Your team starts using AI because everyone is already too busy.
Emails get drafted faster. Proposals appear in minutes. Reports that used to take a day arrive before lunch. People can explore more ideas, create more options and move several pieces of work at once.
Then something strange happens.
The manager gets busier.
There is more to review. More to approve. More decisions to make. More work that looks almost right but needs someone experienced to catch the thing it misunderstood.
The team has become more productive, but the person responsible for the outcome feels like they are drowning in everyone else’s productivity.
AI was supposed to remove the busywork. Instead, it created more work that needs a manager.
AI can increase output without increasing autonomy
Most AI rollouts begin by helping individuals produce things faster.
That is useful. But producing the work was only one part of the job.
Someone still needs to know:
- whether this is the right thing to produce;
- whether it fits the organisation’s goals;
- whether an exception is acceptable;
- whether the quality is good enough;
- whether another team can deliver what has been promised; and
- who is accountable when the answer is unclear.
If those decisions still live in one manager’s head, every faster employee and every new agent sends more work back to that manager.
The organisation has accelerated the outside of the loop while leaving the centre unchanged.
AI does not remove the bottleneck if every agent still has to ask the same manager what to do.
The manager may be the undocumented specification
In many businesses, the real rules are not in the systems or process documents.
They live in the experience of a manager or owner.
The manager knows which customer deserves an exception. They know when a proposal can go below the standard margin. They know which risk is acceptable, which project is worth squeezing in and which wording will worry the client.
Years of experience make these decisions feel obvious to them.
They are not obvious to everyone else.
So the team asks.
The manager answers, sometimes consistently and sometimes according to the pressure of the day. The next situation is similar but not identical, so the team asks again.
The manager becomes frustrated that nobody shows initiative. The team becomes frustrated that it is expected to predict decisions whose logic has never been made clear.
If the manager is the only place the organisation’s decision logic lives, every person and agent has to keep asking them.
Your team may not have an initiative problem
Leaders often say they want their people to take more ownership.
But ownership is risky when the boundaries are invisible.
Imagine a salesperson deciding whether to adjust a quote. The sales manager believes they should use judgement. Delivery believes the margin must be protected. Leadership regularly approves exceptions when a large contract is close to signing.
What does “take ownership” mean in that situation?
If the salesperson follows the calculator, they may lose the deal. If they add scope, delivery may reject it. If they reduce the price, finance may object. If they ask, the manager wonders why they could not decide for themselves.
Escalating is rational when the cost of guessing wrong is higher than the cost of interrupting the manager.
Sometimes the team is not refusing to take responsibility. The organisation has never made responsibility safe to exercise.
More agents can create more interruptions
The same problem becomes sharper with AI agents.
An agent can draft fifty proposals, analyse a pipeline and generate a dozen options before a person has finished their first coffee.
But if it does not understand the commercial goal, the protected margins, the available trade-offs and the limits of its authority, each output comes back for interpretation.
The manager is no longer reviewing one salesperson’s work. They are reviewing the work of every salesperson plus the agents working for them.
This is an AI leash at organisational scale. The technology can move quickly, but it keeps pulling the same person back to review, redirect and explain.
AI can multiply the work that was already not working.
There are names for the two patterns that follow. Approval Addiction is when the organisation adds a human check after every meaningful agent action, so the manager becomes the permanent supervisor of routine work. Permission Inflation is when every unclear decision escalates upwards because nobody has defined who may decide what. Both are among the six AI agent failure modes, and both mean more agent output translates directly into more management load. If decision rights stay where they are, every extra unit of output is another item in the same person's queue.
A quick diagnostic for your own review queue:
- Which of these checks are genuine decisions?
- Which are repeated reassurance?
- Which decisions could be made once and encoded into the specification, so the agent can act without asking?
A specification makes judgement reusable
The answer is not to automate the manager’s judgement or attempt to turn every decision into a rigid rule.
It is to make enough of the recurring logic visible that people and agents can act without starting from zero.
A living specification can explain:
- the outcome the team is responsible for;
- the goals it is working towards;
- the rules that must always hold;
- the trade-offs that are acceptable;
- the decisions the team can make itself;
- the situations that require the manager; and
- the questions the organisation has not yet answered.
This does not capture everything the manager knows.
It captures the part that everyone should not have to rediscover one interruption at a time.
The manager should receive decisions, not unfinished thinking
When an agent reaches a genuine boundary, it should still involve a person.
But it should not simply say, “What do you want me to do?”
It should bring the decision in a form that respects the manager’s attention:
- what happened;
- why the existing specification does not fully answer it;
- which goals or boundaries are affected;
- the viable options;
- the likely trade-offs; and
- who else will be affected by the choice.
The manager provides the judgement and authority. The agent carries the coordination.
That is a very different experience from being dragged into the middle of every task.
What does the manager do instead?
Interestingly, the manager’s role should not change much on paper.
A sales manager, marketing manager or team leader is already supposed to:
- help the team perform;
- develop the people who report to them;
- improve how the work is done;
- translate business goals into team direction; and
- remain connected to the outcomes.
In practice, reactive delivery work takes over.
The manager spends the day answering questions, fixing handovers, approving exceptions, finding information and personally carrying work that the team could not safely complete without them.
In a spec-driven organisation, the manager’s job does not become something futuristic.
It finally becomes the job it was supposed to be.
Agents handle more of the reactive coordination and routine execution. People and agents can make recurring decisions inside clearer boundaries. The manager is brought in for fewer, more important choices.
They spend more time asking:
- What must we achieve this quarter?
- What is preventing the team from reaching that goal?
- Which capability should we develop?
- What are we learning from customers?
- Which part of the process should improve next?
- Does the current specification still reflect our strategy?
They remain close to the work without being trapped inside every action.
The promise is calm, not distance
The goal is not to disconnect leaders from what is happening.
It is to connect them to the decisions and outcomes that deserve their attention.
The manager should still understand what customers are saying, how the team is performing and where delivery is struggling. They should see patterns and receive evidence that helps them steer.
They should not have to manually collect all of that information or approve every step that produced it.
When this works well, the emotional change is a sense of calm and control.
The business may be moving faster than before. More work may be happening. But the manager is making fewer low-value decisions and receiving better context for the high-value ones.
A spec-driven organisation can feel calmer while it moves faster.
How to find where the manager is still the specification
Look at the questions that repeatedly travel upwards.
For one week, ask managers to notice every time someone comes to them for a decision. Do not judge the person asking. Sort the interruptions into three groups.
1. We have already made this decision
The manager is repeatedly explaining an existing rule, preference or trade-off.
That logic should become easier for the team and its agents to find and use.
2. We have never made this decision clearly
The question is exposing a real gap.
Record it as an open decision. The manager decides whether it needs an answer now or whether work should continue around it.
3. This situation is genuinely new
The work has reached a boundary where experience and authority matter.
This is exactly where the manager should be involved. The result may remain a one-off exception, become an experiment or lead to an intentional change in the specification.
The goal is not zero interruptions.
It is to stop paying for the same interruption again and again.
Bottom line
AI makes managers busier when it creates more work without making the organisation’s decision logic easier to use. If the manager remains the undocumented specification, every employee and agent still has to return to them for interpretation, approval and direction.
A living specification makes recurring judgement reusable. It gives people and agents enough clarity to act, makes unanswered decisions visible and reserves the manager’s attention for the situations that genuinely need it.
That is how AI begins to create strategic space instead of another stream of work to review.
BusyWork Dispatch helps businesses turn AI ideas into working software, automations and agents, supported by a Workbook that preserves the context each request needs. See how BusyWork works or book a call with Ben.
