Assigning Work:
Task Design for AI Agents

The new management skill

Managing AI employees requires a skill that managing human employees doesn't: writing unambiguous task descriptions.

When you tell a human employee "clean up this spreadsheet," they ask clarifying questions. "Which columns? What counts as a duplicate? Should I keep the headers?" A human employee fills in the gaps with judgment, experience, and office small talk.

An AI employee doesn't ask questions. It interprets the description literally and starts working. "Clean up this spreadsheet" might produce a beautifully formatted document with all your data deleted (it was "cleaning up" after all). Or it might do nothing because "clean up" is too vague to act on.

The manager's job is to write descriptions that leave no room for misinterpretation. This isn't a limitation of AI — it's a discipline that produces better outcomes than the vague delegation most managers practice with human employees too.

The four rules of agent-shaped tasks

Every successful agent task we've seen follows four rules. Every failure violates at least one.

Rule 1: Clear inputs

The agent needs to know exactly what data it's working with. A file path, a URL list, a PR number, a date range. If the input is "figure out what to work on," the agent will spend 80% of its time and budget exploring instead of executing.

Rule 2: Verifiable outputs

Can you tell in 5 seconds whether the agent succeeded? If yes, the output is verifiable. "Good report" requires reading and judging. "Report with 30 company sections, each with pricing and features" requires counting sections.

Rule 3: Bounded scope

The task has a clear end. A number of items, a stopping condition, a maximum time. Without bounds, the agent either loops forever or decides on its own when to stop — which may be too early or too late.

Rule 4: Safe to fail

If the agent gets it wrong, the damage is containable. A bad draft is deleted. A wrong test result is noticed. A missed invoice is caught downstream. Tasks with irreversible consequences (production deploys, financial transactions, customer communications) need a human approval gate between the agent's output and the action.

The trust ladder: supervised to autonomous

Every new agent task starts supervised and graduates to autonomous. This matches how you'd manage a new human hire:

1. Week 1: Watch everything. The agent runs. You review every output. You find 2–3 issues in the task description and fix them. Trust: zero.
2. Week 2–3: Spot-check. The agent has run 15 times without error. You review 3 out of 10 outputs. The task description is stable. Trust: building.
3. Month 2: Alerts only. The agent runs 30+ times. You review only when it flags an exception or when the weekly summary shows anomalies. Trust: established.
4. Month 3+: Full autonomy. The agent runs on schedule, posts results, handles 98% of cases correctly. You look at the monthly performance summary. Trust: earned.

The graduation speed depends on the task's risk level. A competitive intel sweep can reach full autonomy in 2 weeks (low risk, easy to verify). A financial reconciliation might stay at spot-check for months (higher risk, needs accuracy).

The 20 most common agent tasks

Across every company deploying AI agents, these 20 tasks appear most frequently. All follow the four rules. All are referenced in guides throughout this series.

Your first task: pick from this list

Don't invent a new use case. Pick the task from the table above that resonates most with your team. Then:

1. Write the task description following the four rules. Be embarrassingly specific.
2. Run it once manually (prompt-driven). Review the output.
3. Refine the description based on what went wrong.
4. Run it 5 more times. Consistent? Good. Inconsistent? Refine again.
5. Wire it to a trigger (webhook, cron, Slack command).
6. Graduate it through the trust ladder. Supervised → spot-check → alerts → autonomous.

The first task takes an afternoon. The second takes an hour. By the fifth, you're writing task descriptions in 10 minutes and the agent runs them perfectly on the first try. The skill compounds fast.

What's next