A data protection impact assessment, or DPIA, for an AI review tool is the written risk assessment that UK GDPR Article 35 requires before a practice starts processing personal data in a way that is likely to result in a high risk to individuals. This article is for partners and managers who are deciding whether the accounts files they plan to put through an AI tool trigger that duty, and what the assessment should cover when they do.
Why accounts files hold personal data
The working file behind a set of statutory accounts usually holds the names of directors and shareholders, directors’ loan account balances, dividends paid to named individuals, payroll records, and correspondence that mentions people by name. For a small owner-managed company, the figures and the individual are very close together.
That matters because the duty to carry out a DPIA attaches to processing of personal data, and the practice is the one deciding whether to send that data to a new tool. Our guide to client data, the UK GDPR and AI in an accountancy practice sets out the wider position on roles and contracts. This article deals with one question inside it: when does the practice need to write a DPIA, and what goes in it?
What the law and the ICO say about when a DPIA is needed
Article 35(1) of the UK GDPR requires an assessment of the impact of the planned processing on the protection of personal data where it is likely to result in a high risk to the rights and freedoms of individuals, and it refers in particular to the use of new technologies. The assessment has to happen before the processing starts. The same paragraph allows one assessment to cover similar processing operations that present similar high risks, which is useful for a practice that will use one tool across many client files.
Article 35(3) lists three cases where an assessment is always required: systematic and extensive evaluation of people based on automated processing, including profiling, where decisions with significant effects follow; large-scale processing of special categories of personal data or criminal offence data; and systematic large-scale monitoring of a publicly accessible area. A review tool that reads accounts is unlikely to fall under the first or third. The second needs a closer look, because payroll files can contain health information such as sickness absence, and the question is then one of scale.
The ICO publishes a further list of processing that it says requires a DPIA. Innovative technology, which the ICO says includes AI, appears on that list with a condition attached: it counts when combined with any other criterion from the European guidelines the ICO relies on. Separately, the ICO’s guidance on AI and data protection says that, in the vast majority of cases, the use of AI will involve a type of processing likely to result in a high risk, and that it must still be judged case by case. Where there is real doubt, the ICO’s advice on its DPIA page is to do one anyway.
Checked against the ICO pages cited below on 21 November 2026. The ICO says on several of them that its guidance is under review because of the Data (Use and Access) Act, so check the current text before you rely on a specific passage.
Does your tool trigger one?
The ICO describes a screening step before a full assessment: a high-level test for warning signs of high risk. For an AI review tool, the screening questions below follow the criteria the ICO sets out.
| Screening question | Why it matters | Where the answer sits |
|---|---|---|
| Is the tool an AI system, or a new use of an existing technology? | The ICO treats a novel application of existing technologies, including AI, as innovative technology. | Supplier documentation |
| Does it combine or match personal data from several sources? | Combining, comparing or matching data is on the ICO’s list. | Your own data flow map |
| Do the files include special category data at scale? | Article 35(3)(b) makes an assessment mandatory. | Payroll and HR content in the file |
| Do its outputs feed decisions about individuals? | Automated decisions affecting people raise the risk. | How review points are used |
For most practices the answer to the first question is yes, and the ICO’s position that AI use will in the vast majority of cases involve high-risk processing means the sensible default is to carry out the assessment. A practice that decides it does not need one should record the reasons, so the decision can be explained to a regulator, an insurer or a client.
Questions that are specific to accounts data
These questions are specific to an accounts file, and each one should be answered in the assessment or recorded as a gap.
- Which categories of personal data are in the files that will go to the tool, and could any of them be removed first? The ICO says you should process only the personal data you need for your purpose.
- Does the tool receive whole working files, or only the primary statements and notes? A narrower input puts less personal data in front of the tool.
- Who are the individuals affected: directors, shareholders, employees, and the people named in related party notes? Do any of them belong to a vulnerable group?
- Does the supplier keep the files, the outputs or the logs, and for how long?
- Will the supplier, or anyone it uses, process the data outside the UK? The ICO’s transfer guide requires a transfer mechanism and, where you rely on a safeguard, a transfer risk assessment.
- Is any of the data used to train or improve a model that other customers use?
- What happens to the output? A list of review points that stays with the reviewer is a different risk from one that is stored in a shared system.
