Part 1, Understanding AI agents
Part 3, Introducing an agent
Principle 6
Learning
Scope and limitations of this guide
This guide provides general information on organising work with AI agents, not advice tailored to a particular organisation, system or decision. It is not legal or regulatory advice.
Its examples, checklists and templates do not establish that a proposed use is safe, lawful or suitable. You need to assess and test the proposed use in your own circumstances. Following the guide does not by itself meet the requirements of any law or standard.
The guide is provided in good faith, using information available at the time of publication or revision. AI systems, provider terms and legal requirements can change. Before acting, check the current requirements that apply to your organisation and country, confirm what the system can do and test the proposed use. You are responsible for how you apply the guide and should obtain professional advice where appropriate.
Reading, downloading or using the guide does not by itself create an adviser–client relationship with Tom Clowry or Considered Ideas Ltd.
To the fullest extent permitted by law, Tom Clowry and Considered Ideas Ltd accept no liability for loss or damage arising from use of, or reliance on, this guide. Nothing in this notice excludes or limits liability for death or personal injury caused by negligence, fraud or fraudulent misrepresentation, or any other liability that cannot lawfully be excluded or limited.
How will you and your colleagues develop and maintain the expertise to judge the agent’s work?
When AI performs the work through which expertise has traditionally developed, how do people learn to judge the result?
You and your colleagues need knowledge and experience to assess an agent’s work. Plan how you will acquire and maintain both as the division of work changes.
- What to decide
- How you and your colleagues will develop and maintain the expertise to judge the agent’s work.
- Key point
- People need practice to retain the expertise that oversight depends on.
Why it matters#
The expertise paradox#
AI can replace the expertise needed to perform a task while leaving us dependent on that expertise to make good judgements about the result.
People have traditionally developed that judgement by doing the work AI is now performing for them.
People often learn through the work itself. If an agent takes over that work, the opportunity to learn may change too. This matters at every career stage.
Background#
From psychology
The ironies of automation#
In 1983 the psychologist Lisanne Bainbridge described the ironies of automation. When a system takes over routine work, people are left to monitor it and to take over when it fails. Monitoring gives them little practice, so the skills they need in a difficulty fade.
This resembles the expertise paradox. Bainbridge was writing about industrial process control, long before AI agents. Principle 6 asks you to plan practice, feedback and difficult cases so that you and your colleagues keep the judgement that oversight depends on.
Source Bainbridge, L. (1983), Ironies of automation. Automatica, 19(6), 775–779.
In practice#
6 steps
01Agree with each colleague what they needDifferent stages of a career need different opportunities.
Someone beginning a role and an experienced colleague can both be asked to assess an agent’s work, but they need different opportunities to maintain and extend their knowledge.
Someone beginning a roleNeeds experience of ordinary cases.An experienced colleagueNeeds to understand changes in the work and encounter unfamiliar situations.02Allow time to form a view firstChoose the sequence according to what each colleague needs to learn.
Relevant expertise helps a colleague assess an agent’s draft, but it does not remove the risk of accepting its framing. Forming an initial view before reading the draft can be useful for experienced colleagues as well as those learning the work. Discussing where and why the answers differ can help them understand the subject more fully.
Reviewing an agent’s draftAgent’s draft→Review and refine→DecideForming an independent view firstForm a view first→Compare with agent work→Discuss differencesTwo approaches to reviewing and learning from an agent’s work Choose the sequence according to what each colleague needs to learn.
03Practise on ordinary and difficult casesInclude examples in which an agent’s work is incomplete or wrong.
You and your colleagues should practise on ordinary and difficult cases, compare different approaches and examine examples in which an agent’s work is incomplete or wrong, with feedback from someone who can judge the work.
When a colleague is learning the work, ask them to explain:
Why they preferred one interpretation
What evidence supported it
What would make them reconsider
04Use agents to support learningCheck the material where the learner cannot yet judge it.
Agents can also support development. They can help someone explore a different approach, prepare practice exercises or rehearse how to explain a difficult decision. Where the learner cannot yet judge the material, an experienced colleague should check it and help interpret the results.
05Consider what depends on a human relationshipAgent support can help a colleague prepare and follow up.
