Three questions have started coming up in almost every conversation we have with business owners. Which AI should we be using? How does it fit with Microsoft 365? And what do we do about staff using whatever AI they fancy on free accounts?
They are the right questions, and the third one is more urgent than most people realise. So here are straight answers to all three, in plain English.
Start with the uncomfortable truth: your team is already using AI
Whether or not your business has "adopted AI", your staff almost certainly have. Somebody is pasting customer emails into a free chatbot to draft replies. Somebody is summarising a contract. Somebody is tidying up a quote. Ask around and you will find it, usually done with good intentions by your most productive people.
The problem is not the AI. It is the free personal account. Information pasted into a free consumer AI account leaves your control: it may be stored indefinitely, it may be used to train future models, and it sits in an account your business cannot see, manage or close when that employee leaves. If what gets pasted includes customer details, pricing or contracts, that is company data walking out of the building, and under UK data protection law it is still your responsibility after it leaves.
This has a name, shadow AI, and banning it does not work. People just get quieter about it. The fix is giving them a safe, approved way to do what they were going to do anyway.
Which AI should a business actually use?
Here is the honest answer: for most businesses, the account type matters more than the brand. The big assistants (Microsoft Copilot, ChatGPT, Claude, Gemini) are all capable. The meaningful line is between free personal accounts and paid business accounts.
Business versions commit, contractually, to not training on your data. They give you central control: you decide who has access, you can see it is being used, and the account belongs to the business rather than to whoever signed up. That single difference addresses most of the risk above, and it is why "which AI?" is usually the wrong first question. The right first question is "on what account?"
As for the brand: the good news is that all three of the leading assistants can be made to work properly with Microsoft 365. One does it natively, two need some setting up, and done correctly any of them can be run securely. Which brings us to question two.
How AI fits with Microsoft 365: three routes in
Copilot is the native route. It is built into the tools your team already uses: drafting in Outlook and Word, summarising meetings in Teams, working through numbers in Excel. There is nothing to connect, your data stays inside the Microsoft 365 service, and it is licensed as a per-user monthly add-on, so you can start with the handful of people who would benefit most rather than buying it for everyone on day one.
ChatGPT and Claude are the configured routes. Their business and enterprise versions can be connected to Microsoft 365, so the assistant works with your documents and email where they live rather than staff copying text out by hand. That connection takes additional configuration to set up properly, but done correctly it is secure: access follows your existing Microsoft 365 permissions, and the business agreements ensure your data is not used to train the models.
So the choice between the three is rarely about safety. Set up correctly, all of them can be run securely. It is about fit: Copilot if you want AI inside the Office apps themselves with nothing to bolt on, ChatGPT or Claude if your team prefers those assistants or already works well with them.
One catch that applies to all three and rarely makes any sales pitch: a connected assistant can see whatever the person using it can see. It respects your existing Microsoft 365 permissions exactly. That sounds reassuring until you remember how many businesses have years of loosely shared folders where everyone can technically open everything. AI makes sloppy permissions visible, instantly, because suddenly anyone can ask for what they could always technically reach. Tidying the tenant first is not optional homework, it is the prerequisite, and it is the same housekeeping we covered in our Secure Score guide.
What to actually do, in order
- 1. Write a one-page AI policy. Not a legal tome. What staff may use AI for, what must never be pasted in (customer personal data, financials, contracts, credentials), and which tools are approved. One page, plain words.
- 2. Give people an approved tool. A business AI account your team can actually use beats a ban every time. People stop improvising when the official route is good.
- 3. Use business accounts, never free ones, for anything touching company data. The training and control commitments are the whole point.
- 4. Whichever assistant you choose, tidy the tenant before you connect it. Permissions, sharing links, old accounts. Then pilot with a few people and measure whether it earns its keep.
- 5. Revisit in six months. This field moves quickly. A policy written today should expect to be revised, which is also true of this article.
The question to ask your team this week
Not "is anyone using AI?", which invites silence, but:
"What are you already using AI for, and what would make it more useful?"
Asked without blame, this gets honest answers. You will learn where the shadow AI is, which jobs your team actually wants help with, and where a proper business tool would pay for itself. That is a far better starting point than any vendor pitch.
How we help
We help businesses get this right as part of our Microsoft 365 and cloud work: an honest look at which of the three routes fits how you work, the tenant tidy-up that has to come first, Copilot licensing or the ChatGPT and Claude connections configured securely, and a plain-English AI policy your staff will actually read. It sits alongside the security side, because as of this year the two are inseparable.
If your team is already experimenting and you would rather steer it than discover it, that is exactly the right instinct. We are happy to talk it through, no jargon, no hype.