Is RACI Enough for AI? Building Accountability That Actually Works
RACI has helped organisations clarify responsibilities for decades.
Responsible.
Accountable.
Consulted.
Informed.
It is simple, familiar and useful.
So when organisations begin establishing AI governance, there is an obvious temptation:
Create an AI RACI.
List the stakeholders.
Assign responsibilities.
Agree who needs to be consulted.
Problem solved.
Except AI creates an accountability challenge that is more complicated than allocating responsibility for a conventional project.
An AI system can influence decisions.
It can change over time.
It may interact with sensitive data.
Its outputs can affect customers or employees.
AI agents can increasingly take actions.
Multiple functions may own different parts of the same system.
And the people responsible for implementing AI may be completely different from the people accountable for the business outcome it is expected to create.
A RACI can help.
But on its own, it may not answer the most important questions:
Who owns the outcome?
Who has authority to approve the AI?
Who decides what data it can access?
Who determines how much autonomy it receives?
Who can override it?
Who owns the risk?
Who owns the value?
And who can stop it?
That is why AI accountability needs to go beyond RACI.
What is RACI?
RACI is a responsibility-assignment model used to clarify who is involved in an activity or decision.
The four roles are:
RESPONSIBLE
The person or people performing the work.
ACCOUNTABLE
The person ultimately answerable for the outcome.
CONSULTED
People whose expertise or input is required.
INFORMED
People who need visibility of the decision or outcome.
For many business processes, this is extremely useful.
The problem is not that RACI is wrong.
The problem is that AI creates multiple different types of accountability within the same use case.
Why AI makes accountability more complicated
Consider an AI system introduced into customer service.
Technology may be responsible for the platform.
Customer Service may own the workflow.
Security may own technical controls.
Data Protection may oversee personal information.
Risk may establish requirements.
HR may support employee adoption.
Finance may validate benefits.
A governance committee may approve higher-risk use.
The executive sponsor may own the wider transformation.
Everyone is involved.
A RACI could document all of this.
But leadership still needs to know:
Who is accountable for the customer outcome?
and:
Who has authority to make each material decision?
These are not always the same person.
For more on the distinction between ownership and decision authority, read AI Decision Rights: Who Should Be Allowed to Approve, Override or Stop AI?
The first problem: AI has more than one type of owner
As the Stratify methodology has developed, several distinct ownership roles have emerged.
Executive Owner
Who is accountable for the organisation's overall approach to AI?
Business Owner
Who owns the business process or outcome affected by the AI?
Capability Owner
Who owns the technology or AI capability?
Use-Case Owner
Who owns the particular application of AI?
Risk Owner
Who is accountable for the material risks associated with the use case?
Value Owner
Who is accountable for ensuring the expected benefit is realised?
These roles may overlap.
A business leader might be both:
Business Owner
and
Value Owner.
A CIO might be:
Capability Owner
and potentially the executive AI sponsor.
That is fine.
What matters is that the organisation has consciously determined which forms of ownership exist and who holds them.
For more on why value ownership needs to be explicit, read Who Owns AI ROI? Why Every AI Investment Needs a Business Value Owner.
The second problem: accountability and authority are different
Imagine a Sales Director is told:
“You are accountable for the value created by this AI system.”
But they cannot:
change the workflow;
require employees to use it;
influence implementation;
alter performance expectations;
request changes;
or recommend stopping it.
They have accountability without authority.
That is structurally weak.
Likewise, someone may have authority without accountability.
A technical administrator might be able to increase an AI agent's permissions even though they are not accountable for the business consequences of that decision.
AI accountability therefore needs to align:
ACCOUNTABILITY
with
AUTHORITY.
This gives us a fundamental principle:
Nobody should be accountable for an AI outcome they lack sufficient authority to influence.
And:
Nobody should have material AI authority without corresponding accountability for how that authority is exercised.
The third problem: responsibility changes across the AI lifecycle
AI accountability is not static.
Different decisions occur at different stages.
Consider:
IDENTIFY
Who proposes the AI opportunity?
↓
ASSESS
Who determines whether it is viable and appropriate?
↓
APPROVE
Who authorises investment and use?
↓
DEPLOY
Who owns implementation?
↓
OPERATE
Who owns day to day performance?
↓
MONITOR
Who reviews risk and outcomes?
↓
CHANGE
Who can alter permissions, models or workflows?
↓
SCALE
Who approves expansion?
↓
STOP
Who can suspend or terminate the AI?
A single RACI created at the start of the project may not adequately reflect all of those decisions.
