The AI Middle Manager Problem

There is a reason so many AI management demos look persuasive.

Middle management produces a lot of text. Status updates. Follow‑ups. action lists. meeting summaries. weekly reports. staffing notes. risk logs. roadmap roll‑ups. dependency trackers. executive briefings. If you only look at the visible surface of the job, a large part of it appears to be an information‑formatting problem.

And to be fair, part of it is.

AI can already make management work cleaner in several obvious ways. It can reduce note‑taking overhead, draft updates, cluster themes from meetings, organise actions, identify duplicated work, surface blockers, and compress communication that would otherwise sprawl across half a day. That is real leverage. It is also where a lot of organisations are about to confuse administration with management.

That confusion matters because weak management was already expensive before AI arrived. It just tended to hide under respectable‑looking output. Too many updates. Too many dashboards. Too many meetings. Too little clarity. Too little ownership. Too little decision‑making. AI can make that kind of status theatre much cheaper to produce.

The risk is not that management becomes unnecessary. It is that companies start automating the most visible parts of management while underestimating the parts that actually make teams work.

Why Management Looks Especially Automatable

Management has always contained a large amount of coordination work, and coordination work leaves paperwork behind.

That makes it unusually easy to reduce management to artefacts. A director can look at the calendar load, the number of updates, the length of the project plan, or the volume of reporting and conclude that much of the role should now be automated. In a narrow sense, that conclusion is often correct. Several management‑adjacent tasks are exactly the kind of language and synthesis work that current models handle well enough to save time.

NBER's Shifting Work Patterns with Generative AI is a good example of the pattern. In a field experiment across knowledge workers using AI tools already integrated into email, meetings, and writing workflows, treated workers spent less time on email and less time working outside normal hours. That does not prove management work disappears. It does show that a meaningful amount of coordination overhead is compressible.

The mistake is what usually comes next. If meeting notes, updates, and summaries are more compressible, leaders start treating the management layer itself as overhead. The visible routines become the definition of the role. Once that happens, the part that matters most is no longer what gets measured.

Status Theatre Was Already a Management Failure

Long before generative AI, many organisations had a management problem disguised as an information problem.

You can usually recognise it quickly. Plenty of visibility, weak understanding. Lots of tracking, poor decisions. A well‑maintained dashboard telling you everything that happened last week and very little about what should happen next.

AI does not create that failure mode. It accelerates it.

This is closely related to the output trap in The AI Productivity Mirage. If the organisation already rewards visible artefacts over meaningful outcomes, then AI becomes a cheap way to improve appearances. More polished updates arrive faster. Meeting recaps sound sharper. Risks are categorised more neatly. The reporting machine gets smoother while the quality of prioritisation stays flat.

That is why the middle‑manager question should not start with "How much reporting can we automate?" It should start with "Which parts of this role were actually valuable before we started counting the paperwork?"

What AI Can Automate Usefully in Management

There is no value in pretending AI is bad at everything management‑related. It is not.

Used properly, it can help managers:

• draft status summaries without spending an hour formatting them
• extract action points and owners from long meetings
• compare overlapping plans across teams
• synthesise recurring concerns from one‑to‑ones or project updates
• prepare first‑draft briefings for senior stakeholders
• spot where the same blocker is appearing across multiple projects
• reduce low‑value email churn and repeated explanation

Those are legitimate improvements. Some of them are precisely the kind of augmentation that makes skilled managers more effective. They spend less time converting raw discussion into shareable artefacts and more time making decisions, coaching people, and removing constraints.

The broader evidence on augmentation points in that direction. Anthropic's Economic Index found observed usage concentrated in writing, coding, and knowledge tasks, and overall leaning slightly more toward augmentation than automation. That fits management work fairly well. The role contains many tasks AI can accelerate, but the role is not reducible to those tasks.

What Good Managers Actually Do

Management research is still useful here because it is a reminder that managerial value has never been mostly clerical.

Bloom and Van Reenen's long‑running work on management practices argues that differences in management quality are strongly associated with large differences in productivity and performance across firms. Better management is not the same as better filing. It is a capability that shapes how information is used, how decisions are made, how accountability works, and how teams actually coordinate under pressure.

That matters because the parts of management that produce durable value are not the parts a large language model safely owns.

Good managers decide what not to do. They prioritise across competing truths. They recognise when a clean metric is hiding a dirty reality. They mediate conflict when two teams have locally sensible goals that cannot both be satisfied. They coach somebody through a performance problem that is partly technical, partly behavioural, and partly contextual. They take incomplete information and turn it into a direction that a team can act on.

Those are judgement problems.

The more your organisation speeds up the production of visible artefacts, the more valuable those judgement calls become. Faster updates mean less if nobody is deciding correctly between them.

Weak Managers May Be More Exposed than Strong Ones

There is an uncomfortable side effect to all of this.

AI is likely to expose weak managers more aggressively than it exposes strong ones.

If a manager's main contribution was already forwarding updates, reformatting risks, paraphrasing meeting notes, or asking the team for status so they can repeat it to somebody else, then the role was fragile before AI. The technology does not create the fragility. It removes some of the camouflage.

That does not mean every manager will disappear or that organisations can simply flatten structures and expect everything else to work. It means some managerial roles were carrying less substantive value than their reporting footprint suggested.

The stronger managers, by contrast, often become more obviously useful. They are the ones who:

• turn ambiguous information into a coherent decision
• challenge bad trade‑offs before they harden into roadmaps
• coach engineers through judgement‑heavy work, not only throughput work
• arbitrate sequencing when several teams are all locally rational and globally incompatible
• absorb organisational noise so specialists can do deeper work

That is one reason management quality and worker skill are often complementary in the research rather than substitutable. Management Practices, Workforce Selection and Productivity makes that relationship explicit, showing that better management practices interact with workforce quality and help shape firm performance.

