Every Agent Needs a Human Operating Model

An agent is not a tool, and not a hire

An AI agent creates an awkward management problem: it acts like labor but is funded, procured, and monitored like technology. A business can approve an agent, assign it a license, point it at a workflow, and still have no answer for who owns its work when it makes a judgment, escalates poorly, or produces an output that looks finished and is wrong. That gap is not a model-quality problem. It is a governance problem, and it is where most agent programs quietly fail.

The market shows the pattern. McKinsey reported in 2025 that 88 percent of respondents were using AI in at least one business function, while only 39 percent reported enterprise-level EBIT impact. Gartner projects that more than 40 percent of agentic AI projects will be canceled by the end of 2027, citing escalating costs, unclear business value, and inadequate risk controls. Deloitte puts the root cause plainly: agents are neither capital nor labor — they act like workers but are funded like technology — and that creates governance gaps in decision rights, risk and liability, quality assurance, and accountability. Its 2026 survey found that 84 percent of companies had not redesigned jobs to fit AI. Redesign is exactly the work that produces owners, decision rights, and escalation paths.

The operating move

Give every agent an operating model before it goes live. The test is whether a named human can answer six questions in advance: what the agent may decide on its own, what it must escalate, who owns its output, what evidence it must leave behind, how it is stopped, and what a good result looks like. Those answers assemble into a short canvas — role and scope, decision rights, an escalation and stop mechanism, a challenge protocol, acceptance criteria, an evidence and logging duty, and one accountable human owner.

None of this is new to anyone who has run a program or a portfolio. The discipline is applying it to a non-human actor before it starts working. Decision rights are written as decide, sign-off, or escalate, following Deloitte's split, with the evidence each decision must carry. A stop mechanism exists and has been tested, not assumed. Every agent leaves a trail sufficient to reconstruct what it did, in the spirit of the NIST AI Risk Management Framework and its Generative AI Profile. And one named person — not a committee, not the vendor — owns the outcomes, sitting “above the loop” as McKinsey describes it: setting policy and watching outliers rather than reviewing every action by hand.

The real ceiling is oversight, not capability

The canvas surfaces the constraint leaders underestimate: how many agents one human can actually own. McKinsey observes that a team of two to five people can already supervise 50 to 100 specialized agents — and, in the same breath, that the scale of agentic adoption will be capped by how much oversight capacity humans can provide, making governance itself a potential bottleneck. The practical consequence is that oversight is a real, budgeted cost, and the number of live agents should follow the number of humans who can answer for them — not the other way around. An agent without an operating model is not autonomy. Build the model first, then let the agent work.

Notes and sources

1. Alex Singla, Alexander Sukharevsky, Bryce Hall, Lareina Yee, Michael Chui, and Tara Balakrishnan, "The state of AI in 2025: Agents, innovation, and transformation," McKinsey & Company, November 5, 2025. Verified July 10, 2026. mckinsey.com
2. Gartner, "Gartner Predicts Over 40% of Agentic AI Projects Will Be Canceled by End of 2027," press release, June 25, 2025. Verified July 10, 2026. gartner.com
3. David Mallon, Brad Kreit, and Natasha Buckley, "Rethinking operating models for humans with agents," Deloitte Insights, April 2, 2026. Verified July 10, 2026. deloitte.com
4. Alexander Sukharevsky, Alexis Krivkovich, Arne Gast, Arsen Storozhev, Dana Maor, Deepak Mahadevan, Lari Hämäläinen, and Sandra Durth, "The agentic organization: Contours of the next paradigm for the AI era," McKinsey & Company, September 26, 2025. Verified July 10, 2026. mckinsey.com
5. National Institute of Standards and Technology, "AI Risk Management Framework Core," excerpt from AI RMF 1.0, 2023. Verified July 10, 2026. airc.nist.gov
6. National Institute of Standards and Technology, "Artificial Intelligence Risk Management Framework: Generative Artificial Intelligence Profile," NIST AI 600-1, July 2024. Verified July 10, 2026. doi.org