How to Delegate to People and to Agents

Sarah Hoban suggests starting with a list of everything you are working on or plan to work on, including the small recurring tasks that "only take a few hours". Then sort every item into three groups: ditch, delegate or do yourself. Talking the list through with someone outside your role helps, because most people rationalize keeping everything.

Some work should stay with you. HBS Online names performance reviews and personnel matters as examples. Beyond that, delegate when someone else could do the work better, when it would help their growth, or when it frees you for higher-value work.

When the candidate delegate might be an AI agent, add a few more questions. A research paper on intelligent AI delegation proposes axes for judging any delegated task. Four of them are especially practical:

The paper notes that tasks that are easy to verify allow lighter oversight, while tasks that are hard to verify need either a highly trusted delegate or expensive checking. Irreversible actions with real-world side effects, such as executing a trade, deleting a database or sending an external email, call for stricter controls than reversible ones like drafting an email.

This gives a simple rule for agents: delegate freely where the work is reversible and easy to check, delegate with approval steps where it is not, and keep the irreversible, hard-to-verify decisions with a person.

The single most common delegation failure is that the two sides assume different levels of authority. Full Focus tells the story of a manager who thought he had asked a team member to research and report, while the team member believed they had been asked to decide and act. Neither was wrong. The level had never been stated.

Full Focus describes five levels:

1. Do as I say. Follow the instructions exactly. The research and the decision are already done.
2. Research and report. Gather information and bring it back. I decide.
3. Research and recommend. Lay out the options with pros and cons and recommend one. I approve.
4. Decide and inform. Make the decision and tell me what you did.
5. Act independently. Decide and act with no need to report back.

Humanizing Work uses Jurgen Appelo's seven-level version (tell, sell, consult, agree, advise, inquire and delegate) and adds an important warning. There is a line between the levels where you remain the final decision maker and the levels where you commit not to be. If you hand over a decision and then overrule it when you dislike the result, you are practicing what Humanizing Work calls "delegation theater", which breaks trust and breeds cynicism. If you might need to overrule, stay at a level where you keep the decision.

The same scale maps cleanly onto AI agents:

Say the level out loud, or write it in the brief, every time.

HBS Online quotes Professor Kevin Sharer on the core requirement: "You've got to have real clarity of objective", including alignment on what good looks like, by when, and how accomplishment will be measured. It also stresses providing the training, resources and authority the person needs, and resisting the urge to dictate every step.

SHRM's "Delegator's Dozen" checklist adds the practical details: define the result you want, choose the person deliberately, set the deadline and progress report dates, define the authority level and tell others who need to know, and agree on the control process.

A brief that covers these points fits on half a page and works for a colleague or an agent:

OUTCOME: what done looks like, and how we will measure it
WHY: how this connects to the wider goal
CONTEXT: background, constraints, links, what has been tried
AUTHORITY LEVEL: research and report / recommend / decide and inform / act
LIMITS: budget, systems you may touch, actions that need approval first
CHECKPOINTS: when we look at progress, and what you show me then
DEADLINE: final date
ESCALATE IF: the situations where you stop and ask

Hoban highlights the part managers most often skip. In the handoff conversation, cover what the task is, how it aligns with project and company goals, why you picked this person, how involved you will be and the details of execution. Managers tend to neglect the "why" in favor of the "what" and "how".

Most delegation principles carry over, but four differences matter.

Agents act on what you write, not what you meant. A colleague who is unsure will usually ask. An agent will usually proceed. The brief therefore needs explicit limits and an explicit "escalate if" list. The arXiv paper describes a related risk it calls the zone of indifference: instructions that are accepted and executed without critical scrutiny. As delegation chains grow longer, from person to agent to another agent, small mismatches in intent can pass through unchallenged.

Trust has to be calibrated, not assumed. The paper defines trust calibration as matching the trust placed in a delegate to its actual capability. Frank Goortani draws the parallel with situational leadership: the autonomy granted to an agent should match its demonstrated competence and reliability, just as a good manager adjusts oversight to a team member's readiness.

Plausible is not the same as correct. Data & Society warns that some agent divergences appear as plausible outputs produced through wrong methods or shortcuts that preserve the look of success. Reviewing only the final output can miss this. For important work, ask the agent to show its sources, steps or test results, not only its answer.

Accountability stays with people. Data & Society describes a "moral crumple zone", where the person nearest the system is blamed for a failure even when they lacked the practical ability to see it or redirect it. If you assign someone to oversee an agent, give them the visibility and authority to actually intervene.

Teams that build agents into their work, including those using Quanta Intelligence agents, get the best results when every agent has a written brief with the same fields a person would get.

HBS Online quotes Sharer again: delegation "isn't fire and forget". You need some way to see how things are going without getting in the way. With people, that usually means agreed check-ins and progress reports. With agents, it means logs, approval steps and alerts.

Data & Society argues that effective oversight requires four things: knowledge, observation, control and intervention. A pause button only matters if the person can recognize the need to pause, see the relevant part of the work and redirect what happens next. It also names the practical tension. If you watch every step, delegation loses its value. If you see too little, oversight becomes nominal. The skill is selective attention: deciding in advance which uncertainties you can tolerate.

The arXiv paper offers a useful distinction for planning this: monitoring can be continuous, periodic or event-triggered. A practical oversight plan:

Hoban adds advice for the human side. The person you pick will rarely be fully ready, will struggle at first and will put their own spin on the task. If you hover, ready to step in at the first misplaced comma, they will stop trying to improve. Set regular check-ins, then gradually step back.

Delegation is a cycle, not a single act. HBS Online recommends giving feedback when the task is done, both constructive and positive, and asking the delegate whether your instructions were clear. It also recommends giving credit publicly to the people who did the work. SHRM's checklist ends with identifying lessons learned on both sides, because the person doing the work often finds a better way to do it.

For agents, closing the loop means updating the brief. When an agent's output misses, change the instruction, the limits or the examples, then check the next run. Over time, record which tasks an agent handles reliably and which still need approval.

Then widen the scope deliberately. When a person or an agent consistently succeeds at one level of authority, move them up one level on the next similar task and say so explicitly. When something goes wrong, step back one level for a while instead of taking the work back entirely. This keeps trust calibrated in both directions and builds the team's capacity over time.