Working better with AI

Not: "Write an email about the outage." But: "You are a support employee at an IT company. Write a short, reassuring email to our customers (SME, non-technical) about this morning's email outage, which has now been resolved." The more relevant context, the better the result — feel free to add an example of the style you're after.

End complex requests with: "First ask me clarifying questions if anything is unclear." This stops the AI from making assumptions that leave you correcting it three times. Works surprisingly well for quotes, plans and analyses.

Don't ask for the end result in one go. Have it draft an outline or structure first, give feedback, then let it work it out further. Every intermediate step is a chance to steer — just like with a human colleague.

Do you have to explain what your company does and which tone you want every single time? Use fixed workspaces instead: Projects in ChatGPT and Claude, or Gems in Gemini. There you set context, house style and instructions once — every conversation then starts with that knowledge. Also build a small library of prompts that work well.

Ask by default: "Support this and name your sources." And crucially: verify important facts, figures and quotes in the source itself, not with the AI. A tool checking its own answer is marking its own homework.

Probe further: "What argues against this?", "What assumptions are you making?", "What would an expert object to here?" Or put the same question to a second model or a new conversation as a second opinion. Note: AI tends to agree with you — explicitly ask it to push back.

Use AI as a starting point, not an endpoint. Anything that carries your name — an email, a quote, advice — you must have read, understood and checked yourself. "The AI said so" is no excuse, not even legally.

No personal data of customers or colleagues, contracts, financial figures or passwords in public AI tools. Rule of thumb: don't put anything into an AI tool that you wouldn't put on the internet — unless your organisation uses an approved business variant with clear agreements. Anonymising is fine: discussing a customer case without names, amounts and traceable details in an approved tool is exactly the right way to work.

Which tools are approved, and for what? No policy yet? Raise it — or have Tip ICT draw one up. The European AI Act and GDPR also set legal requirements here. Watch out for copyright too: AI output is building material you edit and check, not a ready-made end product.

The same technology makes phishing emails flawless and voices cloneable. An unusual (payment) request, even if it sounds exactly like your colleague? Verify through another channel. See the recognising phishing training.