The Promotion You Didn't Apply For: Reviewing The Human-Agent Orchestrator

David Golub
23 hours ago
7 min read

Pascal Bornet and co-authors argue that the bottleneck on AI value has moved from the model to the manager. The diagnosis is right and the Orchestration Canvas is excellent.

Buried in The Human-Agent Orchestrator is a number that grabs the spotlight: of 108 senior leaders the authors interviewed, 68% said the gaps in their own management approach, not the AI itself, were to blame for their most significant AI setbacks.

When AI deployments go south, the argument runs, don't blame the model or the vendor. Look in the mirror and recognize flaws in management. The first chapter gives the problem its name. Working with agents that reason, plan and act should be understood as a promotion into AI management, a job change most of us never applied for. Even worse, most leaders are running the new job on an old playbook.

The authors count themselves among the promoted. "This book exists because we got it wrong first," lead author Pascal Bornet wrote on X at the book's launch, describing hundreds of deployments that failed "not because the technology broke," but "because nobody had built the management layer around it."

That sobering outlook forms the spine of this May 2026 book from Bornet and his eight co-authors, with a foreword by Marshall Goldsmith and an afterword by Cassie Kozyrkov. The book emerges as a follow-up to their 2025 Agentic Artificial Intelligence, which my colleague Nancy Wang reviewed on this blog and praised for its real-world grounding and its five-level map of agent capability.

To recap, that previous book explained what agents can do. This one argues that nobody has learned how to lead them, and that this widening gap has become the rate-limiting factor on generating AI value.

Verdict up front: from where our practice sits, the diagnosis seems spot on and the central instrument is timely, even if the evidence base deserves more skepticism than the packaging invites.

The Bottleneck Is the Manager

The book opens with a thought experiment destined for a hundred keynotes. Your company deploys a perfect AI, a real-life Jarvis. No hallucinations, no misunderstandings. For the first few weeks, it's magical. But soon enough, performance stutters and transformation stalls.

The authors' explanation cuts to the core of their adoption thesis: "The bottleneck used to be the technology. Now it's the people running it."

From there the argument turns to an AI readiness assessment of 432 companies across seven industries. Roughly 60% of organizations sit at what the book calls Level 1, tool adoption, harvesting 20 to 30% productivity gains. About 15% reach Level 3, organizational transformation, where gains of 100 to 200% show up.

Across all levels, the same technology yields radically different outcomes, and the difference, again, is management structure rather than model choice.

If that sounds familiar, it should. It is the same value-realization gap we mapped in Why AI Adoption Doesn't Equal ROI, and the external record backs the direction hard. MIT's Project NANDA found that 95% of enterprise generative AI pilots deliver zero measurable ROI despite $30 to 40 billion in investment.

Little wonder Gartner predicts over 40% of agentic AI projects will be canceled by the end of 2027, citing escalating costs, unclear business value and inadequate risk controls. And McKinsey's superagency research found only 1% of leaders call their organizations mature on AI deployment.

Bornet's thesis is not contrarian. It is the consensus of everyone who has watched a pilot die, stated with unusual clarity.

The Supervision Trap Strikes Again

Yet if management remains a choke on AI value, the supervision trap shows the complexities of human collaboration with agents.

Imagine the following scenario: a manager reviews every agent output. The agent makes one mistake. The manager reviews harder, finds more imperfections, trusts less, dials autonomy toward zero and ends up with a system slower than the manual process it replaced.

"Supervision doesn't scale to agent speed," the authors write. "It never will."

Anyone who has run agents in production has lived some version of this loop. A companion concept offered by the book turns even darker: the Human Crumple Zone, the human positioned to absorb blame for oversight that was structurally impossible, grounded in the 2018 Uber autonomous-vehicle fatality in Tempe, Ariz.

Supervision, the authors argue, often functions as liability architecture rather than quality control. And the diagnostic is delightfully practical. They call it the Holiday Test: if your AI work piles up while you are on vacation, you have a tool. If the system runs and flags true exceptions, you have a system.

Part II plots an escape route from this trap, grounded in what the book refers to as the Autonomy Dial. Autonomy, the authors insist, is not a property of an agent. It is a management decision about this agent, on this task, under these conditions, set on a capability-by-criticality matrix and reassessed on triggers.

This is the same shift we argued for when we told marketing leaders to get comfortable being human-on-the-loop: stop approving every action and start engineering the conditions under which actions need no approval.

