AEVOMIND INSIGHTS — AI AT WORK

AI employees and the UK's new automated decision rules: what makes a decision solely automated, which safeguards apply, and how to keep a person in the loop

15 November 2026 · 12 min read · By the AevoMind team

A small glowing green bead travelling along a line of light through a large gold-edged glass ring, standing for an automated decision passing through a point of human review.

Since 5 February 2026 the UK GDPR has new Articles 22A to 22D on significant decisions made without meaningful human involvement, with four safeguards when they apply. The ICO's draft guidance, still under consultation, says involvement must be active, timely, by a trained person with authority, and recorded. Most of what an AI employee does is not a significant decision. Software can keep a person in the decisions that matter, but it cannot decide which ones those are.

The short answer

Since 5 February 2026, UK data protection law has had new rules for decisions made without a person involved. Section 80 of the Data (Use and Access) Act 2025 replaced the old Article 22 with Articles 22A to 22D, and the commencement note on legislation.gov.uk shows the section in force from 5 February 2026. The key definition is short: “a decision is based solely on automated processing if there is no meaningful human involvement in the taking of the decision”.

For a business running AI employees or AI calling agents, the practical question is where the AI only prepares, and a person decides, and where the AI decides on its own. Most of what an AI receptionist does, such as answering, booking and following up, is not a significant decision about someone. Screening, scoring and refusing can be. A connected system can keep a person in the decisions that matter and record that they were there. It cannot decide which decisions are significant. That is a judgement, and a legal one.

A gold toggle switch on a green panel, standing for a choice between a decision made automatically and one a person takes.
Either a person sits in the decision, or the decision goes through on its own.

The three tests

The ICO’s guidance on automated decision-making is a draft: the ICO says it is “consulting on the Data (Use and Access) Act updates” (page updated 31 March 2026), so treat it as the regulator’s current thinking, not settled guidance. The draft says the rules apply only when three things are present together:

One point from the draft is worth underlining for AI buyers. Systems “do not need to involve complex algorithms, artificial intelligence (AI), or other types of advanced processing” to come into scope. A simple scoring rule can be in. And an AI that only follows a fixed human rule may not be.

What the Act says: restrictions and safeguards

The Act builds the regime in four articles. Article 22A gives the definitions above, and adds that when considering whether there is meaningful human involvement, “a person must consider, among other things, the extent to which the decision is reached by means of profiling”. Article 22B sets two restrictions:

The ICO’s summary of the Act says the reforms open up a wider range of lawful bases for significant automated decisions, potentially including legitimate interests, with safeguards. That is right, but the recognised legitimate interests exclusion is part of the picture, so “the full range” overstates it.

Then Article 22C sets the safeguards. Where a significant decision is based solely on automated processing, the controller must have measures that: “provide the data subject with information about decisions” taken about them; “enable the data subject to make representations about such decisions”; “enable the data subject to obtain human intervention on the part of the controller in relation to such decisions”; and “enable the data subject to contest such decisions”. Article 22D lets the Secretary of State add detail by regulations. We have not seen any made.

A line of light splitting into two paths, one passing through a gold ring and glowing bright and the other running on past it unchecked, standing for reviewed and unreviewed decisions.
A decision that is reviewed before it takes effect, and one that is not.

What “meaningful” human involvement means in the draft guidance

This is where the draft is most useful, and strictest. It says that for involvement to count, “you must ensure it is active and not just a token gesture”. A human should:

Three further points matter for how a business sets up an AI. First, the criteria apply every time: “Using ad hoc spot-checking isn’t sufficient” because some decisions will go unchecked. Second, timing: the involvement has to come before the decision is applied to a person, while it can still be changed. Third, the people who built the system do not count: “A human merely designing or building an automated system does not count as meaningful human involvement”, and a person who only inputs data into a system that then decides is still within the rules if the effect is significant. The draft adds: “You should keep a record of how the human was involved in the decision.”

Its worked example is reassuring on the point many businesses worry about. An employer’s automated clocking-in system flags that an employee has been late a set number of times, and an HR manager reviews the data before deciding to issue a warning. The draft says that is meaningful human involvement.

Which of an AI employee’s tasks are decisions that matter?

The draft gives a spread of examples that help sort an AI employee’s work. On the significant side: an automatic refusal of an online credit application, e-recruiting without meaningful human involvement, and a bank account freeze on suspected fraud. And, close to home for some of our sectors, a lettings platform that automatically decides whether someone is eligible to let high-value properties, using their letting history, landlord ratings, payment behaviour and behavioural signals. On the less likely side: a video-on-demand service recommending what to watch next.

Applied to an AI employee, that suggests a rough sorting, which is ours and not the ICO’s:

The draft is cautious about the line. It says context is key, that a decision that has little impact for most people can have a significant effect for some, such as children, and that “Unless you are confident that you can accurately separate out the people who will experience legal or similarly significant effects from those who will not, you should apply the safeguards to all the decisions you make.” The simplest design for most businesses is the one the draft implies: the AI prepares, and a person decides the ones that matter.

The four safeguards, in practice

Where a decision is solely automated and significant, the safeguards are not optional extras. The draft says: “You must apply the safeguards consistently rather than on an ad hoc discretionary basis.” In practice each one is a piece of working process.

All four depend on one thing: being able to find the decision again, with the data and the reasoning behind it. An AI employee whose actions are not logged cannot support any of them.

