# 9.10 One deliverable, from question to plan (/managing-ai-workers/work-that-runs-without-you/one-deliverable)

---
type: Document
title: "9.10 One deliverable, from question to plan"
description: "How to carry one report or plan through four stages, from the question to a decision, and why each stage gets two checks."
status: stable
order: 209.87
ksor:
  owner: team:panaversity
  audience: [ public ]
  approval:
    by: process:panaversity
    at: 2026-11-18T20:32:55Z
chapter: "09"
part: II
expert_status: required
concepts: [ "9.10" ]
last_verified: 2026-11-19
generated:
  at: 2026-11-18T20:32:55Z
  by: process:panaversity
trust_tier: unverified
build_id: sha256:97110dcafc31f6bc3cb7488277330384cf5ceafbae273b570c80ac3362b0d2ec
dirty: true
ksor_version: 0.0.60
---

**In everyday life.** A family plans a move. They look up schools, add up the costs and write a plan with dates. The school list is accurate, and the sums are right. But the list ranked schools for older children, and their children are six and eight.

**What you will learn.**

- Which four stages carry a report or plan from the question to a decision?
- Every stage passed its own check. Can the plan still be wrong?
- What makes a plan one that someone can act on?

You practice it in Task 13 of the lab.

**Why it matters.** A plan built on an unchecked source or a guessed number looks as finished as a good one. So does a plan built on right answers to a different question.

Most of this chapter is about work that repeats. Some work comes once. For that work, [9.1](is-this-an-agent-problem.md) said to build nothing, and to handle it in a conversation. But a large question does not fit in one request. It needs sources, numbers and a decision. A **deliverable** is a report or plan that someone decides from.

### What are the four stages?

A deliverable has four stages, and each one feeds the next.

| Stage | The question it answers | Use | Who decides |
| --- | --- | --- | --- |
| 1. Requirements | Who decides what, by when, and what does done look like? | Chat. Let the AI interview you ([5.4](../the-four-part-brief/decompose.md)), then write the "Done when" sentence ([5.1](../the-four-part-brief/four-parts.md)) | Human-retained. The person who asked agrees the question |
| 2. Research, with cited sources | What do good sources say? | Research, for a question that needs many sources ([2.4](../../ai-worker-paradigm/what-is-an-ai-worker/surface-and-model.md)) | Human-reviewed. The AI searches, and you decide which claims hold |
| 3. Analysis, in code | What do our own numbers say? | Code execution on the file you upload | Human-reviewed. The AI computes, and you decide which numbers hold |
| 4. Synthesis, into a plan | What should we do? | An artifact, such as a document or slides ([5.7](../the-four-part-brief/output-format.md)) | Human-reviewed, then human-retained. The AI drafts, you check, and the person who asked decides |

The last column uses the labels of [9.2](fit-the-workflow.md). No stage is AI-executed: a person checks each one before the next stage uses it. Put a checkpoint where a person's decision changes what follows ([5.4](../the-four-part-brief/decompose.md)). The AI stops there, and waits for you.

### Every stage passed its check. Can the plan still be wrong?

Yes. You already have a check for each stage. For a source, open it, and see whether it says what the report claims ([6.5](../the-review-contract/citation-and-abstention.md)). For a number, recompute it from the source file ([6.4](../the-review-contract/decision-numbers.md)). Both checks ask one thing: is it right?

A stage can be right and still answer a different question. A sound survey can count something else. A correct total can mean less than it seems. A complete plan can leave out the one decision that was asked for. As [6.2](../the-review-contract/good-and-plausible.md) said, a good output matches the outcome you asked for in the brief.

