RACI Matrix
The RACI matrix crosses an initiative's stakeholders with its activities and each cell carries one of four assignments: Responsible, who does the work; Accountable, the single person who decides and answers for the outcome; Consulted, whose opinion is sought before the decision; Informed, who is told the outcome afterwards. It turns a tacit division of responsibility into a written commitment, activity by activity, and it does so while contesting it still costs almost nothing. BABOK describes it among the stakeholder maps; PMI files it under responsibility assignment matrices and treats it as an instrument of governance.
Goal
The RACI matrix assigns, for each activity of an initiative, one of four responsibilities to each stakeholder: doing the work, answering for the outcome, being consulted before the decision or informed after. It thereby puts on the table, while it is still negotiable, a division of responsibility every initiative makes in any case, whether it writes it down or not. That tacit division works until something jams, and it surfaces at the first hard call: the activity two departments each believed the other was holding, the one nobody was holding, the decision taken without asking the person who had the expertise and the refusal issued by the person nobody had asked anything of. The disagreement the matrix provokes in the room is the disagreement that would otherwise have arrived six months later in an expensive form.
The decision it supports is a governance decision: for each activity in the plan, who decides, who does the work, who is asked before, who is told after. Those four questions get answered one way or another. The matrix asks them together, in an order, in front of the people concerned, and it demands an answer. Its value lies in the commitment made out loud by the person whose name goes in the cell, and the grid is the record of it.
The deliverable has three inseparable parts. The filled grid: one row per activity, one column per stakeholder, one letter per cell and one only and cells left blank where the stakeholder has no part. The legend, written on the grid itself, defining the four letters for this project. And the date, with the moments the grid will be replayed: a gate, a scope change, a reorganisation. A matrix with no legend carries four meanings, one per reader. A matrix with no date is a snapshot nobody can still vouch for.
Usage
When to use it
- A team mixing internal and external resources: this is the trigger PMI names explicitly, because the contractual boundary blurs responsibility.
- An activity crossing several departments: fix the single accountable before the boundary becomes a void.
- A decision stalled with nobody knowing who calls it: the grid assigns the decision to a single role.
- A project plan that is written and approved: the activity axis is ready, the grid takes it as it stands.
- Work subject to audit, review or a contractual constraint: establish who checks and who signs, before the first inspection.
- A reorganisation or a change of project manager: hand over an explicit division rather than tacit knowledge that leaves with the person.
- A project run under HERMES: make explicit what the method leaves open, the participating roles and the difference between consulting and informing.
- A communication plan to be written: the C and I columns are its raw material, consulted before which decision, informed after which outcome.
When not to use it
- A question of access rights in a system: that is about functions and not activities, reach for the roles and permissions matrix.
- One co-located team, short scope: the grid costs more than it teaches, keep a dated decision log instead.
- Collective authority written into the statutes: no activity admits a single accountable, describe the body, its quorum and its voting rules in the governance plan.
Description
The four assignments
The responsible (R) is the person or group doing the work of the activity. The plural is normal: several people work on one activity, and nothing obliges you to single one out. The accountable (A) is the person ultimately answerable for the activity's completion and the person who decides. BABOK says so in its own definition: only one stakeholder receives this assignment. The consulted (C) is the person asked for an opinion or information about the activity, typically the domain expert. The informed (I) is the person told the outcome.
The distinction between C and I carries the whole weight of the technique, and BABOK states it in its own terms: with the informed, communication runs in one direction, from the analyst to the stakeholder; with the consulted, it runs both ways. The sentence has a consequence the standard does not draw. A C is a promise. It says that this person's opinion is sought before the decision and can still change it. An I is a report. It says the decision is made and the person learns it afterwards. Marking someone C and then deciding without them costs more than marking them I from the start: it is a broken promise, and it is the commonest way a RACI matrix destroys the trust it was meant to build.
The legend is part of the artefact. BABOK notes the practice of defining each term on the grid itself, so that every stakeholder reading it understands the same thing by the same letter. The observation looks trivial and is not: consulted means, to one person, "my view will be heard" and to another, "I shall have been notified", and a disagreement about the letter is always discovered on the day it is too late to settle.
The single-A rule
An activity carries exactly one accountable. PMI gives both the rule and its reason: the responsibility assignment matrix ensures that only one person is accountable for any one task, which avoids confusion about who is ultimately in charge and holds authority over the work. Two A on a row read as consensus and behave as deadlock: the day the two accountables disagree, the row stops, and the matrix nominates nobody above them to settle it. BABOK states the same constraint, in its definition of the assignment, and it describes it without grounding it.
