The researchers didn’t set out to write a leadership manual. They were building a technical framework for AI agents. But when a Google DeepMind team published Intelligent AI Delegation in February 2026, they arrived at a striking conclusion: the most rigorous protocols for safe AI orchestration are structurally identical to classical delegation training for human leaders.
That irony carries a practical implication. The executives most equipped to deploy agentic AI at scale aren’t necessarily those with the deepest technical knowledge but the ones who already know how to delegate well.
The hard part is applying that knowledge to a new kind of subordinate.
The Shift from Querying to Commissioning
Most organisations still treat AI agents the way they treated the first generation of software: you type something in, you get something back. That mental model breaks down the moment an agent has tools, memory, and the authority to take action in the world.
An agent with access to your CRM, your calendar, and your communications platform isn’t a search engine. It’s a delegatee. And like any delegatee – human or otherwise – it will interpret ambiguous instructions in ways that technically satisfy the brief while missing the point entirely.
This is what the DeepMind paper calls specification gaming: the agent fulfils the prompt but fails the business intent. It isn’t a bug in the model. It’s a delegation failure by the principal.
The antidote isn’t better prompting, but a formal transfer of context, authority, and accountability – the same elements that underpin effective human delegation.
| Delegation Dimension | Classical Human Delegation | Intelligent Agentic Delegation |
|---|---|---|
| Authority | Permission to make decisions within a defined scope | Access permissions and tool-use boundaries |
| Responsibility | The obligation to perform the assigned task | Commitment to goal-state optimisation |
| Accountability | Being answerable for outcomes | Verifiable execution logs and audit trails |
| Trust | Based on past performance and character | Calibrated against certifiable capabilities |
The Five Rights Applied to AI Agents
High-stakes human disciplines have long used structured delegation frameworks to reduce the cost of ambiguity. The American Nurses Association’s Five Rights of Delegation – developed to ensure clinical tasks are assigned safely and clearly – translates directly to agentic deployment.
1. The Right Task
Not every process should be delegated to an agent. The useful filter is a combination of criticality and reversibility. A high-volume, low-stakes, easily-reversed task, such as processing invoices, triaging inbound enquiries, summarising research, is ideal for autonomous agents. A multi-million dollar trade execution, or a communication that cannot be unsent, is not.
If the cost of a wrong decision is hard to recover from, a human needs to be in the loop before the agent acts.
2. The Right Circumstance
Agents perform well inside defined operational envelopes. They struggle when the environment shifts in ways their instructions didn’t anticipate, like a regulatory change, a market dislocation, an internal policy update that happened after deployment.
Leaders must define not just what the agent should do, but the conditions under which it should pause and ask. Adaptive coordination mechanisms – built-in signals that prompt an agent to recognise when it’s outside its operating parameters – are what separate robust deployments from brittle ones.
3. The Right Agent
Matching task to capability is obvious in principle and routinely ignored in practice. A CEO wouldn’t ask a marketing analyst to negotiate an M&A term sheet. The same logic applies to agents: one optimised for structured data retrieval is not the right choice for open-ended strategic reasoning.
Certifiable capability – not general AI quality – is the relevant measure. The question isn’t “is this a good model?” It’s “has this agent demonstrated it can handle this specific class of task reliably?”
4. The Right Direction and Communication
This is where most agentic deployments fail. Vague objectives produce creative interpretations. Creative interpretations produce outcomes that surprise everyone.
Effective direction requires four components:
Clear objectives: the “why” behind the task, not just the “what.” An agent that understands the purpose of an instruction can navigate edge cases. One that only has the instruction cannot.
Specified outputs: the exact format, quality threshold, and acceptance criteria for what success looks like. “A good summary” is not a specification. “A 200-word executive summary highlighting the three material risks, in plain language, suitable for a board audience” is.
Tool guidance: which systems, datasets, and APIs the agent is authorised to use, and which are out of scope.
Boundaries: explicit no-go zones. Sensitive data the agent should not touch. Actions it should never take without escalation.
5. The Right Supervision and Evaluation
Delegation is not set-and-forget. It is a transfer of responsibility, not an abdication of it.
Supervision at scale requires what the DeepMind paper calls structural transparency: the agent’s reasoning process must be auditable, and its outcomes must be verifiable against the original intent. That means logging, not just monitoring. It means being able to reconstruct why an agent made a decision, not just what it decided.
For senior leaders, this has a practical implication: if your agentic infrastructure can’t answer those questions, you’re operating on faith, not oversight.
Managing What You Can’t Watch Directly
In human organisations, span of control describes the number of direct reports a manager can effectively oversee. For most roles, the research converges on somewhere between five and nine. Beyond that, performance degrades.
Deploying hundreds or thousands of agents without addressing span of control is a governance failure waiting to happen. The DeepMind framework addresses this through what it describes as a need for agents to recognise when a technically permissible request is contextually ambiguous enough to warrant stepping outside their zone of indifference – pausing, challenging the delegator, or requesting human verification before acting.
In practice, this means building orchestrator agents: specialised nodes whose job is not to execute tasks, but to manage other agents. They monitor sub-agent performance, calibrate trust based on outcome data, flag anomalies, and escalate to human supervisors when the situation exceeds the system’s defined parameters.
This is the organisational design question that determines whether your agentic infrastructure scales safely or just scales.
What This Means for Governance Now
The organisations that will navigate the agentic transition well are already building governance reflexes now – practising the discipline of clear intent, explicit boundaries, and verifiable outcomes on the systems they already have.
The research is consistent on one point: systems and people respond to the incentives actually present, not the ones intended. An agent given vague instructions and broad permissions will optimise for something. The question is whether it’s optimising for what you actually want.
That answer starts with the quality of the delegation.
References
- Tomašev, N., Franklin, M., & Osindero, S. (2026). Intelligent AI Delegation. arXiv:2602.11865. https://arxiv.org/abs/2602.11865
- American Nurses Association. (2023). The Five Rights of Delegation. https://www.nursingworld.org/