Agentic AI Is a Delegation Problem. Leaders Already Know How to Solve It
The Leadership Question
Agentic AI moves enterprise AI from assistants that answer questions to multi-agent systems that act across workflows such as customer resolution, financial reconciliation, and supply chain routing. Executive coach Michael Rolph argues the central challenge is delegation: deciding which decisions agents may make alone, who owns the outcome, and whether the people in the loop are empowered to stop them. Deloitte's 2026 State of AI in the Enterprise survey found only 21% of organizations have mature governance for agentic AI, and Gartner expects more than 40% of agentic AI projects to be canceled by the end of 2027.
As enterprises move from AI copilots to multi-agent systems that act on their own, the hardest design work isn't technical. It is deciding what leaders will hand over, who answers for the result, and whether anyone feels authorized to say stop.
For the past two years, most enterprise AI has waited to be asked. A copilot drafts the email; a person decides whether to send it. That arrangement is ending. The next wave, agentic AI, consists of systems that take a goal and act on it: resolving a customer case end to end, reconciling accounts, rerouting a shipment. Increasingly they work in teams, with one agent handing work to another across the legacy software and APIs a company already runs.
The AI Summit London's 2026 trends briefing puts this shift first on its list: single assistants giving way to policy-aware, multi-agent workflows running in production with guardrails, human oversight, and audit trails. Practitioners it features, from Lloyds Banking Group to NHS Blood and Transplant, describe agents embedded in everyday processes and spreading across departments. For the European leaders I work with, that conversation is already shaped by the EU AI Act, which makes oversight and auditability design requirements rather than afterthoughts.
The appetite is clear. Deloitte's 2026 State of AI in the Enterprise survey of 3,235 business and IT leaders in 24 countries found that 74% expect their companies to use AI agents at least moderately by 2027. McKinsey's State of AI report, published in August 2026, found that 40% of organizations with more than $1 billion in revenue are now scaling agents in at least one function, up from 27% a year earlier.
The readiness is not. Only 21% of Deloitte's respondents say their organizations have mature governance for agentic AI. Gartner expects more than 40% of agentic AI projects to be canceled by the end of 2027, citing escalating costs, unclear business value, and inadequate risk controls.
That gap is Structural Lag in its sharpest form: the distance between what an organization's tools can do and what its human systems are ready to absorb. When AI only suggests, a weak human system costs you quality. When AI acts, it costs you control.
What actually changes when AI starts acting?
A copilot is a tool. An agent is closer to a new kind of team member. It holds a task, makes intermediate decisions, and hands work to others. That shifts the leadership question from adoption ("Are people using it?") to delegation ("What are we authorizing it to do, and who answers for the result?").
Gartner projects that at least 15% of day-to-day work decisions will be made autonomously through agentic AI by 2028, up from essentially none in 2024. Every one of those decisions currently belongs to a person. Someone has to decide which ones move, and on what terms.
Leaders already know how to do this work. Delegating to people means defining scope, decision limits, escalation paths, and how you will know it is going well. The pattern I see in coaching is consistent: leaders who delegate vaguely to people will delegate vaguely to agents. Unclear mandates, fuzzy authority, and surprise when something goes wrong do not disappear because the delegate is software. This is why I treat AI fluency as a relational skill rather than a technical one.
Why isn't "human in the loop" enough on its own?
Nearly every agentic design promises a human in the loop. Deloitte's research names what most organizations are missing to make that real: clear limits on which decisions agents can make alone and which need approval, real-time monitoring that flags anomalies, and audit trails of agent actions.
Those are the mechanics. The harder part is human. Anyone who has reviewed high volumes of routine work knows how quickly review becomes ratification. And in many organizations, overriding a system the executive team championed feels like a career risk. If the person in the loop doesn't believe they will be backed for pausing an agent, the loop is decorative.
A human in the loop is only a control if that person is willing, and authorized, to say stop.
This is Earned Candor applied to automated work. People flag what an agent got wrong only when leaders have shown, consistently and over time, that raising a problem is safe and gets acted on. No oversight policy creates that. Leadership behavior does.
Where do multi-agent systems break down?
Orchestrating agents across existing systems, rather than replacing core IT, is a pragmatic choice. It also means agents will operate across the same seams where organizations already struggle: the handoffs between finance and operations, customer service and logistics, IT and the business.
In 2023, I designed and facilitated an executive retreat for the seven-person leadership team of a mission-driven organization. The diagnostic work beforehand surfaced a familiar pattern: strategic focus that wasn't shared across the team, accountability that depended on who was asking, and disagreements left unresolved in asynchronous threads. None of it involved AI. All of it is exactly where a multi-agent workflow stalls. An agent can route a decision to "the owner" only if the organization has agreed who the owner is.
Agents don't fix ambiguous ownership. They execute it faster.
McKinsey's data points the same way. The roughly 6% of organizations it classifies as AI high performers are more than three times as likely as others to be scaling agents in most business functions, and nearly three-quarters of them report fundamentally redesigning workflows because of AI, compared with about one-quarter of everyone else. The redesign comes before the scale.
What should leaders decide before agents scale?
Three decisions belong at the leadership level. They should not be delegated to IT or to a vendor.
1. Decision rights. For each agentic workflow, specify which decisions an agent can make alone, which need human approval, and which stay entirely human. Write it down the way you would a delegation of authority for a new executive.
2. Ownership of outcomes. Every agent action should roll up to a named person accountable for its results. Not the vendor, and not "the system."
3. Permission to stop. Make pausing an agent a protected act. The first time someone halts one for a good reason, recognize it publicly. That single moment will teach your organization more about oversight than any policy document.
For organizations operating in Europe, there is a regulatory clock as well. Under the EU's AI Omnibus, published in the Official Journal in July 2026, obligations for high-risk AI systems listed in Annex III of the AI Act, including human oversight, now apply from 2 December 2027 rather than August 2026. The extra time is runway for building organizational readiness, not a reason to wait. My organizational AI coaching work with cross-border leadership teams increasingly starts here.
The leadership opportunity
Agentic AI will reward organizations that already know who decides what. That was always true of good management. Agents simply remove the slack that let ambiguity survive.
AI transformation stalls at the human-systems layer, not the technology layer. With agents, it stalls faster.
If you are deciding what your organization will hand to agents, and who will answer for the result, Book a Conversation.
Michael Rolph is an executive coach who helps founders, operators, and senior leaders lead AI transformation as a human-systems challenge. He currently coaches the C-suite of a publicly traded U.S. company through enterprise AI adoption and developed the Structural Lag and Earned Candor frameworks at the center of his practice. Based in Rabat, Morocco, he works with leaders across the SF Bay Area, Europe, North Africa, and the Middle East.
Sources
Deloitte Insights, "Business and IT leaders report AI agents are scaling faster than their guardrails," April 2026 (State of AI in the Enterprise, 2026).
McKinsey & Company, "The state of AI in 2026: On the road to ROI," August 2026.
Gartner, "Gartner Predicts Over 40% of Agentic AI Projects Will Be Canceled by End of 2027," June 2025.
The AI Summit London, "Top AI Trends Shaping Business and Innovation in 2026," 2026.
Future of Privacy Forum, "The AI Act Implementation Timeline: What Changes Under the AI Omnibus?," July 2026.