[Firm name]: risk assessment (DPIA) for using AI tools
Prepared by: [name, role]. Date: [date]. Data protection lead: [name]. Version: [1.0].
1. What we're doing (description)
Which AI tools and plans: [for example ChatGPT Business, Microsoft Copilot signed in with work accounts].
What we use them for: [for example drafting client letters, summarising documents, checking spreadsheet formulas].
Whose personal data could be involved: [clients, clients' employees, our staff, others].
What kinds of personal data: [names, contact details, financial details, identifiers such as NI numbers or UTRs, any of the most sensitive kinds (special category data)].
Where the data goes: [the AI provider, the country it's stored in, and any sub-processors].
Who has access, and how long it's kept: [roles, and the provider's retention settings].
2. Why we need it (necessity and proportionality)
The problem it solves: [for example, time spent drafting routine letters].
Our lawful basis: [for example, legitimate interests, with a short note of the balancing test].
Why this is proportionate: [why we need personal data at all for each task, and how we keep it to the minimum].
Other options we considered: [for example, not using AI, or using it only with no personal data].
3. Risks to people
For each risk, note how likely it is and how serious it would be: [low, medium or high].
- Personal data shared with a provider that has no contract with us (personal or free accounts).
- Data used to train the provider's AI.
- Data stored outside the UK without the right safeguards.
- Inaccurate output relied on, affecting a client.
- The most sensitive data (special category data) entered without the extra care it needs.
- Staff not knowing the rules, or not being told how their own use is recorded.
4. Measures to reduce the risks
For each risk, the measure and who owns it: [measure, owner, date].
- Approve business plans only, with a data processing agreement, and block or warn about personal accounts.
- Check each provider's training setting and data location.
- Keep personal data out of prompts where the task doesn't need it.
- Publish an AI policy, train the team, and record who has read it.
- Check AI output before it's relied on.
- Tell staff what's recorded about their AI use, and why.
Remaining risk after these measures: [low, medium or high].
5. Sign-off
Measures approved by: [name, role, date].
Remaining risk accepted by: [name, role, date]. If the remaining risk is still high, consult the Information Commission (ICO) before you start.
Data protection lead's advice: [summary, and whether it was followed].
Next review: [date], or sooner if we add a tool or a provider changes its terms.
Worked example: a 10-person accountancy practice (made-up example)
Made-up example, not a real client. Harbour & Lane Accountants has 10 people. It uses ChatGPT Business and Microsoft Copilot signed in with work accounts.
Description: drafting client letters and tax reminders, summarising HMRC letters, and checking payroll formulas. Personal data: client names, contact details, UTRs, NI numbers, bank details and payroll data. Both providers are business plans with a data processing agreement, and neither trains on what's typed by default.
Necessity: the AI saves around [x] hours a week on routine drafting. For most tasks the AI doesn't need real client details, so the practice decides to keep them out of prompts where it can.
Main risks: a member of staff using a personal ChatGPT account for client work (likely, serious); payroll data pasted to check a formula (likely, serious); an HMRC letter summary relied on without checking (possible, moderate).
Measures: personal AI accounts blocked for client work; client details swapped for placeholders before prompts are sent; an AI policy published and acknowledged by all 10; output checked before it's sent to a client. Remaining risk: low.
Sign-off: approved by the managing partner, [date]. Review in six months.