In Beckhard and Harris's responsibility chart (1977), the ancestor of the technique, there is no A in the sense of accountable at all: the four codes are R (responsibility to initiate action), A-V (approval required or the right of veto), S (support, in resources or logistics) and I (inform, with no capacity to influence). Their guideline, "assign responsibility to only one person", applied to the R, which at the time fused execution and accountability. Modern RACI relabelled the A, split it off from the R and moved the singularity constraint onto it. PMI and IIBA have both since confirmed it there. The technique has no single inventor and the accounts that give it one do not check out: what checks out is the 1977 chart and its four codes, which are not today's four.
One field observation, which is ours and which appears in no standard: the total number of A on a grid is an indicator in itself. A grid where one person carries half the A has documented a bottleneck. The corrective move rests on a distinction: the authority to decide can be delegated, accountability cannot. So you split the row until a different person can honestly carry the A for the half that comes off, and you stop short of handing the decision to a delegate while leaving the overloaded person's name in the cell.
The two axes
The activity axis is built from the real plan. PMI is precise about granularity: the responsibility assignment matrix shows the resources assigned to each work package. A row is therefore a work package. The practical test: a row is an activity for which a single accountability is meaningful. Too coarse, and the row produces an A empty of content, because "accountable for the project" is a job title. Too fine, at the granularity of "update the configuration file", and the grid runs to three hundred rows and nobody opens it again.
When the two requirements pull in opposite directions, PMI's answer is two nested matrices: a high-level matrix assigning the responsibilities of a team, group or unit for each component of the work breakdown structure and a lower-level matrix, held inside the group, that descends to specific activities and their levels of authority.
The stakeholder axis is built from the stakeholder list. The matrix consumes that register: if the list is incomplete, it is the list that needs reworking. It answers a different question from the stakeholder matrix, which calibrates engagement effort by power and interest: the one says how much time to spend on someone, the other says what they answer for. Both belong to the family of stakeholder analysis instruments, alongside the onion diagram and personas.
Roles first, names second and a name wherever the accountability is real. A role can be accountable in a design; only a person can be accountable in fact. A grid that reads "HR director: A" has assigned accountability to a box on an org chart.
Running the session
- Build the activity axis from the plan
The work packages of the breakdown structure, the deliverable list or the decision list. One active verb per row, one constant granularity. - Build the stakeholder axis from the list
The roles, then the names wherever the accountability is real. - Write the legend onto the grid
The four terms defined for this project, before the first cell. - Fill the A first
One per row, out loud, with the candidate in the room. An A is a commitment, and a commitment is made in the presence of whoever makes it. - Then the R, the C and the I
Leave the blank cells blank. - Run the structural checks before the room empties
Exactly one A per row; at least one R per row; no column carrying only C and I; nobody carrying an implausible number of A; no activity in the plan without a row. - Read it back column by column, stakeholder by stakeholder
Take an explicit yes. Silence means no objection was voiced, and that turns out to differ from agreement on the day someone has to act. - Publish the grid where the team works and replay it at every gate
Its value is realised at the decision, weeks later.
The variants
The variants circulate with no primary source, and none of them comes from a standards body. What matters is what each one buys and the price it charges: every added code is a training cost paid again by every reader of the grid.
| Variant | The added code | What it buys | When it is worth the letter |
|---|---|---|---|
| RASCI | S: supportive | Makes visible the dependency that sinks the R: the infrastructure team, the data team, the supplier whose capacity the R needs and over whom the R has no authority. | When the R's ability to deliver depends on another team's capacity. S was already one of Beckhard and Harris's codes, and they held it to be critical. |
| RACI-VS | V: verifier, S: signatory | Splits the A into the person who checks the work against criteria and the person who signs. | Regulated, audited or contractual work where the verifier is deliberately not the signatory. If both are the same person, both letters are noise. |
| PACSI | C: control (review with veto), S: suggest (advice, no veto) | Splits the C into the reviewer who can stop the work and the reviewer who can only advise. | Review-heavy environments: security, architecture boards, compliance. The distinction already exists in Beckhard and Harris, whose A-V was approval with a right of veto. |
| CAIRO | O: out of the loop | Marks non-involvement explicitly. | Rarely: the blank cell already says it. It earns its place where the blank is ambiguous, in an organisation where people assume they are consulted by default or on a partial extract of a grid where a blank could mean "not shown". |
| DACI | D: driver, A: approver, C: contributor, I: informed | Changes the unit of the row: a RACI row is an activity, a DACI row is a decision. The driver gets the decision made, convening, framing and chasing, and the work itself belongs to others. | When the bottleneck is decisions. It is a distinct instrument from RACI, and presenting them as interchangeable loses both. |
The rule of conduct is ours: add a code only when its absence is causing a concrete failure you can name. A supplier's capacity nobody answers for calls for the S. A signature nobody can produce calls for the V and the S. A veto nobody knew existed calls for control. Everything else is one more column that makes the matrix look thorough and slow to read, when four letters are already at the limit of what a whole room reads the same way.
