A new kind of employee is arriving. It is software that can plan steps, use tools and complete tasks by itself, known as an AI agent. When it works, delivery speeds up and the cost per task falls. When it fails, those savings come back as supervision, rework and incident recovery.
Google DeepMind’s paper, Intelligent AI Delegation, argues that the bottleneck is managing these agents. Many ways of handing work to them rely on rules of thumb, which struggle when conditions change or something fails unexpectedly. In business terms, the workflow looks automated until the first unusual case, when people become the safety net and costs climb.
The paper’s answer is intelligent delegation, a structured way to hand off work with clear limits. It treats delegation as a series of decisions about who does what, in which each handoff explicitly transfers authority, responsibility and accountability. It also calls for defined roles and boundaries, clear intent and ways to build trust between the parties.
Authority comes first, because deciding what an agent can do without approval sets both speed and risk. Freedom to act without asking is valuable only within limits that match the risk of the task. When authority is vague, teams pay later through costly cleanup, slower decisions and slower recovery.
Responsibility and accountability are where many rollouts quietly lose margin. Responsibility means doing the work, and accountability means answering for the result. When a task passes through several AI systems and people, clear accountability keeps work from falling through the cracks. It also cuts the time spent untangling blame after an incident.
In the paper, trust comes down to checks that build confidence over time. Those checks tie into monitoring, so delegation can widen when performance is stable and tighten when it drifts.
For a business, that means treating agents like colleagues in how work is assigned, reviewed and owned, without pretending they are people. Each agent gets authority tiers covering what it can do alone, what needs approval and what is blocked, along with written role boundaries and escalation rules. Every critical workflow keeps an audit trail, a record of what was done and why.
Without that structure in place before an agent’s freedom to act expands, an organisation pays for reliability through extra headcount, process friction and slower time to market.
