Managing in the AI era: what to delegate — and what you still need to decide

Leading a team that uses AI raises an unusual question: part of the work is now produced by a system that is fast, often convincing, sometimes wrong, and accountable to no one. The reflex is either to ban it or to let it run. Both amount to not managing.

This guide is for managers and business leads who use AI, or who lead people who do, without being technical specialists. It rests on one simple and demanding distinction: AI can assist with or execute a task, but it never takes over responsibility for the decision.

Delegating a task is not delegating a decision

This is the line everything else hangs on. A task produces a result; a decision commits the organisation, a client, a person or a budget. A system can produce; it cannot commit.

What can be assisted, what can be automated

Not every task deserves the same treatment. Three questions place a task: what does an error cost, is it detectable, and can the result be checked in less time than it takes to produce it?

1. 1

   Useful assistance

   low risk

   Drafts, rewrites, meeting notes, preparing a deck, exploring angles before choosing. Errors are visible and carry no lasting consequence, because the output will be reworked anyway.

2. 2

   Assistance under review

   moderate risk

   Customer replies, numerical analysis, minutes that get circulated, translations sent outside. The output leaves the team: it needs real review by someone who knows the subject.

3. 3

   Framed automation

   repetitive and checkable

   High-volume tasks with stable criteria and results verifiable by sampling: triaging incoming requests, extracting fields, routing. Automation implies quality monitoring over time, not only at launch.

4. 4

   Out of scope

   human commitment

   Appraising a person, disciplinary decisions, sensitive arbitration, crisis communication, subjects where the relationship matters as much as the content. AI can prepare material; the decision and the words stay human.

Review proportionate to risk

The mirror trap of letting go is uniform control: reviewing everything the same way is expensive and drains attention exactly where it would matter. Control should be calibrated.

• The heavier the consequence of an error, the more explicit and traceable the check should be.
• The harder an error is to spot afterwards, the more it must be caught before, not after.
• When volume is high and risk is low, sampling beats a full review done badly.
• What leaves the organisation deserves a higher level of review than what stays internal.
• Rare or atypical cases deserve more attention than the nominal case, even when the system looks strong.

Plausible errors and calibrated trust

The specific difficulty with these tools is the shape of their mistakes: well written, coherent and confident. A team under time pressure approves them more readily than a clumsy human draft, precisely because nothing raises a flag.

Managing that means building calibrated trust: neither blanket suspicion, which cancels the gain, nor default acceptance, which pushes the risk onto the customer. In practice, name with the team the situations where the tool is weak — recent information, figures, references, implicit internal context, out-of-scope cases — and make it normal to say that a deliverable was produced with AI.

A simple example: a briefing note that reads perfectly but attributes to a supplier a clause that is not in their contract. Nothing in the form signals the error; only knowledge of the file reveals it.

Data, confidentiality and the rules of use

A manager does not need to be a lawyer, but is often the last person able to stop sensitive information going into an unapproved tool. Three markers are enough day to day.

• Know which tools are approved in the organisation, and for which kinds of information.
• Tell apart public, internal, confidential and personal data before pasting anything into a tool.
• Raise the question openly in a team meeting instead of letting everyone invent their own rule.

Assessing work produced with AI

The temptation is to assess effort; what still matters is the result and the judgement exercised. Fast and accurate work, whose author can explain what they checked and why, beats slow and approximate work.

• The person can say what they produced themselves, what they had generated, and what they verified.
• Committing statements are sourced or checked, not merely well phrased.
• The reasoning holds when you ask a second-level question.
• The output fits the recipient and the internal context, not just generically correct.

Neither micro-management nor blind delegation

What stays managerial

Setting priorities, trading quality against deadline, deciding who does what, shielding the team from an incoherent request, reading a human situation: none of that goes to a system, because those decisions commit people and require context the tool does not have.

What a manager needs to understand

Not the engineering, but the properties of use: a probabilistic output, apparent confidence that is not a guarantee, dependence on the data provided, variability between runs. That base is enough to set credible team rules and to hold a conversation with a technical team.

A manager's role does not shrink because part of the production is assisted: it concentrates where it was always most useful — judgement, priority and accountability.

• AI literacy
• Copilots
• Human-in-the-loop
• Adoption