A plain gold stamp beside a clear glass card on a green surface, standing for a record that a person reviewed a decision.
A record that a person looked, and what they decided.

What software can do, and what it cannot

The chart below uses illustrative figures, not client data, for a business whose AI handles about 10,000 customer interactions a month.

Most of the work is finding decisions and evidencing reviewIllustrative hours a month for a business whose AI handles about 10,000 customer interactions a month, by hand and with a system
By handWith a system
Finding which decisions the AI made, and on what data14 hrs3 hrs
Recording who reviewed each decision and what they changed12 hrs2 hrs
Preparing a decision-specific explanation when someone asks8 hrs3 hrs
Tracking representations, human review requests and challenges6 hrs2 hrs
Deciding whether a decision is significant and how to treat it6 hrs5 hrs

Illustrative inputs, not client data. Totals: 46 hours a month by hand against 15 with a system. Deciding whether a decision is significant falls least, on purpose: that is a judgement for the business, a data protection adviser and, in the end, the regulator or a court. A system removes the finding, recording and tracking, not the judgement. Substitute hours you have measured.

Show the data
By handWith a system
Finding which decisions the AI made, and on what data14 hrs3 hrs
Recording who reviewed each decision and what they changed12 hrs2 hrs
Preparing a decision-specific explanation when someone asks8 hrs3 hrs
Tracking representations, human review requests and challenges6 hrs2 hrs
Deciding whether a decision is significant and how to treat it6 hrs5 hrs

The shape matters more than the figures: finding, recording and tracking shrink, while the decision about significance does not. Our AI employees page describes hard rules for what they may decide and what they must escalate, every action logged, attributed and reversible, and handing work to your people with the full context. Our AI calling agents page describes a warm handover to a person the moment a call needs one, with every call transcribed, recorded and filed against the customer. Used well, that kind of design can:

What it cannot do is decide whether a decision is significant, whether a human’s involvement is meaningful in your circumstances, or whether your processing is lawful. A system also cannot make a reviewer trained or give them authority. Those are for the business, its data protection adviser and, in the end, the ICO and the courts. We do not claim that using any product makes a business compliant. For the wider picture, see the UK rules on AI for businesses, what AI employees are and how to implement AI in a UK business.

A loop of light leaving a glass sphere, passing through a gold ring and returning to it, standing for a challenged decision going back to a person who can change it.
A challenge goes back to a person who can change the outcome.

Where to start this week

  1. List what your automated systems and AI actually decide about people, not what they say or schedule.
  2. For each, ask the three questions: is it a decision, is it significant, is it solely automated?
  3. Where the answer is yes to all three, either put a trained person with authority into the decision before it takes effect, or build the four safeguards.
  4. Check the lawful basis and the data. Does any of it rely on recognised legitimate interests, or involve special category data?
  5. Check you could show the human involvement: pick one decision and find who looked at it, when, and what they could have changed.

This is general information based on section 80 of the Data (Use and Access) Act 2025 and the ICO’s draft guidance (updated 31 March 2026, under consultation) as we read them on 15 November 2026. We have not reviewed any regulations made under Article 22D, the ICO’s final guidance, which was not published when we read it, or its work on a code of practice for AI. The draft may change. It is not legal advice.

Frequently asked questions

What is an automated significant decision under UK data protection law?

Since 5 February 2026 the UK GDPR has treated a decision as based solely on automated processing if there is no meaningful human involvement in taking it, and as a significant decision if it produces a legal effect for the person or a similarly significant effect. These definitions are in Article 22A, inserted by section 80 of the Data (Use and Access) Act 2025. The ICO's draft guidance says the rules apply only where a system is making a decision about a person, the decision is significant, and it is solely automated.

What safeguards must be in place?

Article 22C says that where a significant decision is based solely on automated processing and on personal data, the controller must ensure safeguards that provide the person with information about the decision, enable them to make representations, enable them to obtain human intervention, and enable them to contest the decision. The ICO's draft guidance says the information must be specific to the decision, not just a repeat of the privacy notice, and that the safeguards must be applied consistently rather than on an ad hoc basis.

What counts as meaningful human involvement?

The ICO's draft guidance, which is still under consultation, says involvement must be active and not just a token gesture. It lists that a human should assess and review the decision at an appropriate point so it has actual impact, have the ability to influence the outcome, have discretion and authority to alter the decision, be suitably trained to understand the system's logic, outputs, limitations and risks, and take into account the relevant data and factors. It says occasional spot-checking is not enough and that you should keep a record of how the human was involved.

Can an AI system use legitimate interests as its lawful basis for a significant decision?

The ICO says the reforms open up a wider range of lawful bases for significant automated decisions, potentially including legitimate interests, with safeguards. But Article 22B(4) says a significant decision may not be taken solely by automated means where the processing relies, wholly or partly, on the new recognised legitimate interests basis. Special category data remains more tightly restricted, and such decisions can only be taken solely by automated means on limited conditions such as explicit consent.

Does a business need to use AI for these rules to apply?

No. The ICO's draft guidance says automated decision-making systems do not need to involve complex algorithms, artificial intelligence or other advanced processing to potentially come into scope. A decision needs to involve some evaluation or analysis of personal information, not just the application of a rule that a human has already set. So a simple scoring script can be in scope while an AI that only follows a human's fixed rule may not be.