RACI, roles and permissions matrix, CRUD matrix
Three matrices resemble each other and are constantly confused, because they have the same shape. They do not have the same cell, and the cell is what decides.
| Matrix | Axes | Cells | The question it answers |
|---|---|---|---|
| RACI matrix | stakeholders × activities | R / A / C / I | Who answers for the work? |
| Roles and permissions matrix | roles × functions | create / read / update / delete | Who may do what in the system? |
| CRUD matrix | use cases × entities | create / read / update / delete | Is the process coverage of the data complete? |
Consulted and informed are communication states between people. They carry no meaning in a functional design: a role cannot be informed of a function, and a system function does not ask a role's opinion before executing. Symmetrically, the cells of a CRUD matrix are operations on data, and nobody is accountable for an operation on a piece of data. Folding RACI into an access matrix would produce a table whose cells mean one thing in one row and another in the next.
The RACI on a HERMES project
HERMES, the Swiss Confederation's project management method, does not use RACI: the method ships the responsibility assignment already built. Its role descriptions show, module by module, which concrete tasks the role is responsible for (verantwortlich) and which other roles are participating (beteiligt) in producing the results. The role responsible for a task also carries responsibility for producing the results and for the results themselves.
Three consequences decide what a RACI still adds. R and A are fused: the responsible role does the work and answers for it, HERMES makes no split. The singularity is structural: the method places every role on one of three hierarchy levels, steering, management and execution, and it pre-assigns the decisions. Steering decisions belong to the project sponsor, who carries overall responsibility: project release, phase releases, go-live, project close. Management decisions belong to the project manager, who holds sole management responsibility. In RACI terms, the A column of a HERMES project is largely decided before the project starts. And HERMES draws no distinction between consulted and informed: participating covers both, and the method itself states that the list of participating roles is not exhaustive and must be fixed project by project.
You start from the method's role/task assignment, and the matrix adds exactly three things HERMES leaves open.
| What HERMES fixes | What the RACI adds |
|---|---|
| One responsible role per task, R and A fused | The split of the fused responsible into A and R, wherever the work is done by people other than the accountable |
| Participating roles, an explicitly non-exhaustive list | The list settled, project by project, exactly as the method itself asks |
| No distinction between consulting and informing | The separation of C and I, so the opinion sought before the decision and the outcome communicated after it |
The limit reads off the same movement: a grid that gave the A on a phase release to anyone but the project sponsor contradicts the method the project is run under. That is a defect, and it shows up in the column-by-column read-back.
What makes the exercise fail
Filling every cell
The blank looks unfinished, so it gets filled with I. The grid becomes a distribution list, being informed stops meaning anything and nobody reads the notifications any more. The blank cell is the richest on the grid: it states, in writing, that this person has no part in this activity and no business in the meeting.
Two A on a row
It reads as consensus and behaves as deadlock, for want of a terminus to the escalation path. It is the defect PMI's rule exists to prevent.
No A on a row
It is almost always a decision never taken, surfacing: the activity crosses a boundary nobody wants to own. The right response is to hold the argument and to resist filling the cell with the project manager on the grounds that the cell was empty.
A and R systematically collapsed
Legitimate on short work, and it was Beckhard and Harris's R. It is a defect when it becomes the rule across the whole grid: it means nobody above the doers has committed to anything and the matrix has documented the team's workload instead of the project's governance.
The wrong granularity
Too coarse and the A means nothing. Too fine and the grid runs to hundreds of rows and is never re-read. PMI's answer is nesting two matrices.
Roles with no names
Only a person can be accountable. The name is what makes the workshop uncomfortable, and the discomfort is the sign that the technique is working. BABOK concedes as much in the limitations it states for the family: assessing a given representative's real influence "can be complicated and may feel politically risky". That is an exact description of the session, and it is also why people happily replace it with a grid of anonymous roles, which has the merit of offending nobody and the drawback of committing nobody.
An A who cannot decide
An accountable with the authority but not the availability produces a row with no A, plus extra steps. Test each A against the real calendar and the real delegation rules before writing it down.
