The CEO's Guide to Getting Started with AI

I was speaking to a room full of CEOs, and at the end one of them asked: "This all sounds great. But I don't know where to start."

Start with your own work. Pick an approved tool, bring it a real problem this week and use it every day for the next two weeks. Check what it produces and keep improving how you use it.

You need that experience to lead this well. You can ask someone to handle setup, training and support. You still need to sit down with the tool yourself and find out what it can do for the work you understand.

Lead by going first

If you want your team to take AI seriously, give them something useful to follow. Show them a piece of your own work, explain where AI helped and point out what you had to correct. Be willing to learn in front of them.

That's what I mean by getting your reps in. Work on a real agenda, prepare for a meeting or improve a draft you need to send. Come back to the same kind of task and get better at it. Your questions about training, quality and cost will have something concrete behind them.

Book the practice time now. At the end of the fortnight, bring your team a few useful examples and a recommendation about what to try together. Here is how I'd spend the first week.

Choose the account before you choose the task

Ask whoever is responsible for your technology which AI product, account and uses the organization has approved. Start there if it supports the work you want to do. Check which information you can use with it before you upload a document.

If you need an account, ask your technology owner to approve a product and set it up for work. Agree on which information you can use and whom to ask when you're unsure. Practice on public information while they finish the setup.

The familiar options include ChatGPT, Claude, Gemini and Microsoft Copilot. Start with the one your company already approves if it does the work you need. Try it on a few typical tasks before shopping for another tool.

There's a distinction worth understanding before you pay: a paid personal subscription and an organization's business service can have different data terms. For example, Anthropic's commercial privacy guidance explicitly separates Claude for Work from its consumer Free, Pro and Max plans. Check the plan and account you're actually using. Read Anthropic's guidance.

The providers' current documentation gives these specific commitments:

• OpenAI says ChatGPT Business and Enterprise data is not used for model training by default, with an exception when customers explicitly opt to share data. OpenAI's enterprise privacy page.
• Anthropic says commercial inputs and outputs are not used for training by default, but submitted feedback or an explicit choice to share data can permit training use. Anthropic's commercial training policy.
• Google says content covered by its qualifying Workspace services is not used for generative AI model training outside the customer's domain without permission; its Privacy Hub specifies which editions and services it covers. Google's Workspace Privacy Hub.
• Microsoft says the prompts, responses and Microsoft Graph data covered by its enterprise Copilot offering are not used to train foundation models, while noting that web queries have separate data-handling practices and agents require their own terms checks. Microsoft's enterprise data protection guidance.

Have your technology or privacy owner confirm the exact product, settings and agreement. Ask about retention, administrator access, connected applications and feedback sharing as well as training. OpenAI's own documentation, for example, addresses retention and access separately from its training policy.

Then give people a usable answer about what they may do. A lengthy list of unresolved concerns is no help to somebody trying to complete a task on Tuesday.

I've watched two different CEOs pay a monthly fee for a knock-off "GPT" app they found at the top of a search, believing they had bought the real thing. Use your organization's approved access link. If you're arranging a new account, have the responsible person verify the vendor and workspace before you subscribe.

Monday: choose work you can judge

Choose a task you know well enough to judge. I'd start with preparation or drafting that you do regularly, and check the finished work yourself.

Try preparing questions from a public annual report, improving a non-confidential presentation or drafting an internal agenda from approved information.

For your first task, leave out sensitive personnel details, confidential board material, customer records, credentials and anything whose use hasn't been approved. If a task depends on restricted information, select another task while the relevant permission is resolved.

Write down what you'd otherwise do and approximately how long it takes. Keep an example of the normal finished output if you can use it within the approved boundary. Set a modest aim, such as preparing a useful first draft or surfacing questions you had missed.

Avoid choosing a task solely because it looks impressive in a demonstration. You'll learn more from a piece of work that recurs in your week and has a clear standard. The result should help you prepare for an actual decision, conversation or deliverable.

Tuesday: give it a usable brief

Brief the tool with the care you'd give a colleague doing unfamiliar work. Explain the purpose, the reader, the source material and the constraints. Make the expected result explicit.

Here is a sample brief you can adapt:

I'm preparing for a leadership discussion about the attached public annual report. Use only this document. Draft a one-page briefing for a CEO who wants to understand the organization's priorities and the questions that need further investigation. Separate facts in the report from your interpretations, and include page references for the facts. Don't invent explanations for missing information. End with questions we should investigate before drawing conclusions. If you can't read part of the document, tell me where.

Read the output alongside the source. Check whether it followed the brief, and use what you find to improve both the result and your instructions.

If the answer is generic, identify what's missing. It may need a clearer audience, a better example or a narrower task. Ask it to revise the relevant section and explain what it changed. Avoid repeatedly requesting a better answer without saying what better means.

Keep control of the purpose. If the tool proposes a different decision or adds a persuasive explanation, compare that addition with the evidence you supplied. A fluent response can still move the discussion away from the question you actually needed to answer.

Wednesday: practice checking the answer

Checking is part of the work from the beginning. In its 2024 guidance, the US National Institute of Standards and Technology describes the risk of generative AI confidently producing false content and recommends verifying sources and citations during evaluation and monitoring. Read NIST's Generative AI Profile.

