Using AI at work: where to start

Company rules, a first low-risk task, data you shouldn't paste, measuring your trials: a six-step plan for getting started without a misstep.

4 min read

Tools hung on a workshop pegboard
Photo: Brian Kay · CC BY 3.0, cropped

Camille is an administrative assistant at a company of fifteen people. In a meeting, her director said: "We should look at what AI can do for us." Nobody specified what, or with which tool, or under what rules. She has two options: wait for someone to explain, or open a chatbot and paste in the first document that comes to hand. The second is more tempting and a good deal riskier.

The path comes down to six steps, and it suits any office job.

1. Find out what's allowed before you open anything

Ask whether there's a rule on using AI tools: a list of approved tools, work accounts provided, a ban on entering certain data. In a small company, the answer may be "nothing in writing," which is not a green light. Ask your manager or whoever handles IT.

The sensitive point is data: customer names, salaries, contracts, information about colleagues. Depending on the tool, the plan, and the settings, what you enter may be stored. Personal data falls under data protection rules, and the French data protection authority (CNIL) publishes guidance. The data you should never paste deserve ten minutes of reading before your first trial. If you're unsure about a particular file, ask before sending it, not after.

2. Pick a low-stakes first task

The best candidates are repetitive, easy to check, and free of sensitive data:

  • rephrasing an email you've already written;
  • turning meeting notes you took into minutes;
  • preparing a meeting agenda from topics you list;
  • finding titles or outlines for a presentation.

Avoid these to begin with: anything involving an official figure, an HR decision, a legal letter, or an unhappy customer. Those are the tasks where a mistake costs you, and where you're still learning the tool.

3. Write the context once

Camille chooses the meeting minutes. Rather than explain who she is every time, she writes a short sheet of a few lines: the company's industry, the usual tone of the minutes, the format she wants. She pastes it at the start of each request. What goes into such a sheet is detailed in the article on context.

Prompt
Here are my notes from the team meeting on [DATE]. Participants are identified by their initials.

[PASTE YOUR NOTES, WITHOUT CONFIDENTIAL INFORMATION]

Write meeting minutes of 15 lines maximum with three sections: decisions made, actions to take (who, what, by when), topics postponed. Do not add any decision that isn't in my notes. If a deadline or an owner is missing, write "to be confirmed."

The words "to be confirmed" are the real safeguard: without them, the AI will happily invent a plausible date.

4. Proofread as if you had signed it

Anything that comes out of the AI and goes out in the company's name commits someone, and that someone is you. For minutes, check the names, dates, amounts, and commitments. For an email, make sure the promise being made is one your company can keep. The method is in the article on checking an answer.

5. Measure with three trials, not an impression

After a few days, the temptation is to say "it works" or "it doesn't work" on the strength of a single result. Instead, run three trials on three different cases and note what you had to fix by hand. If the fixing takes as long as the writing, the task isn't suited to AI, or your prompt needs work. Testing a prompt describes a testing method.

Time yourself once with the AI and once without, proofreading included. Without that number, you can't tell whether the tool is helping you, and you have nothing to show your director.

6. File what works and share it

Once a prompt has proven itself, put it in a shared document with the brackets to fill in. Colleagues who want to get started will find a starting point instead of a blank page, and you'll keep everyone from making the same mistakes on their own. The article on reusable templates suggests a simple way to organize it. The professional email prompts make a good first chapter for this library.

What stays your responsibility

The AI prepares, suggests, and rephrases, and the decision stays with you. You choose, you proofread, you sign, and you answer for the result to your management and to the people you write to. If you can't say why the final text is correct, you've published it too soon.

This week

List three repetitive tasks you did this week. Cross out the ones that involve sensitive data or an important decision. For the one that's left, write the context sheet, run your three trials, and note the time. By Friday, you'll have a well-founded opinion to give your director.

Getting started

What Context to Give AI to Get Good Answers

Too little context and the AI makes things up, too much and it loses the thread. The test to apply to each piece of information, with examples and a reusable template.

5 min read

Getting started

Giving the AI a Role: When It Helps, When It's Pointless

"You are an expert with 20 years of experience" doesn't add any skill to the AI. What a role really changes, and how to write one that's worth having.

5 min read

Getting started

Writing a Clear Request in a Single Sentence

One deliverable, one reader, one goal, the facts, a form: the single sentence that's enough for most requests to AI, with an example from start to finish.

3 min read