- Does the engagement letter or privacy notice tell clients that their data may be processed with AI tools?
Our guide to choosing an accounts review tool and testing it covers the supplier questions in more detail.
What the assessment should contain
Article 35(7) sets the minimum content: a systematic description of the processing and its purposes, an assessment of the necessity and proportionality of the processing operations, an assessment of the risks to individuals, and the measures envisaged to address those risks, including safeguards and security measures. The ICO’s AI guidance adds detail on how to apply this to AI. It asks the controller to describe the data flows and the data sources, to consult the people affected and the processors involved, to show that AI is needed and that the purpose could not be met in a less intrusive way, to score each risk by likelihood and severity, and to record the residual risk after mitigation.
A worked example
This is a hypothetical, not a template. A practice wants a tool to read the draft accounts and the supporting disclosure notes of small owner-managed companies and to list inconsistencies for the manager. It decides to send the primary statements, the notes and the trial balance, and to leave out payroll detail and correspondence.
| Element | Hypothetical entry | Residual risk |
|---|---|---|
| Purpose | Reading drafts for inconsistencies before manager review | Low |
| Data in scope | Directors’ names, loan balances, dividends, related party notes | Medium |
| Data removed first | Payroll by employee, emails, bank statements | Reduced |
| Supplier role | Processor under a written contract; no use of the data for its own purposes | Medium until the contract is signed |
| Output | List of review points held on the file | Low |
The figures in a real assessment would come from the practice’s own scoring method. In the example, the data minimisation decisions sit in the assessment, next to the risk they reduce.
After the assessment: consultation, review and records
Three further points from the ICO’s DPIA guidance are worth building into the process.
- Contracts with processors should require them to assist with the assessment, so the supplier agreement is part of the DPIA work and not a separate exercise.
- If a high risk remains after mitigation, the ICO says you must consult it before you go ahead with the processing.
- The assessment is a live document: the ICO says you need to keep it under review, and to repeat it if the nature, scope, context or purposes of the processing change substantially. A new feature, a new supplier location or a new file type is a reason to look again.
The ICO also describes publishing a DPIA, or a summary of it, as good practice. A practice may prefer to give clients a short plain statement in the engagement letter and keep the full assessment on file.
Responsibility for the accounts stays where it was: the board approves them and the partner signs the review, as our guide to signing off accounts that an AI tool has reviewed explains.
Questions practices ask
Do we need a DPIA for every AI tool we try?
Not necessarily for every tool, but it is the safer starting point. The ICO says the use of AI will in the vast majority of cases involve processing likely to result in a high risk, and that you should do an assessment if you are in doubt. A tool that never receives personal data is the exception, and the practice should be able to show that it never does.
Can one DPIA cover several clients?
Yes, where the processing is similar and presents similar high risks. Article 35(1) says a single assessment may address a set of similar processing operations. Keep a record of the files and client types it covers and revisit it when they change.
Who is responsible for the DPIA, the practice or the supplier?
The controller carries the duty. The ICO says organisations that determine the purposes and means of processing will be controllers, and in its worked example of a cloud provider supplying machine learning tools to clients it treats the clients as controllers and the provider as likely to be a processor. The practice should confirm its own position in the engagement terms, and ask the supplier to help with the assessment.
What if the assessment shows a high risk we cannot reduce?
The ICO says you must consult it before going ahead. In practice the more usual response is to narrow the data sent to the tool, change the supplier terms, or decide not to proceed with that use.
Is a pilot on past files covered by the same duty?
Yes. Putting signed files through a tool is processing, so the assessment, the contract and the data minimisation choices come first. Our guide to quality management for accounts preparation covers evidence of review on file.
Sources
- UK GDPR, Article 35: data protection impact assessment, legislation.gov.uk, read 21 November 2026
- When do we need to do a DPIA?, Information Commissioner’s Office, read 21 November 2026
- Examples of processing likely to result in high risk, Information Commissioner’s Office, read 21 November 2026
- How do we do a DPIA?, Information Commissioner’s Office, read 21 November 2026
- What are the accountability and governance implications of AI?, Information Commissioner’s Office, updated 15 March 2023
- How should we assess security and data minimisation in AI?, Information Commissioner’s Office, read 21 November 2026
- A brief guide to international transfers, Information Commissioner’s Office, updated 15 January 2026