of 319 knowledge workers found higher confidence in AI went with less critical thinking, and higher self-confidence went with more.
- Workday's November 2025 survey of 3,200 AI users at large organizations found nearly 40% of AI time savings are lost to rework, and only 37% of high-rework employees get skills training.
- The Three Slips are a wrong fact, invented support and a wrong fit; the first two appear in independent testing of AI assistants, and the third depends on your business.
- Match the check to the stakes: a quick read inside the team, the Three Slips for anything customer-facing, and a second person who passes the Loop Test for money, legal, safety or people decisions.
- The Seeded Error Drill takes twenty minutes: ten alone, five in pairs, five as a group, with the answer key revealed only at the end.
- Track the share of planted errors found and the quality of the 'why it looked right' answers, never a score for individuals.

You train staff to catch AI mistakes by letting them practice on mistakes, not by telling them to be careful. Give the team an AI-drafted piece of their own kind of work with three planted errors: a wrong fact, invented support and a wrong fit. Have them find the errors, then discuss how each one looked right. Twenty minutes, repeated monthly, builds a habit that warnings never do.

That is the short version. The rest of this article explains why it works, how much checking each kind of work deserves, and gives you the sheet to run the drill on Monday. It is written for owners of small and medium-sized businesses (SMBs, or SMEs in the UK and Europe) with roughly 5 to 250 people, where nobody has time for a course on critical thinking.

## Why do people stop checking AI output?

People stop checking because the more they trust the tool, the less they test it, and a polished answer gives no signal that anything is wrong. Researchers at Microsoft surveyed 319 knowledge workers who shared 936 first-hand examples of using generative AI at work. They reported in [The Impact of Generative AI on Critical Thinking](https://www.microsoft.com/en-us/research/publication/the-impact-of-generative-ai-on-critical-thinking-self-reported-reductions-in-cognitive-effort-and-confidence-effects-from-a-survey-of-knowledge-workers/) (CHI 2025, April 2025) that "higher confidence in GenAI is associated with less critical thinking, while higher self-confidence is associated with more critical thinking."

That is self-reported data from one survey, so treat it as a pattern, not a law. My reading is practical. The protective factor is not suspicion of the tool. It is a person's confidence in their own judgment of the work. A warning to "always check AI" does nothing for that, because a warning does not show anyone what a mistake looks like. Practice does.

## What does unskilled checking cost?

It costs time, and it can cost most of the time AI was supposed to save. [Workday's January 2026 research](https://newsroom.workday.com/2026-01-14-New-Workday-Research-Companies-Are-Leaving-AI-Gains-on-the-Table), fielded by Hanover Research in November 2025, surveyed 3,200 full-time employees who actively use AI, across North America, Asia-Pacific and EMEA, all at organizations with $100 million or more in annual revenue. It found that nearly 40% of AI time savings are lost to rework: correcting errors, rewriting content and verifying outputs. It also found that 77% of frequent users review AI-generated work as carefully as human work, if not more, and that only 37% of the employees with the highest rework say they get access to skills training.

Read that together. Staff in these companies are checking. They are checking without a method, and most are doing it without training. The survey covers large organizations, so it does not prove the same split in a 15-person firm. I would expect the same shape there, and nothing in it argues against that, but it is an inference and I mark it as one.

## What kinds of mistakes should people look for?

People should look for three, which I call the Three Slips: a wrong fact, invented support and a wrong fit. Each one can sit inside writing that reads perfectly well.

The first two show up in the independent evidence. In October 2025 the European Broadcasting Union and the BBC published [the largest study of its kind on AI assistants and news](https://www.ebu.ch/news/2025/10/ai-s-systemic-distortion-of-news-is-consistent-across-languages-and-territories-international-study-by-public-service-broadcaste). Professional journalists from 22 public service media organizations in 18 countries assessed more than 3,000 answers from ChatGPT, Copilot, Gemini and Perplexity, in 14 languages. They found at least one significant issue in 45% of answers, serious sourcing problems in 31% and major accuracy problems, including hallucinated details and outdated information, in 20%. The questions were about news, not business, so the percentages are not a business error rate. What carries over is the kind of error: facts that are wrong and support that is missing or false.

