Monday. A sales manager sends a client a proposal. Inside it is a figure that appears in no internal document. The model “calculated” it; the manager didn’t double-check. The client, luckily, didn’t either.
Wednesday. Marina from HR opens her inbox and finds twelve cover letters. Identical, as if photocopied. Same opening sentence, same intonation — as if twelve people had shared one head that morning.
Friday. An analyst pastes the text of an NDA into a public chatbot. “Translate it, it’s in English.” Along with the NDA, the chat receives names, sums, and the terms of a deal the company hasn’t announced yet.
Three people, three mistakes, one and the same tool. Here’s the funny paradox: in all three cases, AI worked exactly as it was told. It wasn’t the neural nets that failed — it was the habits around them. These are the ten habits, from amusing to genuinely expensive, and how to part with each one. If you recognized your own team — no worry, all ten are curable: I help companies with exactly these problems.
Ten mistakes
1. Asking vaguely. “Write a letter to a client” — and you get a letter that resembles all letters at once: polite, even, empty. Now picture forty of them. A client opens the third in a row, recognizes the opening sentence, and stops reading. He doesn’t get angry, doesn’t complain. He just stops. The same prompt with context — who’s the client, what’s the deal, how we usually write — plus a past letter as a sample returns text that differs from the template exactly as much as your business differs from average. If you want to go deeper, Anthropic’s prompt engineering guide is the place to start.
2. Quitting after the first answer. The first draft is weak, and something inside clicks: “well, that’s that, a toy.” It’s like hiring a designer and firing him after the first sketch — money paid, nothing gained. But the interesting part starts right after. “Make it shorter. Now stricter. Add the client’s objection. Now criticize your own answer.” Five minutes of that conversation — and instead of mush you hold a draft you’re not ashamed to send. Iterations, in detail — same guide.
3. Treating the model as all-knowing. “Check what the client wrote in our CRM.” The model isn’t connected to the CRM and says so honestly. But the employee is already offended — and by lunch the whole department knows “your AI doesn’t work.” The tool’s reputation dies for nothing. All it took was attaching an export of the correspondence: in a minute, the model finds what a human would have spent an hour of careful reading to find. No access? Hand over the data yourself. What an assistant with real access actually looks like, I showed in the piece on AI for a CEO.
4. Taking it on faith. The most expensive habit. Lawyers in the US filed a court brief citing precedents — confident, polished, existing only in the model’s imagination. The judge asked for the texts of those rulings. They couldn’t be found in any database. Mata v. Avianca became a textbook case of automation bias, and the lawyers’ names became a footnote in a court order. The annoying part: checking precedents would have taken one evening. Two minutes of verification against the primary source — and the figure in the presentation exists somewhere besides the neural net’s head. How control turns into ceremony, I broke down in the article on personal AI assistant risks.
5. Sending it as is. An experienced HR manager opens a cover letter and reads the first sentence. “Excited about the opportunity to join your team.” That’s it, you can close it: no human wrote this, and not today. If a text can go out without a single edit, a competitor is sending the same text at this very minute. But a draft into which you’ve added your experience, your figure, your intonation stops being average. It starts getting answers. Where text decides the outcome and where it doesn’t, I wrote in the piece on AI for work, not for demo.
6. Feeding it secrets. Samsung engineers pasted code and meeting protocols into a public chatbot — until the company imposed a hard ban. Here the rules are stricter still: citizens’ personal data must not leak into foreign services, and 152-FZ explains why. The way out is simple and unglamorous: “Client A,” “Company B,” sums in round units — and a calm translation. And the company should long ago have stood up a corporate perimeter, so employees don’t have to choose between convenience and safety. The security side of this is covered in the NCSC and CISA Guidelines for Secure AI System Development.
7. Hiding in personal accounts. “AI is banned here,” the manager says at the standup. At that very moment, forty employees in adjacent browser tabs are rewriting their reports in personal chatbots. That’s nearly every other office, and a ban is powerless: convenience always beats a piece of paper. What works is corporate access. People stop hiding, security sees the logs, the data stays home. BlackFog studies this shadow work closely — it’s bigger than most leaders think.
8. Hiding behind AI. “Where’s the number from?” — “ChatGPT said so.” In a meeting that sounds like a confession. In court it sounds like a verdict on yourself: Air Canada tried to blame its own chatbot, which had invented a ticket refund rule, and had to return the customer’s money. The court was laconic: a company answers for what its bot says. Which means the questions will go to a human — the one who pressed “send.” Check the source beforehand, and at the meeting you answer “according to the quarterly report,” not “I was told so.” By the way, Article 16 of that same 152-FZ speaks plainly about a person’s right to contest a decision made by a machine. More on where responsibility actually lives — in the article on personal AI assistant risks.
9. Handing out keys without a manual. Licenses bought, access granted, everyone waiting for a miracle. A quarter later, silence: people tried, understood nothing, gave up. Money spent, enthusiasts burned out, effect — zero. And yet it was all solvable in a day: an hour of hands-on practice on the department’s real tasks, an internal library of prompts that worked, a person you can walk up to. Teams that were taught win many times over teams that were simply handed a password — and the difference is measured, not promised. See the NBER study “The Cybernetic Teammate”.
10. Banning instead of managing. The outcome of bans is always the same: shadow work, and behind it a leak the company learns about from the news. Managing, though — managing is possible and necessary. Sort scenarios by degree of autonomy, keep logs, check results rather than login counts. Speed doesn’t go anywhere in the process; it simply stops being scary. Four levels of autonomy, with the right control for each, I broke down in the article on personal AI assistant risks.
What to read
- Anthropic’s prompt engineering guide — for mistakes one and two.
- Guidelines for Secure AI System Development — NCSC and CISA recommendations for working with AI safely.
- NBER, “The Cybernetic Teammate” — honest measurements of AI’s effect in real workplaces.
- Microsoft Work Trend Index — how AI is changing work, year over year.
In short
AI forgives a sloppy question but never forgives carelessness with the answer. The tool is rarely the one that’s wrong — the habits around it are.
If you recognized your own company in these stories — it’s fixable. I help teams work with AI without these ten mistakes: policies, training, a corporate perimeter, audits. Get in touch and we’ll look at your case.