Common Mistakes

The mistakes people make most when working with AI every day, the patterns underneath them, and the small habits that prevent each one.

14 minLesson 7 of 7
On this page
  1. One afternoon, six mistakes
  2. The patterns underneath
  3. Four mistakes worth a closer look
  4. Try it yourself
  5. Check yourself
  6. What to practise next

Sam has an afternoon to write the quarterly report for their manager, and an AI assistant to help. By 3:30 the report is sent. By 4:00 the manager has replied: the headline growth figure is wrong, and one of the quoted sources does not say what the report claims.

No single step was reckless. Each was a small shortcut that felt reasonable at the time. That is what the common mistakes look like in practice: several small errors that stack.

One afternoon, six mistakes

Step through Sam's afternoon. Each step names the shortcut and the habit that would have caught it.

How a report went wrongUse the arrows or ← → keys
  1. afternoon
    01:00
    11:05
    21:30
    32:15
    43:00
    53:30
    mistakes
    1

    11:00. Sam types "Write our Q3 report." The reply is a tidy template with generic headings. No goal, no audience, no data: the model filled every gap with the most typical guess. Habit: write a brief.

  2. afternoon
    01:00
    11:05
    21:30
    32:15
    43:00
    53:30
    mistakes
    2

    21:05. To give it "context", Sam pastes the full customer export: names, emails, every order. The report needs totals by month. Habit: paste the least that does the job, and never other people's personal data without a reason.

  3. afternoon
    01:00
    11:05
    21:30
    32:15
    43:00
    53:30
    mistakes
    3

    31:30. The draft is still bland. Sam presses regenerate nine times, hoping the next one is better. Each version is bland in a different way, because the prompt never changed. Habit: name what is wrong and change one thing.

  4. afternoon
    01:00
    11:05
    21:30
    32:15
    43:00
    53:30
    mistakes
    4

    42:15. Sam asks the model to work out quarter-on-quarter growth from the pasted export. It gives 18%. The spreadsheet Sam already has would have said 14%. Habit: use a calculator or code for arithmetic, and recompute any number that matters.

  5. afternoon
    01:00
    11:05
    21:30
    32:15
    43:00
    53:30
    mistakes
    5

    53:00. Sam asks for an industry statistic "with a source" and gets a confident sentence and a link. The link is never opened. Habit: open every citation and find the exact sentence.

  6. afternoon
    01:00
    11:05
    21:30
    32:15
    43:00
    53:30
    mistakes
    6

    63:30. Sam reads the report once for tone, likes it, and sends it under their own name. Habit: the person who sends it owns it. Check the claims that would hurt most if wrong before anything leaves your hands.

  7. afternoon
    01:00
    11:05
    21:30
    32:15
    43:00
    53:30
    mistakes
    6
    time spent
    2.5 hours

    7Six shortcuts, two and a half hours, and a report that has to be corrected in front of the manager. A five-minute brief at the start and ten minutes of checks at the end would have prevented the costly ones.

1 / 7

The patterns underneath

Most everyday mistakes fall into four groups. Earlier lessons cover each one in depth; here is the short version, so you can recognise them as they happen.

PatternLooks likeThe habit
Under-askingOne-line requests for big jobs; no goal or definition of doneWrite a brief
Over-trustingFluent text accepted as true; numbers and links not checkedVerify claims in proportion to stakes
Over-sharingWhole files, logs or exports pasted "for context"Paste less, redact the rest
Over-delegatingHuge diffs accepted; learning outsourced; judgment handed overSmall steps; you do the thinking

Four mistakes worth a closer look

The regenerate loop

Pressing regenerate feels like progress because something new appears each time. But if the prompt does not change, you are drawing again from the same spread of likely answers (Sampling and Temperature explains why replies vary). After two tries, stop and say what is wrong.

Naming the problem instead of regeneratingIllustrative example. Real replies vary.
You
This draft is too generic. Three changes: 1. Lead with the one number my manager cares about: revenue for the quarter compared with last quarter. I'll give you both figures below. 2. Cut the "market overview" section. My manager knows the market. 3. Each section should end with what we are doing next, in one sentence. Revenue Q2: 412,000. Revenue Q3: 470,000.

Even here, check the result: (470,000 − 412,000) ÷ 412,000 is 14.1%. If the reply says anything else, the reply is wrong.

Using AI where a plain tool is better

Adding up a column, sorting a list, finding every line that matches a pattern, converting a date format: a spreadsheet, a script or a search box does these exactly, every time. A language model does them approximately. Use the assistant to write the formula or the script, then let the tool do the work.

Set and forget

An automation that worked in testing can drift: the inputs change, the provider updates the model, a new kind of request shows up. Keep a small sample of real outputs under review every week, and rerun your test set when anything changes. Failures that nobody looks at are failures that pile up.

All-or-nothing trust

Some people trust everything after a few good answers. Others give up after one bad one. Both skip the useful middle: trust that depends on the task. An assistant can be very reliable at rewording your email and unreliable at quoting case law. Learn where your tools are strong by checking, and adjust how much you check as you learn.

Try it yourself

Try it: Spot the mistakes

Read this description of how someone used an AI assistant at work. List every mistake you can find and the habit that would prevent it.

Show a sample answer
  1. Pasted the .env file. That file usually holds secrets. Paste only the code involved, with secrets replaced. If a real key went in, rotate it.
  2. "Fix the discount bug" with no detail. No description of the bug, the expected behaviour or the files involved. Write a brief: what is wrong, an example input and the expected output, which files to change.
  3. Accepted a 14-file change. Far too much to review. Ask for the smallest fix, in one place, and read the diff.
  4. "Make the tests pass". This invites the model to change the tests. Edits to tests are changes to the contract and need a reason.
  5. Green meant done. The tests had been changed, so passing them proved little. Check what changed in the test files before trusting the run.
  6. Merged with no time to watch it. Not an AI mistake as such, but AI-assisted changes deserve the same care as any other: merge when someone can notice and fix problems.

Check yourself

Quick check

You have regenerated a reply four times and every version is slightly off in the same way. What is the best next step?

You need the average of 400 values in a spreadsheet. What is the best use of an AI assistant?

An AI-drafted summary you sent to a client contains a wrong date. Who is responsible?

What to practise next

Write your own five-line checklist from this module: brief, paste less, small steps, verify, own it. Keep it next to your assistant for a week. Before you let an AI agent act on its own, reread Guardrails: the same habits apply to every step an agent takes.