Part 1, Understanding AI agents
Part 3, Introducing an agent
Principle 1
Purpose
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.
What should improve for the people who use or do this work?
If an agent completed this task in minutes, what would you change about the rest of the work?
Decide what the work should achieve and for whom. Then consider how delegation could improve the way it is done.
- What to decide
- The improvement you want, who benefits and how you will recognise it.
- Key point
- Judge the improvement across the whole piece of work.
Why it matters#
Has the work as a whole improved?#
A task that finishes sooner has not necessarily improved the work.
So begin with the people who use the work. What should be better for them, and how will you know?
Background#
From economics
When a measure becomes a target#
A measure is a good guide to the work until people start aiming at the measure itself. They then find ways to improve the figure without improving the work. The economist Charles Goodhart observed this in economic policy. It is usually summed up in the words of the anthropologist Marilyn Strathern. When a measure becomes a target, it ceases to be a good measure.
An agent is affected in the same way. If it is asked to improve a measure, it can do so in ways that make the work worse.
Principle 1 therefore asks you to define the improvement for the people who use the work, and to check for side effects that the measure does not show.
Sources Goodhart, C. A. E. (1976), Problems of monetary management: the UK experience. Strathern, M. (1997), ‘Improving ratings’: audit in the British University system.
Where to start#
Choosing where to start#
A suitable first task is one you know well, in which errors are easy to detect and to put right.
Frequent
It is done often enough to justify the effort of setting it up.
Familiar
You know the work well enough to judge the agent’s results.
Easy to check
Its results can be checked quickly and reliably.
Reversible
Any change can be undone, and a mistake would have limited consequences.
A task with these qualities makes it easier to judge whether the agent is helping, and limits the consequences if it is not.
When an agent is not the answer#
Delegation is not always worthwhile. Consider another approach in the following cases.
The task is quicker to do yourself
Once briefing and checking are counted.
Nobody available can judge the result
Errors would then go unnoticed.
A mistake could not be put right
And limits or approvals could not prevent it.
The person needs a human relationship
Someone may need a person to listen, explain a decision or take responsibility for resolving a difficulty.
The information cannot be shared
The rules that apply to it do not allow it to be supplied to the AI service.
The process itself is the problem
A simpler change, such as clearer standard answers, may achieve more.
Understanding the existing work#
Before proposing changes, establish how the work is done and why people do it that way. This account gives you a basis for applying the six principles and judging whether a different arrangement would be better.
Follow the work from beginning to end
Choose a recent example and follow it from the initial request to its eventual use. Ask the people involved what they did, what they needed from others and where they had to wait. Include those who use the result. Their experience can differ from the account in a procedure or a completed record.
Distinguish the task from what it enables
A budget analysis informs a spending decision. Advice helps someone understand their options. Record how the result is used, including any explanation or follow-up that people need before they can act on it.
Identify the judgement and informal checks involved
Note where someone had to interpret an unclear request, resolve a disagreement or ask for advice before they could continue. Note too what helped them. Include the checks and informal enquiries that prevent problems, because these can leave little trace in the record.
Who to involve#
Introducing an agent changes how work is done. It needs the involvement of those who authorise and oversee the work, and of those who do it or rely on its results.
Involve them from the start and keep doing so. Colleagues who use or oversee an agent need to know what it can do, what information it can reach and what is expected of them. Agree how they can report problems and raise concerns, and review their experience alongside the results.
Sources UK Government, AI Playbook for the UK Government, 10 February 2025. National Audit Office, Good practice guide for organisations using AI, May 2026. IMDA, Model AI Governance Framework for Agentic AI, version 1.5, 2026. NIST, AI Risk Management Framework (AI RMF 1.0), 2023. Acas, 28 April 2025.
In practice#
5 steps
01Define the improvementBe specific about what you want to change.
Be specific about what you want to change. Fewer errors, shorter waits and more time for difficult cases each require different arrangements. State who should benefit, and how you will recognise the improvement.
02Check for side effectsCheck for adverse effects as well as improvements.
