Vet an AI tool before your organisation uses it
An impressive demo isn’t due diligence. Before your organisation adopts an AI tool, check what happens to your data, the contract, the controls and whether it does the job.
Video transcript
Your organisation wants to roll out a new AI tool, and the demo looked great. Before anyone uploads real data, here’s what to check.
Start with the data. Does the provider train its models on what you put in? How long does it keep it? And where is it stored and processed?
Get those answers for the plan you would actually buy. The consumer and business versions of the same tool often handle data quite differently.
Then ask for the paperwork: a data processing agreement, the list of sub-processors, independent security evidence such as an ISO 27001 certificate or a SOC 2 report, and admin controls like single sign-on and audit logs.
Next, trial it with a small group on real tasks with nothing sensitive in them, and compare it fairly with how you work now.
Watch for red flags: vague promises about your data, consumer terms only, or a security page full of logos but no reports you can read.
If it will handle personal data, check whether UK data protection law requires an impact assessment before you start. Involve your data protection lead early.
Read the full guide below for the complete checklist, and a prompt that drafts your supplier questions.
In 30 seconds
- Start with data: does it train on your inputs, how long does it keep them, and where?
- Ask for a Jargon busterData processing agreement: A contract between your organisation and a supplier that handles personal data for you, setting out what it may do with the data and how it must protect it., the Jargon busterSub-processor: Another company that a supplier uses to handle your data, such as a cloud hosting firm. Suppliers should list them and tell you about changes. list and current, independent security evidence.
- Trial it on real tasks, check admin controls and exit terms, and do a Jargon busterDPIA: Data protection impact assessment: a documented check of the risks a project poses to people’s personal data, and how you’ll reduce them. UK GDPR requires one for high-risk uses. if needed.
A team has found an AI tool that could save hours a week, and the demo was impressive. Before anyone uploads a customer list, someone has to ask the unglamorous questions: where the data goes, what the contract says, who controls access, and whether it really does the job. This checklist is for that person.
Your data: training, retention and location
Start with three questions. Does the provider use what you put in to Jargon busterTraining: The stage where AI learns, by finding patterns in a huge number of examples. its models? How long does it keep your data, including deleted chats and backups? And where is it stored and processed?
Get the answers for the plan you’d actually buy, because terms often differ between the consumer and business versions of one tool. When Anthropic updated its consumer Claude terms in 2025, it said it would keep data for five years from users who let it train on their chats, rather than the usual 30 days. Its business plans run on separate commercial terms and aren’t used for training by default.
If data will leave the UK, the international transfer rules in Jargon busterUK GDPR: The UK law on personal data. It sets how organisations must collect, use and protect information about people. apply. Ask which countries are involved, and which legal safeguards the provider relies on.
The paperwork and the controls
Ask the supplier for
- A data processing agreementThe contract UK GDPR requires when a supplier handles personal data for you. It limits what they can do with it.
- The sub-processor listEvery other company that handles your data on the provider’s behalf, such as cloud hosts, and how you’ll hear about changes.
- Security evidenceIndependent proof, such as ISO 27001 certification or a SOC 2 report. Check its scope, its date and what it leaves out.
- Admin controlsCentral user management, audit logs and retention settings, so IT can control access and see what happens.
Controls often depend on the tier. On Claude’s business plans, for example, Jargon busterSingle sign-on: Logging in to many work apps with one company account, so IT can give and remove access to everything in one place. comes with Team, but audit logs and custom data retention need Enterprise. Ask for the security reports themselves, not a page of logos, and have whoever handles contracts read the data processing agreement.
Trial it on real work
A demo shows the best case. Run a time-limited trial with a small group, on real tasks that contain nothing sensitive, and compare the results with how you work now. Our guide to testing AI models yourself shows how to keep that comparison fair.
Check what it can connect to, such as email, files or your customer database, and what permissions each connection asks for. Prefer read-only access, and connections your admins can switch off centrally. Then cost the whole thing: per-seat prices, usage limits, charges for heavy use and the minimum contract term. Finally, ask how you get your data out, and how it’s deleted when you leave.
| Good sign | Red flag |
|---|---|
| Written terms: no training on your business data by default | Vague promises, or “we may use your data to improve our services” |
| A data processing agreement and a published sub-processor list | Only consumer terms, or no answer about sub-processors |
| Current ISO 27001 or SOC 2 evidence you can read | Logos on a security page, but no reports |
| Single sign-on, user management and audit logs | Shared logins and no admin view |
| A clear way to export and delete your data | No route out, or deletion nobody will confirm |
Do you need a DPIA?
If the tool will process personal data, check whether you need a data protection impact assessment, or Jargon busterDPIA: Data protection impact assessment: a documented check of the risks a project poses to people’s personal data, and how you’ll reduce them. UK GDPR requires one for high-risk uses.. UK GDPR requires one before processing that’s likely to result in a high risk to people, particularly when it uses new technology. That can include AI tools handling customer, staff or patient data, so involve your data protection lead early and see the guidance from the ICO, the UK’s data protection regulator.
I’m assessing [tool] for [what we’d use it for] at a [type and size of organisation] in the UK. It would handle [types of data, such as customer emails]. Write a due diligence questionnaire covering: training on our data; retention and deletion; where data is processed; the data processing agreement and sub-processors; security certifications; admin controls; integrations and permissions; data export; pricing and limits. Put the questions most likely to rule it out first.
Once you’ve chosen, set the ground rules for using it: see how to write an AI policy. If it will link to email or files, read connect AI to your apps, safely.
Check yourself
3 quick questions nothing is savedTools in this guide
Sources (3)
- Updates to Consumer Terms and Privacy PolicyAnthropic, August 2025
- Plans and pricingAnthropic
- Claude EnterpriseAnthropic
Spotted a mistake? Tell us and an editor will check it.