The courtesy C
Marking someone consulted so as not to offend them, then deciding without them. The C promises that the opinion can still change the outcome: a courtesy C is a promise the project has no intention of keeping.
Built once, never reopened
The matrix's value is realised at the decision, weeks after the workshop. If nobody opens it at the first real disagreement, it was theatre. It is re-read at every gate, every scope change, every reorganisation.
The grid instead of the conversation
A matrix records agreements, it creates none. Circulated by e-mail, approved by silence, never argued over, it documents nothing but its author's assumptions.
AI considerations
The first useful application is building the activity axis. A language model given the breakdown structure, the deliverable list, the product backlog or the process model proposes rows with active verbs at a constant granularity, which is the tedious half of the preparation.
The second application is the most profitable of all: detecting structural defects. Every structural check is verifiable without judgement: the model finds the rows with zero A or two A, the rows with no R, the columns carrying only C and I that betray a stakeholder consulted out of politeness, the person carrying half the A and the activities in the plan that have no row. A model, or even a spreadsheet formula, finds all of that in seconds; a room re-reading a forty-row grid does not. The third application extends the second: give the model the old grid and the new plan and ask for the difference, every new activity with no assignment and every retired activity still carrying one. The fourth is the per-stakeholder view, "here is what you answer for", column by column, which is what the column read-back needs and what nobody has time to write by hand. The fifth is deriving the communication plan from the C and I columns.
What AI cannot do follows from the nature of the assignment. The A is a political act. It is a commitment extracted from a person in a room. A model asked who is accountable for the go-live will answer from job titles, which means it will reproduce the org chart, the thing the technique exists to test: where the org chart already gave the answer, the matrix was not needed. Likewise, the split between C and I is a judgement about real influence. The model cannot know who is capable of stopping a decision in this particular house, and it will over-consult the senior while under-consulting the person who holds the operational veto.
A filled grid is a list of named people with their duties and their powers, so it is personal data under the revised Swiss Data Protection Act. It does not get pasted into a public model, and the assistance is run on a tool whose data processing is controlled or on a grid whose names have been replaced by roles.
Examples
An industrial SME in French-speaking Switzerland, roughly 450 employees, is migrating its payroll and time management to a swissdec-certified solution. The project plan is written, and the grid is filled at role granularity.
RACI matrix
Payroll and time management migration
| Activity | Executive management | HR director | Payroll manager | Finance director | IT project manager | Software vendor | Staff committee | Statutory auditor |
|---|---|---|---|---|---|---|---|---|
| Scope definition | A | R | C | C | R | C | C | |
| Payroll data migration | I | A | R | I | R | R | C | |
| AVS / LPP / LAA configuration | C | A | C | I | R | C | ||
| Compensation fund interface | I | C | A | R | ||||
| Parallel payroll run | I | I | A | R | R | R | C | |
| Acceptance testing | R | A | C | R | C | I | ||
| Go-live | A | C | R | C | R | R | I | I |
| User training | A | C | I | R | I |
- Responsible: does the work.
- Accountable: decides and answers for the outcome. One per row.
- Consulted: their opinion is sought before the decision.
- Informed: receives the outcome after the decision.
The concept the grid makes visible is the single A, and it does so through what it makes impossible: a row with two A would leap out, a row with none would leave a hole in a column that holds eight. Neither external role ever carries an A. The software vendor carries six of the grid's sixteen R, it does most of the technical work and it answers for nothing to the company, because accountability cannot be subcontracted. The statutory auditor carries neither R nor A: it is consulted and informed, which is the position of an independent controller with respect to work it will later have to check. The payroll manager carries three of the eight A, which is the right result for this project and an alert: the grid has just identified the schedule's bottleneck, and her availability becomes a planning constraint.
Two columns carry only C and I, which trips the corresponding structural check. The check asks a question and the answer is that this really is the role of these two players. The staff committee carries a C on scope definition, where working-time recording gives it a genuine say, and I elsewhere: the C commits the project to hearing it before the scope is settled, and marking it I would have been an elegant way of presenting it with a fait accompli. The statutory auditor is consulted on the data migration, on the social-insurance configuration and on the parallel payroll run, that is, everywhere the project touches what it will later have to check, and it is informed of the go-live. A column with neither R nor A is legitimate when the stakeholder's role is consultative by nature; it becomes a defect when it reveals a stakeholder written into the grid so as not to have forgotten them.
Visualisations
The deliverable is the table itself, made of rows and columns. PMI gives the reason the form matters: the matrix shows all activities associated with one person and all people associated with one activity, and those two readings are the real use of the artefact. You read a row to prepare a decision; you read a column to go and tell someone what they answer for. A picture of a table loses both: it does not sort, does not filter, does not read back by column and cannot be checked by formula.