Use a straightforward review routine. First check that the output answers the request. Then check material facts against the source, redo calculations using the underlying figures and identify anything that has been assumed. Finally, decide whether the work is suitable for its intended reader.

For the annual-report exercise, open the referenced pages. Confirm that the figures and dates refer to the right period, and that the interpretation is distinguishable from what the report says. If a citation doesn't support the claim, remove or correct the claim before using the briefing.

Ask the tool to identify uncertainty, but don't make it the sole judge of its own work. When the output concerns something outside your expertise, involve a qualified reviewer or choose a different starting task. Use a second AI answer to raise questions, then check the evidence yourself.

Record one thing the tool did well and one thing you had to repair. Be specific. A note that it omitted an important condition will help you improve the next brief more than a general verdict that the answer was good or bad.

Thursday: repeat the task and compare the effort

Use the same method on another comparable piece of work. Include the time you spend preparing the input, writing instructions, checking the output and revising it. Compare that total with the normal method.

There's good reason to test individual tasks. In a preregistered experiment involving 758 BCG consultants using GPT-4 in 2023, participants completed suitable consulting tasks 25.1% faster on average, but on a separate task chosen to exceed the tool's capability they were 19 percentage points less likely to reach the correct answer. The study was published in Organization Science in March 2026. Read the published research.

Keep a short record of the tasks that work for you and what you need to get a good result. Include the input needed, the checks required and the cases you'd handle another way.

Don't change tools every time an answer disappoints you. First check whether the brief was clear, the necessary information was available and the task was appropriate. If those conditions are met and the result is still poor, try another approved option or stop using AI for that task.

Keep the examples that failed as well. Work through one with a colleague and decide whether a clearer brief, better information or a different approach would help. Use what you learn on the next task.

Friday: turn your experience into a team decision

By the end of the week, you should have actual work to discuss. Bring an example of the original task, the AI-assisted result and the corrections it required. Ask the relevant manager whether this addresses a recurring problem their team wants to solve.

I'd use that discussion to decide whether to run a small team pilot. Choose work that happens often enough to test, give someone responsibility for it and agree on how you'll check the output. Agree on what improvement would justify continuing and what failure would pause the trial.

Here is a sample pilot brief:

We'll test drafting follow-up notes from approved sales-meeting notes. The sales manager owns the trial, and the account owner checks every draft before sending it. We'll compare total preparation and review time with our existing method, and record missing commitments or incorrect details. The tool won't send messages or update customer records. We'll review completed examples before deciding whether to expand the use.

Keep the boundary narrow enough that the team can explain it. If the proposed pilot requires several new integrations, broad access to customer information and autonomous external messages, reduce the scope before beginning.

Continue your own practice during the following week. Invite a colleague to review one of your outputs. Seeing where someone else questions your judgment is useful preparation for setting expectations across the business.

Give people rules they can apply

A short starting guide should answer practical questions: which account to use, what information is allowed, which tasks need approval, how outputs must be checked and whom to contact when something goes wrong.

It should also tell people what to do when they're unsure. For example, pause the upload and ask the named owner, or switch to a public document for practice. Make the answer easy to find in the place people already go for work instructions.

Keep the pilot guide short, and have the right people settle the questions behind it. If a use raises questions about a customer contract, sensitive data or a system connection, name the person who will resolve them and set a date for the answer.

As CEO, follow the same boundary yourself. If you upload information that everyone else has been told to keep out, your demonstration creates an exception people will have to interpret. Bring an approved example and show the checking you did, including what you rejected.

Give learning a place in the working week

Put regular practice in your calendar and ask the pilot manager to do the same for participants. Tell them which existing priority can move to make room. Training and experimentation require effort, and that effort belongs in the cost of deciding whether a use is worthwhile.

Use a short review to discuss completed work and unresolved problems. A useful contribution might be a clearer brief, a discovered error or a decision to abandon a weak use. Don't judge the exercise by prompt counts or enthusiasm.

Give the work a clear purpose, agree on the limits and start. Build your wider AI plan from the tasks you've tried and the results you can show.

Use our AI training walkthrough to help with your practice. When you've a team use worth testing, The Leader's Guide to Implementing AI covers the responsibilities and operating decisions that follow.

Before you finish today, choose the task you'll try and put the next practice session in your calendar. In two weeks, show your team what you've learned and choose a useful task to work on together.

Research checked September 26, 2026. The publication date above is the original article date.

Sources

Enterprise privacy at OpenAI
OpenAI. Updated January 8, 2026; checked September 26, 2026.

Is my data used for model training?
Anthropic Privacy Center, commercial products. August 18, 2026; checked September 26, 2026.

Generative AI in Google Workspace Privacy Hub
Google Workspace Help. Article update August 14, 2026; page footer updated September 24, 2026; checked September 26, 2026.

Enterprise data protection in Microsoft Copilot and Microsoft Copilot Chat
Microsoft Learn. Updated August 18, 2026; checked September 26, 2026.

Artificial Intelligence Risk Management Framework: Generative Artificial Intelligence Profile
National Institute of Standards and Technology, NIST AI 600-1. July 2024.

Navigating the Jagged Technological Frontier: Field Experimental Evidence of the Effects of Artificial Intelligence on Knowledge Worker Productivity and Quality
Fabrizio Dell'Acqua and colleagues, Organization Science, INFORMS. Published online March 11, 2026; experiment conducted in 2023.