Check for adverse effects as well as improvements. Would the change create work or inconvenience elsewhere, including in related processes?
Writing software changes
Can produce changes more quickly than their outcomes can be properly assessed.
Updating records
A mistake can spread to every system and report that uses those records before anyone notices it.
03Ask whether the work should be organised differentlyAgents can change the way work is organised as well as the time it takes.
Revising the analysis before deciding
If an agent can produce an analysis quickly, the team can discuss early results, explore alternatives and request revisions before deciding.
Work that was not practical before
Agents can also make it practical to examine more cases or adapt explanations for different audiences. Allow time for colleagues to use and discuss the results.
Existing sequenceRequest→Compile→Finished report→Discuss at the endProposed sequenceAgree the question→Explore options→Discuss evidence→Revise and decideTwo ways of preparing an analysis 04Decide whether delegation is worthwhileConsider the standard, the time and the review it will require.
Three questions help here. Is the agent likely to meet the standard you need? Will the result arrive in time? How long will it take to review?
Likely to succeedTime the agent takesTime to reviewOutcomeOften falls shortQuickQuickOf little use, as the work is redoneReliableToo slow for the deadlineQuickArrives too late to helpReliableQuickAs long as the work itselfThe time saved is spent checkingReliableIn timeManageableA candidate for delegationThree practical considerations when assessing delegation. Drawing on Ethan Mollick (2026). How long the review takes depends on who carries it out. Understanding the subject helps a reviewer judge whether the result is accurate, relevant and appropriate.
Someone with relevant expertiseCan assess the result more quickly.Is more likely to recognise its weaknesses.Someone with less experienceIs likely to need longer.Is less likely to recognise an error.Estimate the review time for the person who will actually do it.
Then compare the cost of using the agent with the cost of doing the work without it. Include the time spent briefing the agent, reviewing and correcting its work, and any attempts that fail.
Even where the agent would save time or money, you also need to weigh what the improvement is worth and what a mistake would cost.
05Decide with those involved how the work should changeAgree what continues, what changes and what stops.
With the people doing the work, go through what it involves and agree which of these each part falls into.
NecessaryThe records and explanations that others need to understand decisions and rely on the work.
Could changeActivities that could take place earlier or be done differently.
No longer neededActivities that no longer serve a useful purpose and are not required for another reason.
New responsibilitiesPreparing instructions, checking the work and resolving questions referred to colleagues.
If delegation releases time, agree with the people doing the work how it will be used.
See also · 1.1 Model capabilities and trade-offs
Examples#
The opening hours example
Is the analysis worth delegating?#
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.
Test whether the agent can organise the records and calculate costs accurately. The records cover only requests made during opening hours, so they cannot show how many people would use later appointments. The team will need further evidence.
Check that the analysis and its review can be completed before the decision is due.
The service lead checks that relevant records have been included, the calculations are correct and the assumptions are appropriate. The team allows time for these checks.
The team tests the agent on known records before delegating the analysis. Staff and service users help establish what further evidence is needed about demand.
The opening hours example
Comparing the options while they are still open#
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.
Prepares an initial comparison of the three options before the team chooses a preferred approach.
Ask who would cover closing duties after a late session.
Ask whether they could book by phone after 5 pm.
Updates the comparison as these questions are answered.
Preparing the analysis more quickly allows staff and service users to influence the recommendation while it can still change.
Risks#
When objectives conflict#
An objective also reflects assumptions about what is desirable and fair, and two authorised objectives, such as reducing cost and improving service, may conflict. Agree who will weigh them, and consider whose interests are affected.
Who will weigh them?
Whose interests are affected?
Exercise#
Examine a proposed improvement#
Choose an activity that an agent could perform more quickly. Explain what that would allow someone else to do better. If the next stage would remain unchanged, consider whether it also needs to be reorganised.
Identify an activity that could take place earlier, be done differently or be discontinued. Check the implications with the people who do it and those who depend on it.
Describe the improvement you want, who should benefit and how you will judge the result. Explain the proposed changes to the work and any additional responsibilities they create.
Further reading · AI agents and the rebundling of knowledge work · 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.

