Which AI assistant should our organisation use?

Why a ranking will not help you

Comparison tables go out of date quickly. New models arrive several times a year, features move between plans, and a tool that led on one test may trail on the next. Your question is narrower and more useful: which tool fits our work, our data rules and our people?

It also helps to know that the brand on the screen is not necessarily the model underneath. Microsoft, for example, states that it offers third-party models from Anthropic and OpenAI inside Microsoft Copilot, which admins can choose to allow, and that additional terms may apply. Field Guide 03 explains models and thinking modes in plain terms.

Would it help to compare two assistants on your own team's real tasks before you commit? Book a free AI Value Briefing and we will look at it with you.

Start from where your work already lives

An assistant is far more useful when it can see the documents, emails and meetings people actually work with. So look first at your existing platform:

• Microsoft 365: Microsoft Copilot works through Microsoft Graph with your emails, chats and documents, inside apps such as Word, Excel and Outlook, and only surfaces content a user already has permission to view.
• Google Workspace: Google describes Gemini in a side panel across Docs, Sheets, Slides, Drive, Gmail and Meet, retrieving only content the user can access.
• A standalone assistant such as Claude, ChatGPT or Grok can still reach your files and systems through connectors. Our article on connecting AI to other systems covers how to do that safely.

Being built in is an advantage, not a verdict. Some teams choose a standalone assistant because it suits writing, analysis or coding work better in their own pilot.

Check the business terms and the admin controls

Personal and business plans of the same product can follow different data rules, so decide the plan before the brand. Our article on whether your data will train AI compares what vendors publish. As examples of what to look for:

• Microsoft states that prompts, responses and data accessed through Microsoft Graph are not used to train foundation models.
• Google states that Workspace does not use customer data for training models without the customer's prior permission or instruction.
• Anthropic's Claude Enterprise page says prompts and results are not used to train models by default.
• OpenAI's enterprise privacy page says it does not train on business data by default.

For any tool, ask the vendor to show you in writing: training use, retention, single sign-on, user management, audit logs, and who can switch on connectors. Have IT and your data protection lead check the answers.

Languages, connectors and the people who will use it

• Languages: if your teams work in several languages, test each one with real documents during the pilot. Quality can differ by language and by task.
• Connectors: list the three or four systems people need first (shared files, email, CRM) and check which tools connect to them on your plan, and with what permissions.
• People: the tool that staff find easy to open every day often beats a stronger tool they rarely use.

Run a short pilot on real work

A feature list shows what a tool can do. A pilot shows what it does for you. A simple version:

1. Pick five to ten real, repeated tasks from different teams, using non-sensitive or approved data.
2. Give a small mixed group two or three candidate tools on business plans for two to four weeks.
3. Score each task on quality, time saved and how much checking the output needed.
4. Ask users which tool they would keep, and why.

Field Guide 16 helps you decide whether a task needs a new tool at all, or just better use of what you already pay for.

One approved main tool, plus exceptions

Many organisations do well with one approved main assistant for everyone, plus a short list of agreed exceptions: for example a coding tool for developers or a research tool for one team. One main tool means one set of data rules, one training programme and one place to share what works. Exceptions go through a light request: what task, what data, who owns it. Review the choice once a year, because the market will have moved.

Goes deeper in our free AI Field Guide: 3. Choosing a model and a thinking mode, 16. Choosing the right level: use what you have before you build, 33. Claude: the models, the apps and when to use which, 34. ChatGPT: the models, the apps and when to use which, 35. Gemini: the models, the apps and when to use which, 36. Grok: the models, the apps and when to use which.