The Next AI Skill for Managers Is Delegation, Not Prompting

A colleague opens a tool, pastes a messy brief and gets a tidy draft in seconds. Someone else asks the same tool to “chase the actions”. A third person wonders whether the status pack can just update itself.

Prompting still matters. Clear instructions get better drafts. But that is no longer the only leadership skill that counts.

As AI starts doing pieces of work rather than only helping with them, the useful questions change:

• What work should be delegated?
• What authority should AI have?
• Where must human judgement remain?
• How will you know the work was done properly?

Those are delegation questions. They are older than any model release. They are becoming newly important because the boundary between “help me prepare this” and “handle this” is getting thinner.

Delegate the work. Set the boundaries. Keep accountability clear.

That is the practical centre of this piece.

Why this is arriving now

AI use is rising. That does not mean most organisations are already full of autonomous agents.

In July 2026 the UK Office for National Statistics reported that self-reported AI use among UK businesses with 10 or more employees had risen from around 12% in late 2023 to around 35%. It also said adoption has been relatively shallow so far: the average number of AI technologies used by adopting businesses rose only modestly, from about 1.4 to about 1.6. Improving business operations is the most common use among larger firms, and that has not yet translated into widespread headcount change.

UK Government AI adoption research paints a similar picture inside the firms that already use AI. On average about 30% of staff in adopting businesses currently use AI. Agentic AI was the least adopted technology in that research, at 7%, which fits a market that is still early on systems that can carry out several steps with less human input. Among users, three quarters reported improved workforce productivity and over half reported improved processes. Skills and identified need remain major barriers.

So the honest reading is not “the agentic workplace is already here”. It is “enough people are already using AI on real work that managers need a clearer way to delegate it”. The ONS has even flagged future measurement of agentic workflows. That is a signal of direction, not proof that every team already has them.

This is exactly why leaders should learn the skill early: while the stakes of a weak trial are still manageable, and before vague permission becomes informal policy.

Prompting is useful. Delegation is the missing layer.

Prompting answers: How do I get a better output from this tool?

Delegation answers: Should this work be given to AI at all, under what rules and what a person still owns?

A strong prompt can still leave you with a weak management decision. You may get a polished risk note that nobody was authorised to send. You may get a confident summary that skips the one fact a steering group will ask about. You may get a draft action list that quietly changes the decision you still own.

Managers already know this pattern with people. You do not delegate a sensitive stakeholder update by saying “just sort it”. You name the outcome, the limits, the points that need your call and the evidence you want back.

AI needs the same discipline, without treating the tool like a junior employee with feelings or a strategy partner with equal standing. It is work support with growing capacity. Your job is to decide what capacity is appropriate for this task.

A simple spectrum of AI involvement

Not every task needs the same level of machine involvement. A useful spectrum looks like this:

1. Assist
   AI helps you think or draft. You stay in the loop on every sentence. Example: turn rough notes into a clearer outline.

2. Prepare
   AI assembles options, tables or a first full draft from material you trust. A human still decides what goes forward. Example: draft three options for a trade-off discussion.

3. Approval-gated action
   AI can propose or stage an action, but nothing leaves the building until a named person approves. Example: draft chaser emails that only send after you click.

4. Bounded autonomy
   AI may complete a defined task inside fixed limits, with logging and a stop condition. Example: update a non-sensitive tracker from an approved source and flag anything outside the pattern.

More autonomy is not automatically better. The right level depends on:

• consequences if it is wrong
• how reversible the action is
• how clear the success standard is
• how much context and judgement the work needs
• whether the information is safe to use in the chosen tool

If the work is high-consequence, hard to reverse or short on trusted inputs, stay nearer assist or prepare. If it is low-risk, repetitive and easy to check, bounded autonomy may be worth testing. The spectrum is a judgement aid, not a race toward more autonomy for its own sake.

The six-part model for AI delegation

Use this when someone says “can AI just do this?” or when you are tempted to hand something over because you are short of time.

1. Outcome

What exact result do you want by when?

Write it so a stranger could recognise whether it happened.

Weak: “Help with the steering pack.”
Stronger: “A two-page options note with three choices, the delivery risk of each and one recommended ask for Thursday’s meeting.”

If you cannot name the outcome, you are not ready to delegate. You are still clarifying the work. That is fine. Do that first.

2. Authority

What is AI allowed to do, and what is it forbidden to do?

Authority is more than tool access. It includes:

• which sources it may use
• whether it may contact people
• whether it may change shared files
• whether it may send anything externally
• which labels, numbers or commitments it must not invent

A useful test: if this work were done by a capable contractor for one afternoon, what would you put in the email that scopes their brief? Put the same limits on the AI task.

3. Judgement

Where must a human still decide?

Name the judgement points before the work starts. Typical ones:

• which option to recommend
• what risk language is honest enough for this audience
• whether a stakeholder concern is political, technical or both
• whether something is good enough to send
• when to stop and escalate rather than push on

AI can structure choices. It should not quietly make the call you will be asked to defend.

4. Evidence

How will you know the work was done properly?

