Use the current UK framework
The Data (Use and Access) Act changed the UK GDPR’s automated-decision provisions, with section 80 taking effect on 5 February 2026, subject to relevant saving arrangements. The current framework is in Articles 22A to 22D. Older explanations of Article 22 can therefore misdescribe the position for a new UK deployment.
It is not accurate to say that every significant solely automated decision is generally prohibited under the old model. It is equally inaccurate to say that the change gives businesses unrestricted permission to automate decisions about people. Lawfulness, fairness, transparency, accuracy and the specific safeguards remain important. Special-category information brings additional restrictions that require careful assessment.
Separate profiling from the decision it informs
Profiling involves automated processing of personal information to evaluate aspects of a person. It may support a decision without itself being the final decision. A system that sorts enquiries by likely interest is not necessarily equivalent to one that automatically rejects someone for an important service.
Map the full chain: input information, inferred attributes, score, threshold, action and any review. A supplier may describe its product as providing recommendations while your workflow applies every recommendation automatically. The practical effect matters. Ask what staff actually do with the output and whether a person’s circumstances can change the result.
Identify significance and meaningful human involvement
Article 22A addresses decisions producing legal effects or similarly significant effects. It also makes meaningful human involvement central to whether a decision is solely automated. A token approval click does not provide a sensible operational assurance that the decision has really been reviewed.
Our recommended assessment records the consequences for the individual, not just the financial value to the business. Would an incorrect outcome exclude someone from employment or an important service, or materially affect their circumstances? For human involvement, consider whether the reviewer understands the relevant information, has time to evaluate it and has genuine authority to change the outcome. These are practical questions for examining the process, not a substitute for the statutory test.
Treat sensitive inputs as a separate legal issue
Article 22B restricts significant solely automated decisions based entirely or partly on special-category data. One route requires explicit consent for all personal-data processing supporting the decision. The alternative combines contract necessity or legal authorisation with the substantial-public-interest condition in Article 9(2)(g). Contract necessity alone is not enough. Separately, a significant solely automated decision cannot rely wholly or partly on the recognised-legitimate-interests basis in Article 6(1)(ea). These conditions need assessment alongside the other applicable duties.
Look beyond fields explicitly labelled as health, ethnicity or another protected category. Inferences and combinations of information can create sensitivity too. A business should understand what the model is deriving and using, rather than review only the original spreadsheet headings. Where the intended use involves sensitive information or substantial consequences, specialist assessment is prudent before implementation.
Build the safeguards into the service
Article 22C requires safeguards for significant decisions based solely on automated processing. These include information about the decision, the ability to make representations, human intervention and the ability to contest it. A business needs a real route through which those safeguards can operate, not merely a statement that a customer may contact support.
For implementation, decide who receives a challenge, what records they can access and how they can correct an error or revise the result. Test the route with a realistic disputed case. The reviewer should be able to explain the relevant outcome without relying on a supplier’s opaque statement that the algorithm has spoken. Keep the original decision and review trail controlled and intelligible.
Test accuracy and unfair effects before scaling
An automated system can apply an inaccurate rule consistently. That is not the same as making reliable decisions. Examine the quality of input information, missing values, outdated records and the suitability of a model for the population using your service. A pilot should measure errors and consequences, not merely processing speed.
Our practical recommendation is to define unacceptable outcomes before testing. Look for patterns in who is incorrectly excluded or escalated, and consider whether apparently neutral variables operate as proxies for characteristics that deserve closer scrutiny. Data protection assessment does not replace equality, employment, consumer or sector-specific obligations. A deployment may need advice across more than one legal framework.
Make supplier evidence part of procurement
Ask a supplier what can be explained, logged, corrected and overridden. Establish whether your business can retrieve the information needed for an individual challenge and whether material model changes will be notified. A tool that cannot support the necessary safeguards may be unsuitable for the proposed use, however attractive its headline performance.
Also check data protection roles, retention, training uses and overseas access. A demonstration using harmless sample information does not show how the product will behave when connected to live records. Keep the contractual promises and technical capabilities aligned with the actual workflow you intend to introduce.
Keep the decision under review
High-risk processing may require a DPIA before it begins. Record the assessment, alternatives and mitigations, and consult the ICO where the applicable unresolved-high-risk test requires it. Review after changes to the model, input information, thresholds or purpose.
The central management question is whether the organisation can justify and correct the decisions it makes about people. Automation should not create a gap in accountability. If nobody can explain how a consequential outcome occurred or arrange meaningful reconsideration, the business has an operational problem to solve before it expands the system.
Sources and further reading
Primary guidance behind this briefing. Requirements depend on the facts; check the current guidance before making a decision.
This is general guidance, not a legal opinion on your circumstances or a guarantee of compliance. An individual assessment may be needed, particularly where sensitive information, disputed rights or regulatory action is involved.