The Costly AI Mistake: Chasing Copilots, Not Workflows

AI is changing management by shifting attention from supervision to orchestration. The real value comes from redesigning workflows, decisions, capabilities, and accountability around AI-enabled work.

This episode explores how leaders can redefine management as routine coordination becomes increasingly automated.

TLDR / At a Glance

• Supervision giving way to orchestration
• Workflow redesign as the value driver
• Five-part Orchestration Stack
• Rising skill demands in junior roles
• Governance as a management responsibility
• Exception handling and decision ownership

Flattening structures without redesigning management risks relocating friction rather than removing it, while deliberate orchestration creates clearer accountability and stronger AI-enabled performance.

Transcript

Copy Transcript Copied

The Costly Question About Flattening

SPEAKER_00 0:00

The costly AI mistake chasing co pilots, not workflows. Every few years a technology arrives and the same prediction follows. The middle of the organization is about to disappear. This time the prediction has better data behind it, and it is still the wrong question. Gartner expects a fifth of organizations to use AI to flatten structures by the end of 2026, and some firms are already removing layers. But layer removal describes a headcount decision, not a capability shift, and the two are being confused at board level with expensive consequences. The more useful question is what management becomes when machines perform more of the work. This article argues that AI is not eliminating management so much as separating the parts of it that were always administrative overhead from the parts that were always the actual job, and that the organizations converting AI into performance are the ones redesigning the manager role deliberately, rather than letting attrition and automation redesign it by accident. The prediction

What Work Management Really Contains

SPEAKER_00 1:06

that keeps being half right. The flattening argument has a respectable logic to it. Much of what middle management historically did was move information upwards, translate intent downwards, coordinate across silos, and check that work had been done. If software can do all four, the reasoning goes the layer that did them becomes surplus. This is not a straw man argument, and dismissing it makes an article comfortable rather than useful. Where the argument breaks down is in what it assumes about the residual. It treats management as a fixed bundle of tasks that can be subtracted from until eventually nothing is left. What the evidence actually shows is a bundle being pulled apart, with the coordination and reporting components automating quickly, while the judgment, capability, and accountability components become harder and more valuable. Deloitte finds that nearly 40% of managers' time is consumed by firefighting and administration, and the 36% say they feel unprepared for the people leadership parts of their role. Those two numbers sit together for a reason. Managers have been spending their time on the work that AI can do and neglecting the work that only they can do. The signal hiding in plain sight. BCG's

The Rise Of The Agent Boss

SPEAKER_00 2:33

2026 Workplace Research contains the single most revealing figure in this debate. Nearly half of respondents, 47%, now report spending more time managing and directing AI than performing the work themselves. Read that carefully. It does not describe managers being replaced, it describes everybody becoming a manager. This is what Microsoft has started calling the agent boss, and the framing is more than a coinage. If a substantial share of the workforce now spends its day setting direction, checking output, catching errors, and deciding what to escalate, then managerial work has not been eliminated. It has been distributed. The organization has more management happening inside it than before, performed by people who were never trained for it, without decision rights, escalation rights, or quality thresholds to guide them. That distribution is precisely why the manager role becomes more important rather than less. Someone has to design the system in which all this informal supervision takes place. Someone has to decide what good looks like when the first draft comes from a machine. Someone has to own the outcome when it goes wrong at three in the morning across a function that no longer has a night shift. Why

Why Layer Removal Moves The Work

SPEAKER_00 3:48

removing layers does not remove the work? Consider what actually happens when a firm flattens without redesigning. The reporting lines shorten, the org chart looks cleaner, and the coordination work simply relocates. It reappears as longer executive meetings, as cross-functional confusion, as decisions that stall because nobody is quite sure who owns them, and as quality problems that surface late because the checking layer was removed before the checking was replaced. Bain's research on this is unusually direct. AI focused reorganizations underperform other reorganizations, and they underperform because leaders fail to help people work differently. The structural change was the easy part. The behavioral and design change was skipped. McKinsey's numbers sharpen the point considerably. Among organizations seeing meaningful bottom line impact from AI, 55% have fundamentally redesigned workflows compared with 20% of everyone else. Workflow redesign is one of the strongest single contributors to whether AI produces value at all. ServiceNow reaches the same conclusion from a different angle, finding that redesign matters more to returns than budget, talent, or model choice, while enterprise AI maturity actually fell nine points year on year. The market is getting better tools and worse results because tools were never the constraint. IBM's finding that only 25% of AI initiatives have delivered their expected return and only 16% have scaled across the enterprise is not a story about immature technology. It is a story about mature technology dropped into unredesigned work. The orchestration