Decide the evidence standard before you see the output. That might be:

• source links or citations you can open
• a short change log of what was edited
• a checklist of constraints still met
• a human review of the two highest-risk claims
• a comparison against the original numbers you trust

Without an evidence standard, speed becomes a substitute for quality. You only discover the gap when someone senior asks a basic question.

If you are using AI inside a trial or pilot, be equally clear about what the trial can and cannot prove. A tidy demo is not the same as proof that the work is safe to scale. For that boundary, see What Should a Pilot Actually Prove Before You Run It?.

5. Escalation

When must the AI stop and bring the work back?

Define stop conditions in plain language:

• missing data
• conflicting sources
• anything that would commit budget, date or scope
• personal or commercially sensitive content
• a confidence drop on a material claim
• any action outside the approved channel

Escalation is not failure. It is how you keep autonomy bounded. A system that never stops is not efficient. It is unsupervised.

6. Accountability

Who owns the result when it leaves the room?

This is the line managers sometimes blur under time pressure.

You can delegate drafting, assembly, checking and even some routine action. You cannot delegate accountability for a decision you still own, a message sent in your name or a risk accepted on the programme’s behalf.

Write the owner in one sentence: “I own the recommendation. AI prepared the options pack under the limits above. I checked the evidence and approved the final version.”

If nobody can say that sentence, the work is not ready to go.

Putting the model to work on real management tasks

Here is how the same six parts look on common delivery work.

Decision paper
Outcome: one-page recommendation with options and ask.
Authority: use only the redacted notes provided. No email. No new commitments.
Judgement: recommended option stays with you.
Evidence: every figure traceable to the source pack.
Escalation: stop if a material number is missing.
Accountability: you present and own the ask.

Status update
Outcome: a short narrative of change since last time, risks and decisions needed.
Authority: pull from the approved tracker only. Do not invent progress.
Judgement: risk wording and what to escalate stay human.
Evidence: each status line maps to a tracker field or owner note.
Escalation: stop on contradictory owner updates.
Accountability: the workstream lead signs off before circulation.

Stakeholder chaser
Outcome: draft follow-ups ready for review.
Authority: draft only. No send without approval.
Judgement: tone and timing stay human.
Evidence: each chase names the open action and due date from the log.
Escalation: stop if the action owner has already replied in another thread.
Accountability: you send, or you explicitly approve the send.

Notice the pattern. The tool may do more of the assembly over time. The management questions stay stable.

How this fits current risk thinking without turning you into a governance function

You do not need a new department to use this model.

Public risk guidance is moving in a compatible direction. NIST’s AI Risk Management Framework is voluntary guidance organised around Govern, Map, Measure and Manage. It is being extended with more operational profiles, including work in 2026 on higher-stakes settings. Separate research building on that framework, including the UC Berkeley Center for Long-Term Cybersecurity’s February 2026 Agentic AI Risk-Management Standards Profile, focuses on systems that can pursue goals and take actions with little human oversight. The practical warning is simple: once software can act, not only generate text, organisations need clearer limits, monitoring and accountability.

For a delivery manager that translates to ordinary habits:

• know what the tool is allowed to touch
• know what “good” looks like before you accept the output
• keep a human on the judgement points that carry consequence
• make sure someone named still owns the result

For a manager, these are ordinary management disciplines applied to a tool that can now do more.

A 20-minute way to practise this week

Pick one real task already on your desk. Prefer something recurring, low to medium sensitivity and owned by you.

1. Write the outcome in one sentence.
2. Write four authority limits: sources, actions, people contact and send rights.
3. Circle the judgement points only a human should make.
4. Name the evidence you will check in five minutes or less.
5. Write two escalation stops.
6. Write the accountability sentence in your own name.
7. Choose the lowest safe point on the spectrum that still saves time.
8. Run the task. Keep what worked. Tighten any limit that felt vague.

If you cannot complete steps 1 to 6, do not hand the task to AI yet. Clarify the work first. For programme and transformation leaders choosing a sensible first piece of work, How to Choose AI Support for Programme Leadership Work walks through that selection problem in more detail. If you want a short browser recommendation shaped to your situation, use free Where AI Could Help.

What good looks like

You are getting this right when:

• people on the team can explain what AI is allowed to do on a task without guessing
• drafts arrive faster, but risky claims get checked before they travel
• autonomy expands only where consequences are understood and reversible enough
• nobody is surprised by a message, number or commitment that “the tool sent”
• you can still answer “who owns this?” in one breath

You are drifting when:

• speed is treated as proof of quality
• prompts get longer while ownership gets vaguer
• AI is used to avoid a decision that still needs a human
• the first time limits are discussed is after something awkward has already gone out

Closing

As AI does more of the work, good prompting will not be enough. Managers will also need to delegate clearly without losing accountability.

Prompting helps you ask well.
Delegation decides what may be asked for, under what authority and with what evidence.

Start small. Keep the six parts visible. Move along the spectrum only when the work, the consequences and the checks justify it.

Delegate the work. Set the boundaries. Keep accountability clear.