Secure AI means weaker AI

Until your people learn where its real power is hidden

Force Copilot on people who never truly learned it, and they'll see a weaker AI and work around it. Let them use the tool they prefer, and they take your data with them.

Stop pretending Copilot is the best AI. Sell it on what it is: the same AI as ChatGPT and Claude, with a Microsoft layer around it. That layer makes it feel slower, but it's where the value sits: your email, files and meetings, inside Word, Excel and PowerPoint, safely.

The one-hour training everyone got never showed that. Learned properly, Copilot comes close to working with ChatGPT and Claude. Not equal, but close, and it keeps your data in.

When a team needs what Copilot can't do, buy them the business version of an AI tool that can, with the same data protection.

The legal consequences

Right now, every employee decides where your data goes, one prompt at a time.

And the legal part lands on you. Since February 2025, the EU AI Act requires you to support your people's AI literacy and keep a record of what you did. When personal data ends up in a public AI tool, the company answers for it under GDPR, not the employee.

Buying the tool your employees like sounds fair

Some leaders skip the Copilot licence and the expensive training, and buy the business version of ChatGPT or Claude. People already like it and know how to work with it. The contract protects the data. And no AI tool fits every job anyway, so why force one?

Fair. But in the companies I work with, the approved tool is almost always Copilot or Gemini. The platform you already run on made that choice for you, and undoing it is expensive: a second AI tool means a second contract, a second data agreement and a second training.

→ So what matters is whether your people learn what the one you've got can do. And it works differently from ChatGPT or Claude.

Out of the box, Copilot loses

Most people met AI at home first. In PagerDuty's 2026 survey, 89% of office professionals who use AI for work first came across the tool in their personal lives. So they learned ChatGPT or Claude: open a chat, ask, paste.

Then they try the same thing in Copilot. It feels slower. The answers feel thinner. The language feels generic and corporate. Back to the familiar tool. The same survey found that two-thirds have used AI tools at work they believed weren't allowed.

And let's not sugarcoat it: ChatGPT and Claude feel ahead. People aren't comparing the AI. They're comparing Microsoft's packaging, your company's settings and how well they learned to work with it.

→ Judged as a consumer AI tool, Copilot loses. That's the wrong test.

Its power sits in your own systems

Copilot's strength isn't the model. It's what it can see: your email, your Teams meetings, your files in SharePoint. It works inside Teams and Outlook, and moves between your apps without you copying anything across.

Ask it to prepare a meeting from the email thread, last week's notes and the deck in SharePoint. Ask it to find a file, an email or a vague conversation from months ago. A personal ChatGPT account doesn't, because it sees none of it.

You think your people are trained, but they never saw what Copilot can do

That power stays hidden until you learn the deeper layer and change how you work: setting it up for your work, digging into the corners of SharePoint and your inbox, using it as a sparring partner on confidential work. The more you work with it, the better you get, and the more of your work it has to draw on.

Standard training doesn't get there. Usually it's one session for everyone: the novice, the sceptic and the power user, finance and marketing. In Docebo's 2026 research, 85% of employees say their AI training doesn't help them use AI in their role. Nearly 60% say the programmes aren't designed with people like them in mind.

Bespoke training costs money. And you can't see the value you're not getting, so it rarely makes the budget.

→ A training for everyone fits nobody.

You never have to think twice before you paste

When I work in Copilot, I don't think about what I paste. When I use ChatGPT or Claude, I stop every time and check what's in it. For confidential business work, I don't use them at all. I don't want to be the one who causes a leak by accident.

Everyone knows they should pause. Nobody plans to skip it. Then, on a busy day, they paste the one document they shouldn't have.

Safety that's always on is the part that's undersold. It works before anyone's trained.

Start here

Ask three people at different levels one question: what do you use Copilot for, and what do you still do in ChatGPT or Claude?

If the first answer is meeting summaries and the second is everything else, you know two things. Where your data goes. And what your training missed.

For most of your people, close enough and safe is the better trade. Until they learn it, the tab they keep open decides where your data goes.

If your people still go around Copilot, I'll help you build training that fits their work, and be honest about what Copilot can't do. Email me at fleurwillemijnvanbeinum@gmail.com.

9 AI dilemmas

This is the second of nine AI dilemmas for the C-suite. I'm publishing them over the coming weeks. Each piece names the decision, what it costs to leave it open, and what I'd do.

The one you recognise immediately might be the one you've left open too long.