Some parts of the work depend on a human relationship. A service user may need someone to listen, explain a decision or take responsibility for resolving a difficulty. Agent support can help the colleague advising them prepare and follow up. Consider what the service user needs from the person advising them before changing who provides that advice.
06Make development part of the working dayAllow time for practice and feedback.
Agree time for practice and feedback within the working day.
Discuss changes with colleagues
Explain to colleagues how their work will change and how they will learn the parts that are new to them.
Make clear where help is available
Make clear where colleagues can obtain help and allow time for questions.
Examples#
The opening hours example
Learning to lead future reviews#
A local service is open from 9 am to 5 pm on weekdays. Some service users have asked for evening appointments, and the team must decide whether longer hours would improve access. An agent examines twelve months of request records and compares the costs of the options. See chapter 1.5.
Interprets a small selection of request records on their own. They are preparing to lead future service reviews.
Compares the colleague’s account with the agent’s and discusses the differences with them.
Also speaks with service users about their experience of arranging and attending appointments.
Connecting what the records say with the experience of the people using the service can show why a recommendation may need to change.
In other work
Risks#
From the research
AI assistance can affect learning during a task#
In a randomised study reported in a 2026 preprint, 52 developers learned an unfamiliar programming library with or without an AI assistant. The AI-assisted group achieved a lower average score on a subsequent unaided quiz, with no statistically significant difference in completion time.
Within the AI-assisted group, those who asked conceptual questions or sought explanations tended to score higher than those who delegated the coding. These patterns of use were observed during the study; participants were not assigned to them. This was a short study of learning during a task, rather than a test of long-term expertise.
Source Shen, J. H. and Tamkin, A. (2026), How AI impacts skill formation, arXiv:2601.20245 (preprint).
From the research
Practice with AI can affect later performance#
Another 2026 preprint reported three randomised experiments that recruited 1,222 participants, with 1,058 included in the reported analyses after exclusions.
Following brief AI-assisted practice on mathematical or reading tasks, participants performed less well without assistance than those who had practised without AI. They also skipped subsequent problems at a significantly higher rate in two of the three experiments.
Source Liu, G. et al. (2026), AI assistance reduces persistence and hurts independent performance, arXiv:2604.04721 (preprint).
From the research
AI training had reached most staff in relatively few businesses#
In a survey published in July 2026, 11% of UK businesses with 10 or more employees reported that more than half of their workforce had received AI-related training.
Source Office for National Statistics, Artificial intelligence in UK businesses: 2023 to 2026, 20 July 2026. Self-reported; official statistics in development. Some industries excluded.
Exercise#
Plan how you and your colleagues will develop your judgement#
Identify one judgement that you or a colleague will still need to make after delegation. What experience currently helps develop it? Decide how that experience will continue, or what you will put in its place.
Agree how you will recognise progress. For example, can someone explain an alternative, identify an unsupported conclusion and say when they need further help?
Describe the expertise you and your colleagues need, how you will practise and who will provide feedback. Review the division of work as you gain experience and responsibilities change.
Further reading · PACED and the expertise paradox · clowryai.com/thinking
© 2026 Considered Ideas Ltd. Licensed under the Creative Commons Attribution 4.0 International licence (CC BY 4.0). This guide provides general information and is not legal or regulatory advice. The guide’s home page gives the suggested credit.
Scope and limitations of this guide
This guide provides general information on organising work with AI agents, not advice tailored to a particular organisation, system or decision. It is not legal or regulatory advice.
Its examples, checklists and templates do not establish that a proposed use is safe, lawful or suitable. You need to assess and test the proposed use in your own circumstances. Following the guide does not by itself meet the requirements of any law or standard.
The guide is provided in good faith, using information available at the time of publication or revision. AI systems, provider terms and legal requirements can change. Before acting, check the current requirements that apply to your organisation and country, confirm what the system can do and test the proposed use. You are responsible for how you apply the guide and should obtain professional advice where appropriate.
Reading, downloading or using the guide does not by itself create an adviser–client relationship with Tom Clowry or Considered Ideas Ltd.
To the fullest extent permitted by law, Tom Clowry and Considered Ideas Ltd accept no liability for loss or damage arising from use of, or reliance on, this guide. Nothing in this notice excludes or limits liability for death or personal injury caused by negligence, fraud or fraudulent misrepresentation, or any other liability that cannot lawfully be excluded or limited.

