What to delegate to an agent — and what not to

A mid-size Brazilian facilities-engineering company (figures in Brazilian reais), 310 people, 40 job sites running at once. The operations manager gets roughly 90 messages a day that call for some reply: a material release request, a deadline question, a delay notice, an overtime approval, a conflict between two crews over who gets the crane on Tuesday. He answers almost all of them, and answers well. By the end of the day, there's no hour left for the things only he can do.

The temptation is to delegate all of it to an agent. The useful rule splits the 90 messages into two very different piles, and the split isn't by difficulty.

01The scarce resource is attention #

Herbert Simon wrote, in 'Designing Organizations for an Information-Rich World' (1971), the formulation that became the foundation of the attention economy: a wealth of information creates a poverty of attention. If information is abundant, what it consumes is the attention of whoever receives it — and attention, unlike information, is finite.

The design consequence, for Simon, is that a good information system isn't the one that delivers more: it's the one that filters. Applied to agents, it changes the question. It isn't 'what can the technology do', but 'which part of the manager's attention is being consumed by work that doesn't require human attention'.

Andy Grove, in High Output Management (1983), offers the complementary criterion: managerial leverage. A manager's output is the output of the organization under their influence, and high-leverage activities are the ones that affect many people or future decisions. Delegating well, for Grove, is choosing where your hour pays off most — not handing off the boring task.

02The rule: repetitive, verifiable, reversible #

Three tests, all of them have to pass. If one fails, the task doesn't go entirely to the agent — at most, its preparation does.

1. Repetitive. It happens often enough for a pattern to exist. A task that comes up three times a year doesn't justify the cost of scope, testing and maintenance.
2. Verifiable. There's an objective way to know if the result is correct, and someone can check it by sampling without redoing the whole task.
3. Reversible. If it goes wrong, it can be undone at low cost and within a known timeframe. It's the operational version of the type-1/type-2 decision distinction Jeff Bezos described in his 2015 shareholder letter: one-way doors demand greater care.

Of the engineering manager's 90 messages, the three tests classified about 60 as delegable: material release within a pre-approved limit, deadline replies based on the current schedule, chasing open items, consolidating job-site status. All repetitive, verifiable and reversible within 24 hours.

03What doesn't get delegated #

• Judgment about people. Evaluation, promotion, warnings, assignments that change someone's career. It fails all three tests at once, and the cost of error is asymmetric: it falls entirely on the person with less power.
• Strategic priority. Which job site slips if a crew is short is a decision about what the company is willing to lose. The agent can prepare the comparison; the choice belongs to the manager.
• Irreversible decisions. Canceling a contract, telling a customer about a serious failure, committing to a deadline with a penalty attached. A one-way door.
• Conflict between people. Who gets the crane on Tuesday looks like logistics and is really politics. A correct automated reply can cost more than a five-minute conversation.
• Anything whose error only surfaces much later. If verification takes 40 days, the verifiable test failed, even if the repetitive test passed.

There's an in-between category that often pays off more than either list: the agent prepares and the human decides. It gathers the history, lays out the options and their known consequences, and stops there. The manager decides in two minutes what used to take twenty of digging. That's where Grove's leverage and Simon's filter meet.

Want to see how this looks inside a real operation? Explore the platform.

One warning before the math: the rule shifts over time, in both directions. A task that's fragile today can become safe once the process behind it stabilizes and the history is clean; and a task that looks safe today can go back to a human after a costly mistake reveals an edge case nobody had anticipated. That's why the decision to delegate needs a review date, and doesn't hold forever. Review what's delegated every quarter, correction rate in hand: where it climbs, the agent's scope shrinks; where it stays near zero for two straight quarters, it can safely expand.

04The math on attention #

Stated assumptions, at the engineering company's scale; redo it with your own numbers:

Two and a half hours a day don't automatically turn into two and a half hours of strategic work — part of it gets consumed by interruptions again, as described in the manager's shattered day. But the difference between a manager who decides 30 things a day and one who decides 90 shows up in the quality of the 30.

05How to apply the rule this week #

The last item is the check the NIST AI Risk Management Framework 1.0, from 2023, calls continuous management: delegated scope isn't a permanent decision. If you missed seeing something, bring it back; if you never did, the scope can grow.

At Relevanti, what the agent decides, what it prepares and what it hands back to a person is declared by scope — see the platform, the breakdown by department under solutions, or bring your three-day list to a conversation.

Simon warned in 1971 that abundant information impoverishes attention, and AI increased the abundance. The delegation rule exists so the technology gives attention back instead of consuming more of it.