Copilot, ChatGPT, Claude or Gemini: how to choose for your team

There’s no best AI assistant for every team, whatever the adverts say. Microsoft 365 Copilot, ChatGPT, Claude and Gemini are all capable, and the order changes every few months, so any ranking we printed would be stale by the time you read it. Choose on things that last: where your work lives, how your data is handled, how each one does on your own tasks, what it connects to, and what it costs per person who uses it. We’ve left out prices and plan names on purpose, because they change faster than articles do.

Start with where your work already lives

Convenience matters, because an assistant nobody opens saves nobody any time.

• Microsoft 365. Copilot works inside apps such as Word, Excel, PowerPoint and Teams. Check which version you’re being offered: Microsoft offers more than one, and only some of them draw on your organisation’s files automatically. With the others, people upload files by hand.
• Google Workspace. Gemini is built into Gmail, Docs, Sheets and Meet.
• Either, or both. ChatGPT and Claude are standalone assistants. Both connect to Google Drive and to Microsoft services such as SharePoint, Outlook and Teams, through what their makers currently call connectors or apps.

Being built in is an advantage, not a verdict. It doesn’t tell you which one writes the better client letter. That’s what the trial is for.

Check the data terms yourself

At the time of writing, all four vendors say on their own privacy pages (linked below) that they don’t use business customers’ data to train their models by default. Read those pages yourself, for the plan you’re actually buying, because the details differ and they change. Questions worth asking each vendor:

• Is our data used for training, and does anything we do, such as sending feedback, change that?
• How long are conversations kept, and can we change that?
• Where is our data stored, and where are our prompts processed? They can have different answers.
• What can an admin control: who has access, which features are switched on, what gets logged?

Free and personal accounts come under different terms, so make sure work happens in the business account. The ICO’s advice on third-party tools is blunt: if you can’t tell whether one is compliant, choose a different one. Our UK GDPR checklist goes further.

Test them on your own work

Demos and benchmark tables tell you little about your Tuesday afternoons. Run a two-week trial instead.

1. Pick three to five real tasks your team does every week, such as summarising a long email thread, drafting a letter from notes, pulling figures out of a spreadsheet or turning meeting notes into actions.
2. Pick a small group of people who do those tasks, including at least one sceptic.
3. Run each task through two or three assistants with the same inputs.
4. Score every output the same day, on a scorecard like this one.

Score each from 1 to 5	Assistant A	Assistant B	Assistant C
Accuracy: no mistakes we had to catch
Edits needed (5 means none)
Time saved compared with doing it ourselves
Works where we work: our apps and files
Admin controls and data terms
Still being used in week two
Total

Treat accuracy as a gate rather than one score among six: an assistant that’s quick but wrong is still wrong. And write the numbers down as you go. When the UK government gave out 20,000 Copilot licences across 12 organisations in late 2024, users reported saving an average of 26 minutes a day. That’s a useful finding, but it was self-reported, and memories can be generous. Your own scorecard, filled in on the day, is better evidence. We cover measuring in how to tell if an AI project worked.

Check the integrations you need

Beyond email and documents, think about the systems your team spends its day in: a CRM (your customer database), case management, accounts, project tools. Connections to other software vary a lot between vendors, change often, and many need an admin to switch them on. Write down the three systems you’d most like the assistant to read from, and check each vendor’s current documentation for those specifically. Ask whether each connection respects people’s existing permissions, so nobody can read through the assistant what they couldn’t open directly. Our piece on how retrieval works explains why that matters.

Work out the cost per active user

The price is per seat. The real cost is per person who uses it. If you buy 20 seats and eight people use them in a typical week, you’re paying two and a half times the list price for each active user. Check usage after the first month and move seats to the people who use them. Count the time it takes to train people, too.

You can mix tools

You don’t have to choose one assistant for everyone. One sensible pattern is the built-in assistant for most people (Copilot or Gemini, depending on your suite), plus a few seats of a standalone assistant for people who do a lot of drafting or analysis. The cost of mixing is admin: two sets of terms to check, two places to manage access, and a policy that covers both. Keep that policy short. Our guide to an AI policy your team will read shows how.

Before you buy, will this tool pay for itself? puts the hours it takes to adopt next to the subscription.