The third slip is the one a general study cannot measure, because it depends on your business. An answer can be right in general and wrong for you.

| Slip | What it looks like | Hypothetical example (a 14-person bookkeeping practice) |
|---|---|---|
| Wrong fact | A number, date, name or rule that is incorrect or out of date | A client email states a filing deadline a month too late |
| Invented support | A source, case, quote or "guidance" that does not exist or does not say that | The email cites a named guidance page the model made up |
| Wrong fit | Correct in general, wrong for this customer, policy or promise | The email offers a same-day callback the practice does not provide |

## How much checking does each job deserve?

Each job deserves checking in proportion to what happens if the output is wrong, not a flat rule for everything. A flat rule produces either a bottleneck or a signature on a form nobody read.

| Stakes | Examples | The check |
|---|---|---|
| Low: stays inside the team | Meeting notes, a first draft for yourself, brainstorming | A quick read. Anyone can do it. |
| Medium: leaves the building | Customer emails, proposals, social posts, quotes | The Three Slips, every time, by the person who sent the request |
| High: money, legal, safety or people decisions | Payment changes, contract terms, hiring and firing, regulated advice | The Three Slips plus a second person who passes the Loop Test below |

For the top row, I use the Loop Test. The Loop Test is three questions I ask before accepting "a human checks it" as an answer: Would that person notice a wrong result without being told to look for it? Would they notice in time to stop it mattering? And if they missed it, could the work be put back? Oversight that survives all three is a safeguard. Oversight that fails any of them is a signature. You can see it applied to a customer-facing business in [AI for retail and hospitality](https://business.theaieducator.io/posts/ai-for-retail-and-hospitality-small-business), and the payment version is in [how to stop AI voice and payment scams](https://business.theaieducator.io/posts/ai-scams-small-business).

## What is the Seeded Error Drill?

The Seeded Error Drill is a twenty-minute team exercise in which you give people an AI-drafted piece of their own kind of work with three planted mistakes, one of each slip, and ask them to find the mistakes before you tell them how many there are. It trains the skill most AI training skips: noticing what is wrong when nothing looks wrong.

Here is how to run it.

1. **Pick a real task.** Choose one the team does every week, such as a customer reply, a quote summary or a job description.
2. **Draft it with AI, using invented details.** Never paste real customer or employee data into the exercise. Change names and numbers.
3. **Plant three errors.** Edit the draft so it holds one wrong fact, one piece of invented support and one wrong fit. Keep a private answer key. If you already caught a real AI mistake last month, use that as one of the three.
4. **Run it in three parts.** Ten minutes alone, marking anything doubtful. Five minutes comparing in pairs. Five minutes as a group, with you revealing the key. Do not say how many errors there are until the end.
5. **Debrief on one question.** For each error found or missed: what made it look right? That answer is the lesson, and it is usually "it had a source," "it sounded like us," or "I was in a hurry."
6. **Write it down and repeat.** Use the sheet below, and run a fresh drill next month with a different task and subtler errors.

The sheet to copy:

| Line | What to write |
|---|---|
| Task and date | What the draft was for, and when you ran the drill |
| Run by | The person who planted the errors |
| Planted | 1 wrong fact: ____ / 2 invented support: ____ / 3 wrong fit: ____ |
| Found by the group | How many of the three, and which |
| Missed, and why it looked right | The honest reason, in the team's own words |
| The check we add | One concrete habit, such as "open the source before we send" |
| Next drill | Date, task and who plants the errors |

### What does a worked example look like?

Take a hypothetical 14-person bookkeeping practice. The partner asks AI to draft a reply to a client about a late-filing penalty, then plants three errors. The deadline is stated a month late. The reply cites a named guidance page that does not exist. It promises a same-day callback the practice does not offer. In round one (every outcome in this example is illustrative, not data), most people catch the date, because dates feel checkable. Hardly anyone catches the guidance page, because it has a source and sources feel safe. In the debrief the team writes: "we only check things that look like numbers." Their added check is to open every source before a client email goes out. By round three, people are hunting for the third slip as well, because they have learned that a fit error is the one that costs a client.

## What I See When I Train Business Teams

Across the business teams I train, nearly everyone can tell me that AI makes mistakes. Far fewer can tell me what the last mistake they caught looked like. Knowing that errors exist is not the same skill as recognizing one, and the gap between them is where the damage happens. My governing principle for AI at work is to outsource the doing, not the thinking. The drill is the cheapest way I know to keep the thinking in the building: it gives people a trained eye for their own work, not a rule to obey.

## How do you know the drill is working?

You know it is working when the share of planted errors found goes up and the "why it looked right" answers get more specific. Track three things on the sheet: how many of the three the group found, how long it took, and how many real catches people report between drills. Do not turn this into a score for individuals. The point is the group's eye, and anyone who fears being caught out will stop admitting what they missed. If you already measure how AI is performing, add the catch count to what you track, as described in [how to measure whether AI is working](https://business.theaieducator.io/posts/measure-whether-ai-is-working).

## What should you do this week?

1. Pick one customer-facing task your team already does with AI.
2. Draft it with invented details and plant the Three Slips, one of each.
3. Book twenty minutes in a team meeting you already hold, and run the drill as above.
4. Fill in the sheet, and agree one check that goes into the routine.
5. Put the next drill in the calendar for four weeks' time.

