A RACI matrix is a table that maps each task or decision to four kinds of involvement: Responsible (does the work), Accountable (answers for it, only one person), Consulted (asked before) and Informed (told after). For AI agents, it settles who owns each agent, who reviews its held actions, who approves changes and who answers when something goes wrong.
Why use one for agents?
Because agents create tasks that nobody owned before. Who approves a change to the agent's brief? Who decides whether a capability moves from depth 2 to depth 3? Who reads the record each week? Who is told when a certificate lapses? In a team without agents, none of these existed. Unless someone writes them down, each falls to whoever notices first, or to nobody.
A RACI also exposes a common gap. Teams often list the agent as Responsible for routine tasks and then leave Accountable blank. An agent can do the work; it cannot be accountable for it. Every row needs a named person in the A column.
What does a RACI for agent oversight include?
Typical rows:
| Task | R | A | C | I |
|---|---|---|---|---|
| Change the agent's brief | Agent owner | Process owner | Second line | Reviewers |
| Raise a capability's depth | Agent owner | Process owner | Risk, control owner | Team |
| Decide held actions | Certified reviewer | Team lead | Specialist owner | None |
| Weekly record review | Team lead | Process owner | Agent owner | Second line |
| Contain an incident | Incident commander | Process owner | Agent owner, risk | Executive |
The RACI template for AI agent operations gives a fuller, ready-to-copy version.
An example
Take a hypothetical bank's AML team. An agent assembles alert files and drafts dispositions, which investigators decide. The team builds a RACI. Investigators are responsible for decisions on held alerts; the head of financial crime is accountable; the money laundering reporting officer is consulted on anything that may lead to a report; the second line is informed of weekly patterns. Building it shows that nobody was accountable for keeping the agent's sanctions list source current. The head of financial crime takes it, with the agent owner responsible. What investigators themselves need to learn first is covered in what AML investigators should learn before working with AI agents.
How is a RACI different from a skills matrix or training matrix?
A RACI says who does what. A skills matrix says who can do what: each person's level in each skill. A training matrix says who must complete which training, and whether they have. The three should line up. If the RACI makes someone Responsible for deciding held payments, the skills matrix should show they can, and the training matrix should show a current certificate. Where they do not line up, the RACI is assigning work to people not yet qualified for it.
Related terms
Process owner, agent owner, control owner and separation of duties all appear in a well-built RACI. For a version focused on oversight rather than operations, see the RACI matrix for AI agent oversight. The guide to AI agent skills by role describes each role in more depth.
Where the Academy fits
Pryme Intelligence Academy gives teams a way to check the R column against real competence. The admin view shows each person's progress, results, certificates and expiry dates, so a team lead can confirm that everyone marked Responsible for review holds a current Certified Pryme Operator credential.
Questions people ask
Can a row have two people accountable?
No. Accountable means one person who answers for the outcome. If two people share it, neither does. Several people can be Responsible.
How often should the RACI be updated?
Whenever an agent gains a capability, changes depth or a role changes hands, and at least once a year as a check.