Your team is already using AI, and most of them haven't told you. That's what happens when a useful tool arrives faster than a policy and people have work to finish by Thursday. You can't stop it from happening; what you can decide is whether you find out while you're still able to shape it.

We've sat in enough of these conversations to know how they usually go. Someone senior asks whether staff are using ChatGPT. There's a pause. Someone says they assume a few people are. A policy gets drafted, circulated and ignored, and the organisation ends up carrying all of the risk with none of the visibility.

A Tuesday night, a tender and a free chatbot

Take a composite case. Priya is the bid manager at a forty-person engineering consultancy. A tender is due on Friday, the scoring criteria run to nine pages, and at 9pm on Tuesday she's on the sofa with her phone. She pastes the criteria into a free chatbot, then the draft method statement, then (because the pricing section reads badly) the table of subcontractor day rates. Twenty minutes later she has a much tighter response and goes to bed relieved.

Nothing about this is reckless from where she sits. She used a tool that worked, on her own time, to do the part of her job she's judged on. She won't mention it on Wednesday, partly because nobody asked and partly because it feels like admitting she needed help.

From where the firm sits, commercially sensitive rates are now in a consumer account with retention settings nobody at the firm has read. Whether that matters depends on the tool and its settings, but nobody at the firm knows which tool it was.

How common is this?

Very. Microsoft and LinkedIn's 2024 Work Trend Index, a survey of 31,000 knowledge workers in 31 countries, found 75 per cent using generative AI at work, and 78 per cent of those users bringing their own tools rather than waiting for one to be provided (Microsoft).

The figure that should change your policy is the third one. In the same survey, 52 per cent of people using AI at work said they were reluctant to admit using it for their most important tasks. Priya is the typical case. People don't hide that they asked a chatbot to reword an email; they hide that it helped with the tender, the board paper or the client report, which is exactly where your risk and your opportunity both sit. A policy that assumes hidden use is rare will be wrong about the cases that matter most.

What a ban would have done

Suppose the firm had banned AI tools in March. Priya's Tuesday night looks exactly the same, because the office firewall doesn't reach her phone on mobile data. The ban changes three things, all for the worse.

She now has a reason to stay silent even if she later wonders whether pasting the rates was wise. The question that might have prevented a problem ("is it all right to put pricing into this?") never gets asked, because asking means confessing.

If something does go wrong, say a rate turns up somewhere it shouldn't, the firm hears about it from a client weeks later rather than from Priya the next morning. You want near misses reported on the day, and bans teach people to bury them.

And the firm loses its map. It can't see that bid writing is where AI is helping most, which is the information it would need to decide what to provide properly.

Bans appeal because they're quick, they signal seriousness and they cost nothing to write. In practice they move the use somewhere you can't see it.

What would have helped instead

Asking about tasks. People won't volunteer "I use a chatbot for bids". They'll happily tell you the scoring matrix takes four hours to answer and most of that is restructuring text. Ask about the shape of the week, anonymously and in writing:

  • Which parts of your week feel like copying and pasting?
  • Which task would you hand over tomorrow if you could?
  • What would you never put into an AI tool, and why?

The last question is the most useful, because it shows where people's instincts are already good and where their line differs from yours. It often differs in both directions: staff tend to be more careful than you'd expect with client data and less careful with internal pricing and strategy. Publish a summary and say thank you. If it feels like an audit, the answers become useless.

A rule about data, not brands. A list of approved products is out of date within a quarter, and it doesn't answer the question Priya had at 9pm, which was "can I paste this?" A one-page rule built on three tiers does:

  • Public: already published, or would be fine if it were. Use any tool.
  • Internal: strategy, pricing, unpublished plans, rates. Only in a tool the firm has chosen and can account for.
  • Personal and confidential: client details, employee records, health information, anything under a confidentiality clause. Never in a general tool. Use placeholders such as "Client A", "Subcontractor B" or "the candidate", and put the real names back yourself.

Under that rule, Priya would have known the method statement was fine and the rate table wasn't, and she'd have swapped the figures for placeholders in thirty seconds. There's more on retention and where data goes in what happens to your data.

A tool she'd choose at 9pm. The firm's approved tool has to beat the free one on her phone for convenience, and a chat window is a low bar to clear. Anything that needs a new login, a new app and a new habit has to be clearly better. We hold Able° to the same test, which is why the agent it gives you lives in a chat app you already use. If the approved option is slower or worse at the ordinary task, people will keep using whatever lets them finish, and your policy becomes something everyone signed and nobody follows.

Shouldn't we just tighten up?

The case for a harder line is real. Stanford's 2026 AI Index counted 362 documented AI incidents in 2025, up from 233 in 2024, and that only covers the ones that became public (Stanford HAI). One paste of a client list into a consumer tool can be a reportable breach under GDPR or your local data-protection law. In regulated work, the professional body's rules apply to the output whatever produced it.

There's also a fair charge against the permissive approach: it can be abdication dressed up as trust. Staff aren't data protection officers, and asking them to classify data with no examples and no training leaves them guessing.

We agree with all of that, and it still points away from prohibition. Risk exists either way; a ban only removes your view of it. The workable middle is more effort than either extreme: a short data rule, a usable default, a named person to ask, and a habit of discussing near misses without punishing whoever reports them.

Where to start

Send the three questions this week and give people a week to answer. While you wait, draft the one-page data rule with one example per tier and the placeholder trick spelled out. Then use the approved tool yourself on a real task for a few days; if you wouldn't reach for it under time pressure, fix that before asking anyone else to.

Last, name a person people can ask, someone who'll reply within a day and won't treat the question as an admission. The employer check covers what to ask before standardising on a tool, and your first week with an AI agent is written for whoever ends up owning it.

Common questions

Should we pause AI use until we have a policy?

A short pause while you write one page of rules is reasonable, as long as it has an end date. An open-ended ban just moves use onto personal devices. Set the date, write the data rule and lift the pause when it's done.

Is using an unapproved tool a disciplinary matter?

Usually not, unless someone ignored a clear rule about confidential data. Treating ordinary use as misconduct guarantees the next person hides the next mistake. Keep discipline for the data rule and treat tool choice as a conversation.

Do we need to train everyone?

Everyone needs the data rule, which takes fifteen minutes and fits on one page. Deeper training pays off most for the two or three people whose work changes most, and for the managers who'll be asked "is this allowed?"

How do we know the policy is working?

Count questions asked rather than violations found. If nobody has asked the named person anything in a month, people are still deciding alone. A rise in questions is the signal you want.