Three conventions separate a usable grid from a drawing. The legend sits on the grid, with the four terms defined for this project. The A is visually distinct from the other three letters, because it is the only assignment under a cardinality constraint and therefore the only one that can be checked by eye. Blank cells stay blank, with no dash and no dot, since the blank carries the information and a polite filling destroys it.
The asymmetry between the consulted and the informed does not fit in a table, because it is a sequence around the decision. It has to be drawn: the consulted's opinion goes out before the decision and comes back, the informed's notification goes out after it and comes back with nothing. Finally, the grid has a readability ceiling. Past thirty or so rows it stops being re-read.
Cost
| Phase | Level | Justification |
|---|---|---|
| Preparation | Medium | The stakeholder axis already exists, it comes out of the stakeholder list. The activity axis is the substance of the work: it is derived from a genuine plan and set at a granularity where a single accountability is meaningful. |
| Execution | Medium to high | Filling cells is fast. Negotiating the A is not: the whole session goes to it, a facilitated workshop with the real decision-makers. The cost turns high as soon as an A crosses an organisational boundary, an external supplier or another company, because it is then a negotiation. |
| Documentation | Low | The artefact is a table and its legend, and producing it is immediate. The real cost is upkeep: re-reading the grid at every gate and every scope change. A matrix that was cheap to produce and never maintained has cost more than it returned. |
Tooling
The spreadsheet is the honest choice, and it is better than most dedicated tools: the structural checks are formulas. A rule that turns red any row where the count of A is not one enforces the single-A rule automatically, and it catches the defect the moment it is created rather than at the review, when it has become a subject. The same formulas count the R per row, spot the columns carrying only C and I and flag the person accumulating the A.
A table in the team's workspace, in the wiki or in the repository, next to the plan, puts the grid where the team already reads and versions it. That is the only real answer to the upkeep failure: a matrix that requires opening another tool is a matrix that will not be reopened.
Jira, Azure DevOps and their kin carry an assignee and watchers, which approximate R and I. They have no A and no C, and pretending otherwise loses the two assignments the technique exists for. Smartsheet, Monday, Asana and Wrike ship RACI templates that are spreadsheets with better sharing. Modelling and enterprise-architecture tools, where the process is already modelled with its lanes and roles, allow the assignment to be generated from the model instead of maintained beside it, which removes the drift; the investment only pays where a maintained process model already exists.
On a HERMES project the tool is first of all the method: the role and task tables and the project handbook already carry the assignment, and the RACI matrix pins down what they leave open. That leaves the anti-tool, the slide. A RACI in a presentation is a photograph of a decision: it cannot be filtered, checked or diffed from one version to the next and it is wrong the week after kick-off.
Sources
- IIBA, A Guide to the Business Analysis Body of Knowledge (BABOK Guide) v3, §10.43 Stakeholder List, Map, or Personas: the four assignments, the single accountable in the definition, the asymmetry between consulted (two-way communication) and informed (one-way), the practice of defining the terms on the artefact and the acknowledged limitation that assessing a representative's real influence can be complicated and may feel politically risky.
- PMI, A Guide to the Project Management Body of Knowledge (PMBOK Guide), 8th edition, §5 Tools and Techniques, "Responsibility assignment matrix": the responsibility assignment matrix, the work-package granularity, the two nested matrices, the two reading directions, the rule of a single accountable per task and its reason and the team of internal and external resources as the trigger. Also §1.3.2 (governance and its decision-making models) and §2.6 (Resources performance domain).
- HERMES 2022, Project Management Reference Manual, Swiss Confederation, Federal Chancellery, ch. 6 Roles (§6.1 role model, §6.3 role descriptions) and ch. 5 Tasks (§5.1.2 decision tasks): the pre-built role/task assignment, the responsible role that also carries responsibility for the results, the participating roles whose list must be fixed project by project and the pre-assignment of decisions to the project sponsor and the project manager.
- Beckhard, R. and Harris, R. T., Organizational Transitions: Managing Complex Change, Addison-Wesley, 1977: the responsibility chart, its four original codes (R, A-V, S, I) and the guideline assigning responsibility to only one person. It is the primary source of the responsibility chart that RACI descends from, and its four codes are not today's four.
- NHS England, Responsibility charting (Online library of Quality, Service Improvement and Redesign tools): a public body's restatement of Beckhard and Harris's chart, with the four codes and their effectiveness guidelines.