If you do the Friday Four conversation already, the drill slots in as a monthly replacement for one session; the Friday Four is covered in [how to manage a team that uses AI](https://business.theaieducator.io/posts/how-to-manage-a-team-that-uses-ai). For the wider question of what a team needs before any exercise, start with [what AI training your team needs](https://business.theaieducator.io/posts/what-ai-training-does-your-team-need), and for turning practice into a plan, [AI training for small business teams](https://business.theaieducator.io/posts/ai-training-for-small-business-teams). If you want to see whether people are really using AI on their own work first, see [is your team really using AI](https://business.theaieducator.io/posts/is-your-team-really-using-ai). Prompting well is half the skill and checking is the other half, which is where [prompting as a job skill](https://business.theaieducator.io/posts/prompting-skills-at-work) fits.

## Where to go from here

If you want your team to practice checking on their own work, with examples from your own business, that is a core part of my [AI training for teams](https://theaieducator.io/ai-training-for-business?utm_source=business.theaieducator.io&utm_medium=referral&utm_campaign=train-staff-to-check-ai-output). If you want to know where you stand before spending another month experimenting, take the [free AI readiness check](https://theaieducator.io/business-ai-readiness?utm_source=business.theaieducator.io&utm_medium=referral&utm_campaign=train-staff-to-check-ai-output). And if you are planning an event for a business audience, my [business keynotes](https://theaieducator.io/business-keynotes?utm_source=business.theaieducator.io&utm_medium=referral&utm_campaign=train-staff-to-check-ai-output) cover this ground for leadership teams.

## Sources and further reading

- Lee, H.-P. et al., [The Impact of Generative AI on Critical Thinking: Self-Reported Reductions in Cognitive Effort and Confidence Effects From a Survey of Knowledge Workers](https://www.microsoft.com/en-us/research/publication/the-impact-of-generative-ai-on-critical-thinking-self-reported-reductions-in-cognitive-effort-and-confidence-effects-from-a-survey-of-knowledge-workers/), Microsoft Research, CHI 2025, April 2025 (n = 319 knowledge workers, 936 examples; self-reported).
- Workday, [New Workday Research: Companies Are Leaving AI Gains on the Table](https://newsroom.workday.com/2026-01-14-New-Workday-Research-Companies-Are-Leaving-AI-Gains-on-the-Table), 14 January 2026 (3,200 full-time AI users, North America, APAC and EMEA, organizations with $100M or more in revenue, fielded November 2025 by Hanover Research).
- European Broadcasting Union and BBC, [Largest study of its kind shows AI assistants misrepresent news content 45% of the time](https://www.ebu.ch/news/2025/10/ai-s-systemic-distortion-of-news-is-consistent-across-languages-and-territories-international-study-by-public-service-broadcaste), 22 October 2025 (more than 3,000 answers, 22 public service media organizations, 18 countries, 14 languages; news questions).

*Dan Fitzpatrick, The AI Educator, is a Forbes contributor, international keynote speaker and bestselling author who has trained more than 150,000 people across 30+ countries. [About Dan](https://theaieducator.io/about).*

## Frequently asked questions

### How do you train employees to check AI output?

Train them by practice, not warnings. Give the team an AI-drafted piece of their own kind of work with three planted errors, a wrong fact, invented support and a wrong fit. Ten minutes alone, five in pairs, five as a group. Then agree one new checking habit and repeat monthly with a fresh task.

### How often should staff check what AI produces?

It depends on the stakes. Anything that stays inside the team needs a quick read. Anything customer-facing deserves a check for the Three Slips every time. Money, legal, safety and people decisions also need a second person whose check would catch an error in time and allow the work to be put back.

### What are the most common AI mistakes at work?

The three to train for are a wrong fact, invented support and a wrong fit. Independent testing of AI assistants on news questions found sourcing and accuracy problems in large shares of answers. A wrong fit, correct in general but wrong for your policy or promise, depends on your business.

### Why do people stop checking AI answers?

A polished answer gives no signal that anything is wrong, and trust in the tool lowers effort. A Microsoft Research survey of 319 knowledge workers reported that higher confidence in generative AI was associated with less critical thinking, while higher self-confidence was associated with more. The finding is self-reported.

### What is the Seeded Error Drill?

It is a twenty-minute team exercise. You give people an AI-drafted piece of their own kind of work with three planted mistakes, one of each slip, and ask them to find them before you reveal how many there are. Then debrief on why each error looked right and add one habit.

### Should I tell staff how many errors are planted?

No, not until the end. Real AI mistakes arrive without a count, so knowing the number turns the exercise into a puzzle and hides whether people would notice anything unprompted. Reveal the answer key after the pairs discussion, then ask what made each error look right.

---
Source: [How Do You Train Staff to Catch AI Mistakes at Work?](https://business.theaieducator.io/posts/train-staff-to-check-ai-output)
Publisher: [The AI Educator](https://theaieducator.io)