---
url: /handbook/roadmap/just-in-time-task-capture.md
description: >-
  Planned: when an employee requests an agent and no adequate Task Card exists,
  capture just that one task, then govern the request as usual.
---

# Just-in-time task capture

Just-in-time task capture is the bottom-up way into DutyGraph. An agent request starts the capture of the one task it needs, instead of waiting for a full company map. See [Three ways in, one model](../start-here/three-ways-in.md).

::: warning
**Planned.** This page describes proposed behavior. Only the request tools in steps 1 and 2 are built today.
:::

## The problem it solves

A governed agent needs a confirmed Task Card. In a large company, most employees do not have one yet. Mapping a whole department first can take weeks. The request should not wait for that, and it should not skip the evidence either.

## How it would work

### 1. The employee requests an agent

The employee asks for an agent in DutyGraph or from their own AI chat over the Model Context Protocol (MCP). **Built:** the request tools exist today. See [Request from your AI](../governance/request-from-your-ai.md).

### 2. DutyGraph looks for an adequate Task Card

DutyGraph checks for a confirmed Task Card that covers the work. **Built in part:** today a request must name Task Cards that are already set up. If DutyGraph cannot choose one, it asks for the Task Card ids. Judging whether a card is adequate is Planned.

### 3. DutyGraph opens a short capture

If no adequate card exists, DutyGraph opens a capture of that single task. It does not start a department-wide project.

### 4. The employee describes the task

The employee answers by voice or text, on a phone: the task, its steps, the systems it uses, its rules and its exceptions.

### 5. DutyGraph proposes a Task Card and its requirements

DutyGraph drafts one Task Card. It also proposes the business capabilities each step needs. Both are proposals until people confirm them.

### 6. The manager confirms the work

The manager confirms the work is real and belongs to this person. This confirms the work exists. It is not approval of any access.

### 7. An application owner validates the permissions

An authorized application owner, or an IAM owner, validates which technical permissions in their system the work needs. This is evidence about the system, not a grant.

### 8. The evaluator calculates authority

The evaluator checks each requested scope. Each one is allowed, narrowed, denied or escalated for a human decision. In the target design, a scope must be needed by the validated requirements, held by the owner and allowed by every rule.

### 9. Normal human authorization proceeds

The request enters the normal [approval chain](../governance/approval-chain.md). Each decision goes into the [ledger](../governance/ledger.md). The company's IAM provisions the agent if it is approved.

## Confirmation is not approval

Three different people give three different things.

| Who | What they give | What it is not |
| --- | --- | --- |
| The manager | Confirmation that the business work is real | Not approval of privileged access |
| The application or IAM owner | Validation of the technical permissions the work needs | Not a grant |
| The approval chain | The authorization decision, recorded in the ledger | Not provisioning; the company's IAM does that |

**No access is issued because someone submitted a task description.** A submitted description is a proposal until the manager and the application owner confirm it, and access still needs the approval chain.

## Short, and proportionate to risk

The capture should stay short. The design goal is about five minutes for the employee and about two minutes for the manager. These are design targets, not measured results.

The depth of capture should match the risk. A read-only agent on low-risk data needs less evidence than an agent that can change financial records. A higher-risk request would ask for more steps, more detail on exceptions and application-owner validation of each permission.

## What stays the same

* The agent's scopes stay within the owner's holdings. That is a ceiling, not a justification. See [Task Cards](../discovery/task-cards.md#holdings-are-a-ceiling-not-a-justification).
* The approval chain, its three decisions and the ledger do not change.
* The person who does the work describes it. People confirm it. DutyGraph does not write the Task Card on its own.

## What it builds over time

Each request adds one confirmed Task Card to the same Company Repo that top-down discovery uses. Over time, a department's most important work is mapped one task at a time, by the people who do it.

::: info
**Status.** The employee request tools, the evaluator's ceiling and rule checks, the approval chain and the ledger are **Built**. Limiting scopes to what the Task Cards need is **Partial**. The identity systems in the samples are **Simulated**. The single-task capture, manager confirmation, application-owner validation and risk-proportionate capture are **Planned**.
:::