The Orchestration Stack Explained

SPEAKER_00 5:38

stack. If supervision is going and orchestration is arriving, the role needs a structure rather than a slogan. The orchestration stack sets out five accountabilities that replace what supervisory management used to do. They are layered deliberately. Each one only works if the one beneath it holds. Workflow designer. The manager maps where AI compresses cycle time, where steps can now run in parallel, and where human review genuinely adds value rather than adding delay. The discipline here is refusing to automate the existing process. The existing process encoded constraints that no longer exist. Capability builder, less time policing routine output, more time developing judgment, domain fluency, validation instinct, and escalation discipline. This matters more than it sounds. PWC's 2026 barometer finds that AI exposed junior roles are seven times more likely to demand traditionally senior skills, such as leadership, and that AI is professionalizing work rather than simply thinning it. Junior people are being handed senior problems earlier. Somebody has to develop them for it. Decision Architect The manager states explicitly which decisions are human-led, which are machine assisted, which are machine executed under human oversight, and which are reserved for escalation. Most organizations have never written this down for any decision, let alone for the ones AI now touches. Governance owner, role appropriate control over data use, transparency, override rights, auditability, and exception review. In the UK and Europe, this is no longer a legal department concern. The EU AI Act treats several employment-related uses as high risk and places AI literacy duties on deployers, and the ICO's guidance puts accountability and bias mitigation squarely on the organization using the system. Exception manager. As workflows automate, managerial attention migrates to the edge cases. The ambiguous, the cross-functional, the escalated, and the failed. This is where the residual value of management concentrates and it is the accountability most often left unassigned. The stack is not a job description, it is a way of asking whether the management work in a given function has been designed or merely inherited. What

What Orchestration Looks Like In Practice

SPEAKER_00 8:10

this looks like in practice. Telstra offers one of the clearest structural illustrations. Deloitte cites the company separating the role of leader of people from the role of leader of work, which is management disaggregated rather than management deleted. The capability building and the delivery orchestration became distinct jobs, because doing both badly had become the default. A mining operation cited in the same research makes the exception manager concrete. Automated, remotely operated trains took over the driving. Drivers were retained for recovery when automation failed, and their roles expanded to include training new recruits. Notice what happened there. The routine work automated, the exception work stayed human, and the released capacity went into capability. That is the orchestration stack running end to end in an operational setting, and it was designed rather than discovered. In professional services the pattern is subtler but the same. Thomson Reuters finds organizational use of generative AI in the sector rose from 12% to 22% in a single year, concentrated in summarization, document review, risk reporting, and drafting. For a partner or practice lead, that does not mean supervising each piece of work. It means setting review thresholds, defining which matters get human eyes and why, and owning the quality of a system rather than the output of an individual.

Leadership Support Drives Real AI Value

SPEAKER_00 9:43

The barrier is leadership, not willingness. There is a comfortable story in which AI transformation stalls because employees resist it. The evidence does not support it. Adoption is already broad, with Microsoft reporting three quarters of knowledge workers using AI at work and Stanford's AI index putting organizational adoption at 78%. People are using it. They are often using it without being told how. Gallup's finding is the one worth putting in front of an executive team. Employees whose manager strongly supports AI use are twice as likely to use it frequently, and nearly nine times as likely to say it helps them do what they do best. The manager is the conversion layer between technological possibility and actual performance. Remove that layer or leave it untrained and unclear about its own remit, and adoption produces activity rather than value. McKinsey's superagency work reaches the same conclusion in blunter terms. Employees are ready, and the constraint is leadership. That should be read as good news. Constraints in leadership design are fixable in a way that cultural resistance is not.

Limits, Risk And Governance Lines

SPEAKER_00 10:57

The honest limits of this argument two overclaims would weaken everything above. The first is that management is safe, it is not in its current form. Some layers will go, some functions will shrink, and a good deal of managerial task content will be automated outright. The distinction that matters is between the work disappearing and the work changing hands, and the evidence points overwhelmingly to the latter. The second is that agentic AI is ready to run the enterprise. It is not. Gartner expects more than 40% of agentic AI projects to be cancelled by the end of 2027 on grounds of cost, weak value, or inadequate controls. Any argument for orchestration that assumes the things being orchestrated are reliable has skipped a step. The oversight is not a transitional inconvenience. For now it is the design. There is also a governance line that should not be crossed casually. AI-enabled performance management and workforce monitoring are technically available and legally fraught. The OECD's work on algorithmic management shows managers perceiving decision quality gains while reporting real concerns about accountability and explainability. That tension is not resolved by better models. It is resolved by clearer human authority over where the model's word stops. What

Five Board Level Questions To Audit

SPEAKER_00 12:20

this means for leaders. The board question is not how many managers you need. It is whether the management work in your organization has been designed for the way work now happens, or whether it is still shaped by coordination bottlenecks that stopped existing 18 months ago. Most executive teams have never audited this. They have audited headcount, which is a different exercise producing a different answer. Five questions are enough to start. Which managerial tasks should disappear entirely? Which decisions need redesigning rather than accelerating? Where should AI, Act, recommend, or stop? Which managerial capabilities matter most now? And are you building them? What governance and exception routes actually exist as opposed to existing on a slide? If those cannot be answered function by function, the operating model has not caught up with the technology, whatever the adoption dashboard says. IBM's finding that 77% of CEOs see talent and technology leadership converging, and that 76% of organizations now have a chief AI officer type role, against 26% a year earlier, tells you the reinvention is running from the boardroom down as much as from the front line up. This is not a middle management story. It is a management story. Technology creates possibility. Management creates value. The organizations that win with AI will not be the ones that remove managers fastest. They will be the ones that redefine management earliest. This

Redefine Management Early, Not Fast

SPEAKER_00 13:57

concludes the article. You can also read this article on my LinkedIn page where I share regular insights on AI, strategy, and emerging technologies.