15 AI mistakes professionals make

All 15, in one place

Every weekday issue ends with one mistake to stop making. This is the whole list, so you can read them together instead of one at a time.

• Don't ask the model "is this correct?"
• Don't paste a citation you haven't opened
• Don't ask about a document you haven't given it
• Don't cram five jobs into one prompt
• Don't paste anything you would not email to a stranger
• Don't ship the first draft
• Don't skip who it is and who it is for
• Don't use it as a search engine for current facts
• Don't say "make it shorter"
• Don't conclude the tool is useless after one bad answer
• Don't outsource the judgement, only the typing
• Don't describe the format when you could show it
• Don't retype your context every morning
• Don't fight a refusal — reframe the ask
• Don't let it do arithmetic unsupervised

• 01 Don't ask the model "is this correct?"

  Models are agreeable by default. Ask "is this right?" and you will almost always be told yes — including about the parts that are wrong.

  Do this instead: Ask it to attack the work instead: "What are the two weakest assumptions here, and what evidence would disprove them?" You get a critic rather than a cheerleader.

• 02 Don't paste a citation you haven't opened

  A model will produce a case name, a study, or a page number in exactly the right shape and with total confidence, and the source may not exist. The formatting is not evidence.

  Do this instead: Open every link and search for the exact quoted sentence before it leaves your desk. If you cannot find the source in thirty seconds, cut the claim.

• 03 Don't ask about a document you haven't given it

  Asking "what does our returns policy say?" without pasting the policy invites the model to invent a plausible one, because answering is what it does.

  Do this instead: Paste the source first and add: "Answer only from the text above. If the answer is not there, say 'not in the source'." You turn a guesser into a reader.

• 04 Don't cram five jobs into one prompt

  "Summarise this, draft a reply, build a plan, and make a slide outline" reliably produces four mediocre things, because quality drops as the instruction count climbs.

  Do this instead: One task per message. Do the summary, check it, then ask for the reply built on the summary you approved.

• 05 Don't paste anything you would not email to a stranger

  Customer lists, unreleased numbers, salary data and personal details routinely get pasted into consumer chat tools whose retention settings nobody in the room has read.

  Do this instead: Redact before you paste — swap real names for Client A, real figures for round placeholders. Ask your IT team which tool is approved before you use it on anything real.

• 06 Don't ship the first draft

  The first output is the model's average answer, not its best one. Shipping it is why AI writing is recognisable — it reads like the midpoint of everything ever written on the topic.

  Do this instead: Say: "Rate this draft out of 10 against the brief, name the three weakest lines, then rewrite it." The second pass is usually the one worth sending.

• 07 Don't skip who it is and who it is for

  "Write a project update" gives you generic corporate mush, because you have not said whose voice it is in or who reads it.

  Do this instead: Give it three things every time: a role, an audience, and a format. "You are a delivery lead. Write for a sceptical finance director. Five bullets, no adjectives."

• 08 Don't use it as a search engine for current facts

  Unless a tool is explicitly browsing the live web, its knowledge has a cutoff date and it will still answer questions about last week with total assurance.

  Do this instead: Use it for reasoning, structure and drafting; use a real search for anything time-sensitive. If you need both, paste the search result in and ask it to reason over that.

• 09 Don't say "make it shorter"

  "Shorter" is not a measurement, so you get a version that is slightly shorter and just as flabby, and then you ask again.

  Do this instead: Give it a number and a constraint: "120 words. Lead with the decision. Keep exactly one specific figure." Countable instructions get followed.

• 10 Don't conclude the tool is useless after one bad answer

  Most "AI is overhyped" verdicts come from a single thin prompt on a task the person never specified properly. The tool was never told what good looked like.

  Do this instead: Before you write it off, run the same task once more with a role, the real source material, and an example of the output you want. Judge it on the second attempt.

• 11 Don't outsource the judgement, only the typing

  "Which vendor should we choose?" hands a decision you are accountable for to a system with no stake in it and no knowledge of your politics, budget, or history.

  Do this instead: Ask it to build the comparison — criteria, trade-offs, what would change your mind. You still make the call, but on a better-laid-out table.

• 12 Don't describe the format when you could show it

  Explaining the structure you want in prose takes longer than showing it and works worse, because your description is ambiguous and an example is not.

  Do this instead: Paste one good example from your own files and say "match this structure and tone." One sample beats a paragraph of instructions.

• 13 Don't retype your context every morning

  Pasting the same background about your team, product and constraints into every new chat is the single biggest hidden time cost of using AI at work.

  Do this instead: Write the context block once, save it somewhere you can reach in two keystrokes, and paste it at the top. Most tools now have a memory or custom-instructions setting that does it for you.

• 14 Don't fight a refusal — reframe the ask

  Arguing with a model that declined a request burns time and usually fails, because the refusal is about how the request was phrased, not about you.

  Do this instead: State the legitimate purpose and the audience plainly: "I'm the account owner writing a policy summary for our internal team." Genuine context usually resolves it in one turn.

• 15 Don't let it do arithmetic unsupervised

  Language models predict text, so a total can look completely reasonable and still be wrong — and a wrong number in a confident table is harder to spot than a wrong sentence.

  Do this instead: Have it show the working, or better, have it write the formula and run the sum in your spreadsheet. Never paste an AI-generated figure into something financial without recomputing it.

If a word here was new

The terms these mistakes lean on — context window, hallucination, system prompt — each have a plain-English page in the glossary. You can also read the issues themselves, where one of these runs at the bottom of every weekday email alongside a prompt, a word and a skill.

One of these every weekday

Three minutes. One prompt, one word, one skill. You're now better at AI than you were yesterday.

Free. We send a confirmation link first, so nobody can sign you up without their say-so.