In other words, strong managers do not merely pass information around. They increase the value of the people around them.

Dashboards Do Not Resolve Conflict

One of the easiest mistakes a leadership team can make is assuming that better visibility creates better alignment.

It does not.

Dashboards make disagreements easier to observe. They do not resolve it. A portfolio view can show that three teams are all depending on the same platform work, but it does not decide which team takes the delay. A meeting summary can capture that legal, product, and engineering disagree on a release risk, but it does not create the trust or authority needed to settle it. An AI‑generated plan can list all the steps in the rollout, but it cannot take accountability for the trade‑offs if two of those steps become mutually incompatible once reality arrives.

This is where a lot of "AI for management" material becomes oddly shallow. It assumes coordination is mostly an information‑distribution problem. In real delivery organisations, coordination is often a conflict‑resolution problem.

That point becomes even sharper when AI starts reducing the friction of implementation work. If engineers can draft faster, support teams can respond faster, and content teams can publish faster, then the cost of weak prioritisation rises. A bad decision can now propagate more quickly through the system.

Why Strong Managers May Become More Valuable

The middle‑manager problem is therefore not that AI makes management irrelevant. It is that AI changes what the role needs to be good at.

As implementation, drafting, and routine coordination get cheaper, managers who add value through judgement, clarity, and team development become more important. The leverage shifts upward from artefact production to direction‑setting.

You can see a similar pattern in collaborative work more broadly. The Cybernetic Teammate field experiment found that AI could replicate certain benefits of collaboration for professionals working on real innovation tasks, with individuals using AI matching the performance of teams without AI on some measures. That is a striking result, but it does not imply human management no longer matters. It implies that when AI takes over some parts of ideation and synthesis, the scarce part shifts further toward framing, coordination, judgement, and execution discipline.

Good engineering managers are already living in that shift. They are less valuable because they personally produce status documents, and more valuable because they know when to push back on a date, when to intervene in a code‑review dynamic, when a "small" platform change will create downstream operational pain, and when the team needs more context rather than more pressure.

That is closer to technical leadership than to bureaucratic reporting. It is why the distinction between senior individual contributor work and leadership work remains worth understanding.

Coaching Under Uncertainty is Still the Hard Part

One of the least automatable parts of management is helping somebody improve when the problem is not yet neatly defined.

An engineer may be missing signals in review without making obviously poor technical decisions. A product manager may be writing clear updates while repeatedly framing the wrong problem. A designer may be strong on interface craft and weak on trade‑off negotiation with engineering and commercial stakeholders. In all of those cases, the manager's job is not to summarise performance. It is to interpret patterns, challenge them carefully, and help the person change how they work.

That requires memory, trust, timing, and judgement. It also requires enough proximity to the work to know when a problem is genuinely individual and when it is structural. AI can help capture notes or synthesise themes across one‑to‑ones. It cannot safely own the responsibility for deciding what the person most needs to hear next, or how hard to push.

That is why management remains a high‑judgement role even as more of its visible paperwork becomes automatable.

How Engineering Managers Should Use AI Without Outsourcing Accountability

The useful question is not whether managers should use AI. They should, when it clearly removes low‑value friction. The useful question is how to use it without handing over the parts of the role that require human responsibility.

A practical rule is to separate compression from commitment.

AI is usually safe for compression tasks:

• summarising discussion
• drafting communications
• extracting actions
• clustering themes
• reducing repeated explanation

AI is much less safe for commitment tasks:

• deciding priorities between teams
• setting performance judgements
• resolving interpersonal conflict
• interpreting weak signals from a struggling team
• approving risk acceptance
• presenting a decision as if it has an accountable owner when it does not

If you manage engineers, product teams, or transformation programmes, a good pre‑use checklist is:

• Am I using AI to save time on representation, or to avoid making a hard judgement?
• If this summary is wrong in a subtle way, who will notice before it shapes a decision?
• Does this dashboard create clarity, or just the appearance of clarity?
• What part of the issue still requires human context that the model does not have?
• If something goes wrong, is the accountable owner still obvious?

If the answer to that last question gets blurry, the tool is being used too far up the chain of responsibility.

Management Still Decides What the Team Learns from Pressure

Another part of the role that becomes more important, not less, is translating stress into learning rather than drift.

When delivery pressure rises, weak teams tend to normalise ambiguity, skip difficult conversations, and mistake quiet compliance for alignment. Strong managers notice when a rushed plan is teaching the wrong lesson, when a senior engineer is quietly absorbing too much coordination debt, or when repeated context switching is degrading judgement even though the metrics still look acceptable.

AI can help show the pattern. The manager still has to decide what to do about it.

That decision is part of leadership because it shapes what the team will consider normal the next time pressure arrives.

Conclusion

AI can remove a meaningful amount of management overhead. That part is real. It can cut down status churn, compress email work, and make routine coordination less expensive.

But management was never worth paying for because it generated paperwork.

It is worth paying for when it improves judgement, sequencing, conflict resolution, coaching, and accountability under uncertainty. Those are the things that stop teams from mistaking activity for progress. They are also the things a dashboard cannot safely own on its own.

That is the real middle‑manager problem. AI can automate parts of management work. It can also expose how much of some management roles was only ever administrative theatre. The organisations that benefit will be the ones that use AI to strip out reporting friction while making the human parts of management more explicit, not less.