# The Delegation Economics Worksheet

> A free one-page worksheet: decide whether a task is worth handing to an AI agent once brief and rework time count, then give AI spend a budget line.

Use part one before handing a recurring or high-stakes task to an AI agent. It weighs the saving against the time on either side of the agent's work: refining the brief first, and reworking the output afterwards. Use part two when agents run on the project at any volume, so their spend has an estimate, a reserve and an alert before the cap rather than an explanation after it.

## The frame

Part one, once per task:

```
Task: [the piece of work]
Type: [High volume, low stakes / Low volume, high stakes / In between]
Time to do it yourself: [minutes or hours]
Time to refine the brief: [context, scope, examples, before it runs]
Agent cost per run: [estimate, or "unknown, check after first run"]
Chance of rework: [Low / Medium / High, and why]
Time to rework if needed: [including finding what's wrong]
Saving if it goes well: [time to do it yourself, less brief time and agent cost]
Saving after rework risk: [does it still hold?]
Decision: [Delegate / Refine the brief further first / Keep]
```

Part two, once per workstream or agent:

```
Workstream or agent: [name]
Estimate for the phase: [spend, sized to the work, not to the subscription]
Reserve: [sized to how much this work varies]
Alert: [the level that triggers it, below the cap, and who receives it]
Default model tier: [the cheapest tier that does the job; when a higher one is allowed]
Change trigger: [a new agent, wider scope or more autonomy reopens the estimate]
Cost per finished deliverable: [track this, not only spend per month]
Owner: [who approves a raise]
```

## Have an AI help with part one

```
I'm deciding whether to hand a task to an AI agent. Help me fill in this worksheet. Ask me one question at a time for anything you can't infer, and don't invent times or costs.

Fields: Task, Type, Time to do it yourself, Time to refine the brief, Agent cost per run, Chance of rework, Time to rework if needed, Saving if it goes well, Saving after rework risk, Decision.

When the fields are filled, tell me which single field the decision is most sensitive to, and what would change the decision.

Task: [describe it]
```

## Worked examples

All figures below are illustrative. Use your own.

### Weekly status pack

```
Task: Draft the weekly status pack from the tracker and meeting notes
Type: High volume, low stakes
Time to do it yourself: 90 minutes
Time to refine the brief: 60 minutes once, then 5 minutes a week
Agent cost per run: Small
Chance of rework: Low, the format is fixed and the inputs are consistent
Time to rework if needed: 15 minutes
Saving if it goes well: About 70 minutes a week after the first week
Saving after rework risk: Holds
Decision: Delegate, with "AI drafts, person approves"
```

### Stakeholder risk assessment for a phase gate

```
Task: Assess stakeholder risks to the go decision at the phase two gate
Type: Low volume, high stakes
Time to do it yourself: 4 hours
Time to refine the brief: 90 minutes: decision criteria, stakeholder history, what's out of scope
Agent cost per run: Higher, long context and several passes
Chance of rework: High without the brief work, Medium with it
Time to rework if needed: Hours, plus a sponsor conversation if it's acted on wrong
Saving if it goes well: About 2 hours
Saving after rework risk: Holds only with the full brief
Decision: Refine the brief further first, then delegate with a full review
```

### Budget line for a programme's reporting agents

```
Workstream or agent: Reporting agents (status drafting, RAID extraction, weekly digest)
Estimate for the phase: Set from the first month's actual use, not the subscription fee
Reserve: Larger than the programme's usual percentage, because agent use grows as people find new jobs for it
Alert: At 70 per cent of the monthly estimate, to the PMO lead
Default model tier: Mid-tier for drafting and extraction; frontier only for the digest's risk narrative
Change trigger: Any new reporting agent, or an agent given write access
Cost per finished deliverable: Tracked per status pack and per digest
Owner: Programme manager
```

## Notes on the parts

**Time to refine the brief** is a cost you choose, to bring the chance of rework down. Skipping it moves the cost to the end of the task, where it is larger and harder to plan for.

**Type** points to the usual answer. High-volume, low-stakes work almost always passes. Well-defined high-stakes work is often a strong candidate too, once the brief carries the decision criteria.

**Reserve and Alert** treat AI spend like any other volatile cost: estimate it, price the variation in, and act before the cap.

**Cost per finished deliverable** is the figure that tells you whether a cheaper model or a price change helped.

## Check it before you decide

- Did you count the brief time and the rework time, not only the agent's run?
- For a high-stakes task, is the rework cost measured in consequences as well as hours?
- Does every agent with meaningful spend have an alert set below its cap?
- Is there a named owner for raising the estimate?

## When it does not apply

Quick one-off prompts, where the brief is a sentence and checking takes a minute, don't need part one. Part two can wait until agents run on a schedule or several people use them on the project.

## What this feeds

A Delegate decision becomes an entry in the [AI Delegation Register](/toolkit/delegation-register/). Part two belongs in the project's cost plan, with any raise handled like other budget changes. The budget-line reasoning is in [AI Token Budget Controls](/articles/ai-token-budget-controls/).

---

From ProjectorPM: https://projectorpm.xyz/toolkit/delegation-economics/