The wider literature is converging on the same role. Harvard Business Review argued in February 2026 that companies now need dedicated agent managers the way the software era needed product managers. Ethan Mollick put it more bluntly in Management as AI superpower: "the skills that are so often dismissed as 'soft' turned out to be the hard ones."

The book's contribution is not discovering the role. It is writing the job description.

The Orchestration Design Canvas

Strip away the clever parables and the book is a delivery vehicle for one instrument: the Orchestration Design Canvas, six layers the authors brand the 6S.

Source (intent and the irreplaceable human capabilities)
Success (outcomes with good and bad output examples)
Safety (a named owner, hard constraints, a decision log)
Steering (a decision-rights map)
Switch (exception triggers with default states)
Sharpen (drift monitoring and feedback loops)

The appendix populates the full Canvas for a fictional client-briefing agent, down to escalation response times and a manual fallback protocol, a legitimately useful operational artifact. The field-failure stories that motivate each layer are the argument at its most persuasive.

A UK mortgage agent with no default state confirms eight wrong rates in four hours, at a cost just over £300,000, because the human notification sat unread from 2:17 pm to 6:44 pm. A staffing agency's agent autonomously reassigns an ICU nurse's shift 63 times over 11 weeks before the union notices.

Those stories distill into the one-liner that should outlive the book: "Agents do not take control. Humans abdicate it."

From a European perspective, with the EU AI Act's obligations biting from August 2026 and penalties reaching €35 million or 7% of global turnover, this kind of documented governance is becoming less a best practice and more a legal artifact.

The US offers the mirror image. Texas and California put new AI governance laws into effect on Jan. 1, 2026, Colorado's AI Act followed on June 30 and a Dec. 11, 2025 executive order established a federal task force to challenge those same state laws in court. In Mobley v. Workday, a federal court allowed discrimination claims to proceed on the theory that an AI screening vendor acts as an agent of the employers using it.

For American readers, the Canvas is what legal discovery will ask for. On the regulatory moment, the book is well aimed.

Where the Numbers Get Too Tidy

Now the grain of salt. The book's headline statistics are almost entirely self-sourced, and the methodology behind the 432-company assessment and the 108-leader interviews is sketched rather than shown.

Orchestrators averaged 73% productivity gains while supervisors averaged 28%. A Harvard study allegedly finds a 0.69 correlation between human-management skill and agent-management success. Roughly 70% of management transfers; the other 30% causes most of the failures.

These numbers are suspiciously round, arrive without confidence intervals and do a lot of the book's heavy lifting.

The anonymous case studies have the same texture: precise-sounding figures, unnameable clients, resolutions that land a little too cleanly. None of this means the claims are wrong. Independent research points the same way; the Harvard Business School Cybernetic Teammate field experiment with 776 professionals at Procter & Gamble found individuals working with AI matched the performance of full teams without it, which is the orchestration dividend measured properly.

These nuggets are compelling, but a book this insistent on decision logs and audit trails for agents should hold its own evidence to the standard it demands of its readers.

Two smaller gripes. The framework count borders on taxonomy fatigue: by our tally Parts I and II alone coin more than 20 capitalized concepts, and not every reader needs the Benevolence Trap, the Anthropomorphism Trap and the Method Trap filed separately. And the sample edition's structure, with Parts III and IV compressed into citable summaries plus a steady drumbeat of pointers to the orchestrator.coach companion product, makes the reading experience feel like the top of a funnel.

Nancy noted the predecessor's practitioner instincts when she reviewed it last year. Those instincts survive intact here. The packaging around them has grown considerably more ambitious.

Should You Read Human-Agent Orchestrator?

Read it if you own an agent deployment and your governance today is just a prompt and good intentions. Part II and the appendix alone are worth the cover price; treat the Canvas as a starting template and populate it against your own workflows.

On the other hand, skim the book if you want proof, because the proof here is directional and the verified numbers live elsewhere. Skip the parables on a second pass; you will have absorbed the pattern by chapter five.

Kozyrkov's afterword closes on a reworking of the famous Transformer paper title: "Attention is all you need." She means the leader's attention, deliberately directed, as the scarce resource agents cannot replicate.

That's a clever end note, and it matches what we see in production every week. The organizations feeling good about their agents aren’t necessarily the ones with the most elaborate models. Rather, they are the ones running agentic systems with a management discipline that is documented, staffed and accountable.

This book will not do that work for you, but it will hand you the clipboard.