So the first stage does more than start the job. It writes the question down, and the person who decides agrees it. After that, check every stage twice.

| Stage | Is it right? | Does it answer the question? |
| --- | --- | --- |
| Requirements | The person who decides agrees the question, in writing | This stage writes the question |
| Research | The source behind each claim the plan rests on says what the report claims | What did the source count, and whom did it ask? |
| Analysis | Each number the decision rests on recomputes from the source file | What is it a number of: which year, which currency, on what assumption? |
| Synthesis | Every action has an owner, a date, a cost and a risk | Does it give the decision that was asked for, first? |

When a stage cannot answer the question, do not fill the gap. Write "not known", and keep those words in everything that follows.

### Monday at Brightline

Maria is the office manager at Brightline, the book's example company. Dave Kowalski, the controller, approves its payments. On Monday, November 30, he had a question for her.

- **9:10 a.m.** Dave asked whether Brightline should ask six of its suppliers for 2 percent off for paying within 10 days, from January. He wanted a yes or no for each of four supplier groups: packaging, freight, office and cleaning, and pallets. He also asked what it would save, and how much sooner Brightline would pay. He needed it before Thursday's budget meeting.
- **9:30 a.m. Requirements.** Maria let the AI interview her about the ask. Then she sent Dave four lines: "You decide. A yes or no for each of the four supplier groups. By Thursday, December 3. Done when one page gives your decision first, then the saving and how much sooner we would pay." Dave replied, "Yes."
- **10:15 a.m. Research.** A new summary said most suppliers agree, as the report in [6.5](../the-review-contract/citation-and-abstention.md) had. It named its sources. Maria opened the source behind "most." It was a sound survey, recent and large. So it passed the first check. Then she asked what it had counted. It had counted suppliers that already offer a discount to some customer. Dave's question was about suppliers that are asked for one. Under the four lines, Maria wrote, "How many will agree when asked: not known."
- **11:00 a.m. Analysis.** The AI ran code on the 36 invoices Brightline paid those suppliers in 2025. It gave each group's total, 2 percent of it, and the days Brightline had taken to pay. Maria added up one group herself, from the file, and her total matched. Then she named the number. Two percent of the year was $1,916.78, with CAD 840.21 kept apart, because one supplier is paid in Canadian dollars. That is the most the discount could have saved, if every supplier had agreed and every invoice had been paid within 10 days. It is not what Brightline will save.
- **2:00 p.m. Synthesis.** The AI drafted the plan as one page, and Maria checked it before it went to Dave. It started with his decision: a yes or no for each group. Each action had an owner, a date, a cost and a risk. It ended with what could not be verified: how many suppliers will agree. So it counted no saving as certain.

On Thursday, December 3, Dave decided from that page.

The survey was real, and the total was correct. Each passed its first check. Only the second check showed what each one could not say.

### What makes a plan actionable?

A plan is **actionable** when someone can start each action tomorrow without asking a question. Start with the decision the reader must make. Give each action an owner, a date, a cost and a risk. End with what could not be verified. The plan does not wait for those answers. It names them, so the reader knows what the decision rests on. What is not known often gives the plan its first action: test it on a small scale before you count on it.

### The price

Four stages and two checks took Maria from morning until two in the afternoon. For an answer you will only read, one request and one check are enough. Use the stages when someone will decide from the result.

When the same deliverable comes back every month, it has become recurring work, and the gates of [9.1](is-this-an-agent-problem.md) apply. [Case 4, Nisha's Month-End](../../claude-and-chatgpt-2/case-4-nishas-month-end/overview.md), shows a month-end report built from a project, a skill, code and artifacts, and fixed in the setup, not the result.

> **Outside finance: the training plan.** Brightline's warehouse manager needs a training plan for new forklift rules. Requirements: who must be trained, by when, and what counts as done, agreed with the safety officer. Research: what the rules require and how other warehouses train, with each source opened and checked. Analysis: from the staff list, code counts who needs which course, and how many sessions fit in each shift. The manager checks the count against the list, and it matches. Then comes the second check. The list holds everyone in the warehouse, and the rules cover only the people who drive a forklift. The count is right, and it answers a different question. So the code counts again, for drivers only. Synthesis: a plan with the safety officer's decision first, then each session with an owner, a date, a cost and a risk.
