What your agent is allowed to decide on its own, since February.

If you are building anything that decides — who gets a callback, which application is progressed, what a customer is offered — the rule you were told about is no longer the rule.

The Data (Use and Access) Act 2025 received Royal Assent in June 2025, but an Act does nothing until it is commenced. The Commencement No. 6 Regulations brought the bulk of the data protection provisions into force on 5 February 2026, including section 80, whose heading is simply Automated decision-making. Section 80 substitutes Article 22 of the UK GDPR with four new articles.

The four articles, in plain terms

Article 22A defines the terms: what counts as a decision based solely on automated processing, and what counts as a significant decision — one producing legal effects or similarly significant effects for the person.

Article 22B sets the restrictions, and this is where special category data keeps its tighter treatment. A decision that turns on health, biometrics, ethnicity, religion, trade union membership, sex life or sexual orientation is not in the permissive regime.

Article 22C sets the safeguards, and it is the operative article for anyone building. The controller must, as a minimum, provide information about decisions taken, enable the person to make representations, enable them to obtain human intervention, and enable them to contest the decision.

Article 22D allows further provision to be made by regulations, which is Parliament reserving the right to change the shape later.

What "solely automated" is doing, and why people get it wrong

The old comfort blanket was the human in the loop: put a person at the end and the rule does not apply. That was always weaker than it sounded, and it is weaker still now that the safeguards are the point rather than the exception.

A person who approves ninety-eight of a hundred recommendations in a shift, without the information needed to disagree, is not meaningful human involvement. They are a signature. The question a regulator or a court will ask is not whether a human clicked, but whether the human could have decided otherwise — whether they saw the inputs, understood the basis, and had time and authority to overrule.

That is a design question and it has a design answer. Show the reviewer what the system saw. Show them what it rejected. Record what they changed. Measure the override rate and be suspicious when it is zero. This is what the audit layer is for, and it is why we insist a system records what it refused to do as well as what it did.

Four things to do this month

1. Find your significant decisions. Not your AI. Your decisions. A rules engine written in 2019 that declines applications is in scope; a large language model that drafts an email is not.
2. For each, answer Article 22C in one line each. How does the person find out? How do they make representations? How do they reach a human? How do they contest it? If any line is blank, that is the work.
3. Check for special category data in the inputs, including the inferred ones. A model that infers a health condition from a claims history has brought special category data into a decision without anyone entering it.
4. Update the privacy notice. The Article numbers in it are wrong as of February, and a notice citing a repealed article is the cheapest possible finding against you.

What this does not do

It does not make automated decisions safe, and it does not make them lawful on its own — the rest of data protection law still applies, as do sector rules. For an FCA-regulated firm the Consumer Duty is the binding constraint long before Article 22C is, and the two ask overlapping questions in different vocabularies.

It also does not settle what "meaningful" human involvement is. That will be settled by cases, and the firms that come out of those cases well will be the ones that kept the override rate and can show it moved. If you are choosing where to start, the triage tool asks whether a decision should be automated at all before asking how, and the evidence work is the part most firms skip.