How to pick an AI writing assistant for a small team: a checklist

1. Define the jobs

Write down three to five tasks the tool must do, in plain terms: first drafts of support replies, editing grant applications, rewriting product pages, summarising long documents. Different jobs need different things. Editing your own text needs a good editor with tracked changes. Drafting from scratch needs strong prompting and templates. Working across long documents needs generous input limits. Our guide on what to check before you sign up covers the individual version of this step.

2. Classify the data

List what staff will paste in. Public marketing copy is low risk. Client names, contracts, health or financial details and anything under an NDA are not. The answer decides which plan tier you can use at all. If you handle personal data of people in the UK or EU, the ICO's guidance on AI and data protection is the regulator's own starting point, with chapters on accountability and DPIAs, transparency, lawfulness, accuracy and fairness. The ICO notes the guidance is under review following the Data (Use and Access) Act, so check the date when you read it. For the legal questions members raise most, see the thread on client data under GDPR.

3. Compare plan tiers, not products

The same product often has a personal plan and a workplace plan with different data terms. These are the questions to put to each shortlisted vendor, with what the vendors' own pages say today:

Question	Why it matters	What to look for in vendor documents
Is our content used for training by default?	Confidentiality	OpenAI's enterprise privacy page says business data is not used for training by default; Anthropic's commercial page says the same for its commercial products; Microsoft's Copilot documentation says prompts and responses are not used to train foundation models
Who can read our conversations, and why?	Internal and vendor access	OpenAI describes admin access to workspace conversations and limits on its own staff and contractors
Can an admin set retention and delete data?	Policy and GDPR	OpenAI says workspace admins control retention on its business plans; Microsoft says admins can set retention through Purview
Will the vendor sign a DPA?	GDPR processor terms	OpenAI says a DPA is available for Business, Enterprise and the API
Where is data processed?	Transfers and residency	Microsoft states EU traffic stays inside its EU Data Boundary, and says models from one outside provider are currently excluded from it
Single sign-on and member management?	Offboarding	OpenAI lists SAML SSO among its enterprise controls

Vendor pages change. Print or save the version you relied on and note the date. Our thread on what happens to your text in AI writing tools compares the training and retention settings in more detail.

4. Decide who may use it and for what

A short written policy beats a long one nobody reads. Cover: approved tools, banned data types, whether AI-assisted work must be disclosed to clients or readers, who checks facts and numbers before anything goes out, and what happens to drafts. Members have shared what theirs included in the thread on company AI writing policies.

5. Run a two week trial with real work

1. Pick two or three people and one tool per person, so you can compare.
2. Give each the same three tasks from step 1.
3. Score the outputs on accuracy, tone, editing time saved and how often a fact had to be corrected.
4. Check the admin side during the trial: add and remove a user, change a setting, export data.
5. Decide with the scores, not the demo.

6. Understand the pricing model before the price

Do not compare list prices from memory. Check each vendor's pricing page for the actual figures on the day you decide, and compare the model:

• Per seat: predictable, but you pay for people who barely use it.
• Usage-based or API: cheaper at low volume and flexible, but costs scale with use and someone has to build or buy an interface.
• Bundled in a suite you already pay for: convenient, but check which tier unlocks the AI features and the admin controls.
• Free tiers: fine for public text, but the data terms are often the personal ones.

7. Plan the exit

Ask how you export your data and conversations, what happens on cancellation, and how deletion works. If a writing assistant includes process tracking, such as the Authorship feature in Grammarly, note that Grammarly's documentation says admins on Business accounts can switch Authorship on or off for the whole team, and that tracked data is kept on its servers for up to 10 years. Whatever you pick, decide in advance who owns the records.