RACI Matrix: Meaning, Example, and Where AI Agents Fit

In project management, RACI means Responsible, Accountable, Consulted and Informed. A RACI matrix is a table with tasks or decisions down the side and people or roles across the top, and a letter in each cell saying how that person is involved. Responsible is whoever does the work. Accountable is the one person who answers for it and approves it as finished. Consulted are the people whose advice is sought before it is done, and Informed are the people told once it is. The rule that makes it work is simple: every row has exactly one A.

What R, A, C and I mean

Usage varies a little between organisations, and there is no single standard that owns the term. The definitions in NHS Education for Scotland’s guidance↗ (opens in a new tab) are as clear and neutral as any, and match how most teams use the letters:

• Responsible: the person or people who complete the task or deliverable. There can be more than one.
• Accountable: the person with final authority or approval for the task’s completion. There is only one.
• Consulted: an adviser to the task, usually providing subject expertise or guidance. They are asked before the work is done or decided.
• Informed: people kept up to date on progress on the task. They are told, not asked.

The difference people most often blur is R and A. Responsible is about doing; Accountable is about answering for the result. On a small team the same person is often both, which is fine: write A for them, because accountability implies they will see it done.

An example RACI matrix

Here is a matrix for a small agency relaunching a client’s website. Five people, one AI assistant, seven pieces of work. Read across a row to see who does what on a task; read down a column to see one person’s load.

Task / decision            Client  Lead(Priya)  Designer  Developer  AI assistant  Support
-------------------------  ------  -----------  --------  ---------  ------------  -------
Agree scope and budget       A         R           C          C           -           I
Page designs                 C         A           R          C           -           -
Build the pages              -         A           C          R           R           -
Write alt text and meta      C         A           -          C           R           -
Test on phones and browsers  -         A           -          R           R           -
Go-live decision             A         R           I          C           -           I
Tell customers it changed    C         A           -          -           -           R

Three things to notice. Every row has one A, and it is always a named person. The AI assistant appears only in the R column. And the client is Accountable for the two decisions that are theirs to make, scope and going live, while Priya is Accountable for the work, which is how most agency relationships actually run.

The rules

1. Exactly one A per row. Two people accountable means each assumes the other will check; no A means nobody will.
2. At least one R per row. A task somebody answers for but nobody does is a wish.
3. A can also be R. On small teams it often is; mark it A.
4. Keep C small. Every Consulted person is someone the work waits for. The NHS guidance warns that too many consulted roles cause delays.
5. I is not a request for input. If you want someone’s opinion before the work is done, they are C, not I.
6. The A must have the authority to say done. If the Accountable person cannot actually approve, reject or stop the work, the matrix is describing an organisation that does not exist.
7. Rows are tasks or decisions, not people. Name the outcome: “Go-live decision”, not “Priya’s bits”.

How to build one in half an hour

1. List the work as outcomes: ten to fifteen rows at most. A matrix with sixty rows is never read again.
2. List the people or roles across the top. Use roles if people change often.
3. Fill in the A column first, row by row. Most arguments happen here, which is the point.
4. Add R, then C, then I.
5. Read each column. Someone with a column of As is a bottleneck; someone with no letters may not need to be in the meetings.
6. Show it to everyone named and change what they dispute. A RACI nobody agreed to is a document, not an agreement.

Common mistakes

• Several As on a row, usually to avoid an awkward conversation about who is really in charge.
• Every senior person marked C “to be safe”, so every task waits for three approvals nobody asked for.
• The matrix is made at kick-off and never updated, while the people and the work change underneath it.
• Rows that are too big (“Deliver the website”) to tell anyone what to do.
• Using it to assign blame after the fact rather than clarity before the work.
• Giving the A to a tool, a team or a committee. Accountability that cannot be put to one person’s name is not accountability.

Variants, one line each

None of these has a single owner either, and the letters are used a little differently from one organisation to the next.

• RASCI (or RACIS): adds Support, for people who help the Responsible person without owning the work.
• RACI-VS: adds Verify, who checks the result against the requirement, and Sign-off.
• CAIRO: adds Omitted, or Out of the loop, to record deliberately who is not involved.
• DACI: Driver, Approver, Contributors, Informed; used mostly for decisions rather than tasks.
• Responsibility assignment matrix (RAM): the general name for the family, of which RACI is the best known.

Where AI agents fit

An AI agent can be Responsible. It can write the alt text, run the browser tests or draft the release notes, and on a board it can move the work along like anyone else. It cannot be Accountable. Accountability means someone can be asked why, can be held to the result, and can decide to stop; an agent can do none of those in any sense that matters to a client, a regulator or a court. The OECD AI principles↗ (opens in a new tab) put it plainly: the organisations and individuals who develop, deploy or operate AI systems should be held accountable for their proper functioning.

• R: yes. Give the agent the task, and a role that lets it do that task and nothing more. Roles and permissions for humans and AI agents covers how to size that role.
• A: never. For every row an agent is R on, a named person is A, and that person reviews the agent’s evidence before the work counts as done. Human in the loop for AI agents covers where that review belongs.
• C: sometimes. An agent can be asked for a review or a second opinion, but its advice is input to a person, not a sign-off.
• I: useful. An agent that reads the board before it starts is, in effect, Informed of every decision written there.

Regulation points the same way. For high-risk AI systems, Article 14 of the EU AI Act↗ (opens in a new tab) requires that they can be effectively overseen by people while in use, including being able to override the output or stop the system. A practical test for any task: if the agent got it badly wrong, whose name would be on the apology? That person is the A. If the answer is nobody, the task should not be given to an agent yet.

RACI on a fenbs board

fenbs has no RACI field and does not need one; the letters map onto things the board already records. A simple habit is to put the RACI line at the top of a task’s note, for example “A: Priya. R: Claude via Priya. C: Tom.”, and let the board do the rest:

• Responsible shows in the History: every change is recorded under the name of the person or AI assistant who made it, an assistant’s as “Claude via” the person it acts for.
• Accountable is the person who moves the task to Completed. Moving between lanes is its own permission, separate from editing, so the rule holds: the assistant leaves finished work in In Progress with its evidence, or connects with read and comment scopes only, and a person makes the move. For decisions, the board’s Decisions page records who decided, and the decider is always a person, never an assistant.
• Consulted is Ask someone on a team board: it emails the person a link to the board, and the question becomes a comment on the task.
• Informed is Follow this task, which emails a follower when the task moves.