# Agent RACI Worksheet

> **From The Agentic TPM, Chapter 5.** Traditional RACI was designed for human work. When an agent produces the work, the framework needs adapting.
>
> **Core principle:** Agents can be Responsible. Only humans can be Accountable.

---

## Program metadata

| Field | Value |
|---|---|
| Program name | |
| Workflow being mapped | |
| Date | |

---

## Roles in agentic RACI

| Role | Who holds it | What it means |
|---|---|---|
| **Responsible (R)** | The agent | Does the work — generates, executes, produces output within bounded autonomy |
| **Accountable (A)** | Named human | Owns the outcome. Can be asked "why did this happen?" and must answer |
| **Validator (V)** | Meta-agent or named reviewer | New role. Checks the agent's output before it propagates |
| **Consulted (C)** | SMEs, domain owners | Same as classical RACI, sharper scope |
| **Informed (I)** | Stakeholders | Same as classical RACI, faster cadence, AI-generated artifacts |

---

## Workflow mapping

For each activity in your workflow, name the R, A, V, C, and I.

### Activity: __________________________

| Role | Who / What |
|---|---|
| R (Responsible) — the agent | |
| A (Accountable) — named human | |
| V (Validator) — meta-agent or reviewer | |
| C (Consulted) — SMEs | |
| I (Informed) — stakeholders | |

### Activity: __________________________

| Role | Who / What |
|---|---|
| R (Responsible) — the agent | |
| A (Accountable) — named human | |
| V (Validator) — meta-agent or reviewer | |
| C (Consulted) — SMEs | |
| I (Informed) — stakeholders | |

### Activity: __________________________

| Role | Who / What |
|---|---|
| R (Responsible) — the agent | |
| A (Accountable) — named human | |
| V (Validator) — meta-agent or reviewer | |
| C (Consulted) — SMEs | |
| I (Informed) — stakeholders | |

*(Copy this block for every activity in the workflow.)*

---

## The single most important rule

> Every workflow that involves an agent has a **named human Accountable**. Not "the team." Not "the platform." A **person**, with a name, who answers when something goes wrong.
>
> If you cannot name that person for any agentic workflow in your program, you do not have a program — you have a liability.

---

## Review questions

- [ ] Every agentic activity has a named human Accountable (not a team, not a role — a person).
- [ ] Every activity with customer impact has a Validator step before propagation.
- [ ] SMEs (Consulted) are scoped to judgment calls only humans can make.
- [ ] Informed stakeholders receive AI-generated updates at appropriate cadence.
- [ ] No activity relies on the agent being Accountable.

---

## Sign-off

- **Owner:**
- **Date:**
- **Next review:**