What we are reading on AI agents in HR: ten pieces worth your time in 2026

We read a great deal about AI agents so that we build the right thing. Most of it is not worth passing on: a list of vendor features under a headline about transformation. A few pieces are different. They define their terms, say what a number measured and give the limits a section of their own.

These are the ten we would hand to an owner or an HR lead at a UK employer, with a note on what each is good for. We are not affiliated with any of the authors, and they do not all agree with us or with each other.

Start here: what an agent is

1. Simon Willison on what “agent” means

Simon Willison collected a long list of competing definitions and settled on one line: an agent runs tools in a loop to achieve a goal. He is also blunt that the “agent as a member of staff” idea falls down on accountability, because you cannot hold software responsible for anything. Read it before any vendor demo.

2. Anthropic, “Building effective agents”

Written for engineers, readable by anyone. It separates workflows, where the steps are fixed in advance, from agents, where the software chooses the steps, and argues for the simplest thing that works. Much of what is sold as an HR agent is a workflow, and for a pay run that is often the right choice.

The UK regulator’s view

3. The ICO on automated decisions in hiring

The piece every employer using AI in recruitment should read. After talking to employers, the ICO found that many were making solely automated decisions without realising it. Its expectations are short: monitor for bias, be open with candidates, and explain how to get a human review.

4. DLA Piper’s reading of that report

A clear legal summary with the point that matters most to anyone relying on an approval step: human involvement has to be real and active, not a signature on whatever the software produced.

5. The ICO’s Tech Futures report on agentic AI

Not formal guidance, but the regulator’s early thinking. The message is that giving software more initiative does not move responsibility. The organisation that deploys the agent answers for what it does with personal data.

What it means for HR people

6. Personnel Today, “What is agentic AI and should HR be using it?”

Jo Faragher’s explainer is the best UK trade piece we have found. It sets out who is selling agents and what they claim, then spends as long on the practical, ethical and legal problems. It ends on cautious adoption with human oversight, which is where we ended up too.

7. Josh Bersin, “The great reinvention of human resources has begun”

The boldest version of the argument: agents take the administrative half of HR and leave people who work on talent, organisation and running the agents. It is written with large enterprises in mind and is more confident about shrinking HR teams than we are. A business with one HR person reads it differently: the admin goes, and the person stays.

8. CIPD, “How AI is changing the application of HR skills”

The CIPD’s research argues that the limit on getting value from AI is not the technology but ordinary HR capability, such as job design and workforce planning. Organisations that are good at those report much better results from AI. Useful if you are being told the tool will fix everything.

Where the person sits

9. Deloitte, “AI and the future of human decision-making”

From the 2026 Human Capital Trends. It opens with two failures, a biased CV screener and a service bot making promises nobody authorised, and argues that how much freedom an AI gets should depend on what is at stake. Its distinction between a person being in the loop and merely on it is worth borrowing.

10. Ethan Mollick, “Agency and agents”

Mollick looks ahead to agents that do most of the work and bring people in at the right moments. He names four: when approval is needed, when expertise is needed, when something unusual happens, and when the person would simply want to know. On approval he is firm that an agent should not decide for itself to spend money, contact outsiders or open sensitive material.

The numbers we trust

Three figures from sources that say what they measured.

• About 35% of UK businesses with ten or more employees were using at least one AI technology in June 2026, according to the Office for National Statistics.
• 54% of firms told the British Chambers of Commerce they were actively using AI, in a survey published in March 2026, most of them small and mid-sized, with limited effect on headcount so far.
• 84% of AI-using businesses have a person check what the AI produces, and only 2% do no checking at all, in research for the Department for Science, Innovation and Technology. The fieldwork is from 2025, so adoption has moved since. The instinct to check has not.

What we took from all of it

• Define the job. An agent with a clear goal and a clear test of success does well. “Help with HR” is neither.
• The approval has to be real. The approver needs to see what changed and why, and be able to say no.
• Responsibility does not move. Whoever deploys the agent answers for it.
• Almost nobody is writing for small employers. The examples are companies with thousands of staff. That is the gap we are trying to fill here, starting with what AI agents do in HR in 2026 and how to hire one.

If you have read something better, send it to us.