AI accountability therefore needs to operate across the lifecycle, not merely the project.
The Stratify AI Accountability Stack
I would structure AI accountability across six layers:
OUTCOME → OWNERSHIP → AUTHORITY → RESPONSIBILITY → OVERSIGHT → EVIDENCE
Each answers a different question.
1. OUTCOME
What are we accountable for achieving or protecting?
This comes first.
Examples might include:
increasing sales capacity;
reducing processing cost;
improving customer service;
maintaining fair employee decisions;
protecting customer data;
reducing operational risk.
Without a defined outcome, accountability becomes vague.
2. OWNERSHIP
Who is ultimately answerable for that outcome?
This should normally be a named role or person.
Not:
“Technology.”
Not:
“The AI committee.”
Not:
“The business.”
A name or clearly defined role.
For more on this foundational problem, read Who Owns AI in Your Business? If the Answer Isn't Clear, You Have a Problem.
3. AUTHORITY
What decisions can that owner actually make?
Can they:
approve the use case?
change the workflow?
require remediation?
adjust AI authority?
escalate concerns?
recommend additional investment?
stop the initiative?
Ownership without appropriate authority creates accountability theatre.
4. RESPONSIBILITY
Who actually performs the work?
This is where traditional RACI remains extremely useful.
Technology might implement.
Security might test.
HR might train.
Managers might reinforce adoption.
Finance might validate benefits.
Multiple people can be responsible.
But the accountable owner remains clear.
5. OVERSIGHT
Who independently challenges or reviews the activity?
This might involve:
Risk;
Legal;
Compliance;
Security;
Internal Audit;
Board oversight;
specialist governance.
Oversight is different from ownership.
A Risk function may challenge the use case without owning the business outcome.
A Board may oversee AI without operating it.
For more on this distinction at Board level, read Board Oversight of AI: What Directors Actually Need to Know.
6. EVIDENCE
How do we know the accountable parties are fulfilling their responsibilities?
This is the layer RACI often does not address.
Evidence might include:
approved business case;
risk assessment;
named owner;
testing;
human oversight records;
adoption data;
performance metrics;
incident records;
benefits measurement;
review decisions.
Accountability without evidence can become little more than an organisation chart.
From RACI to accountable AI
This gives us a broader model.
Traditional RACI answers:
Who is involved?
The Stratify AI Accountability Stack asks:
What outcome matters, who owns it, what authority do they have, who performs the work, who provides oversight and what evidence proves the system is working?
That is a more complete accountability architecture.
It does not replace RACI.
It extends it.
The Stratify AI Accountability Map
For every material AI use case, leadership should be able to complete something like:
Accountability question | Named role |
Executive sponsor | ? |
Business owner | ? |
AI capability owner | ? |
Use-case owner | ? |
Risk owner | ? |
Value owner | ? |
Approval authority | ? |
Human oversight owner | ? |
Monitoring owner | ? |
Stop authority | ? |
The same person can appear more than once.
The point is not to create ten new jobs.
The point is to expose accountability gaps.
If leadership cannot fill a field, it has identified something that needs resolving.
Accountability should be proportional to risk
Not every AI use case needs the same organisational structure.
An employee using an approved AI tool for low-risk internal drafting does not need ten named owners and a governance committee.
A high consequence AI system influencing customer eligibility might.
This means accountability requirements should increase with:
impact;
data sensitivity;
decision significance;
autonomy;
scale;
reversibility.
For more on determining the appropriate level of governance, read AI Risk Classification: How to Classify AI Use Cases by Risk.
What happens when AI agents enter the model?
AI agents make accountability more complicated because software can increasingly perform actions previously carried out by employees.
Imagine an agent can:
access customer records;
decide which workflow to follow;
send communications;
update systems;
issue refunds;
trigger additional processes.
The organisation now needs to distinguish between:
who performs the action
and
who is accountable for the action being permitted.
The agent can execute.
It cannot ultimately hold organisational accountability.
This gives us another important principle:
AI can receive delegated authority. It cannot receive ultimate accountability.
A human or organisational role must remain accountable for the boundaries within which the agent operates.
For more on how authority changes as AI becomes more autonomous, read AI Assistants vs AI Agents: What Business Leaders Need to Know Before Giving AI Autonomy.
The Delegation Chain
When AI acts, leadership should be able to trace:
EXECUTIVE AUTHORITY
↓
BUSINESS OWNER
↓
USE-CASE AUTHORITY
↓
AI PERMISSION
↓
AI ACTION
↓
MONITORING
↓
HUMAN ESCALATION
This creates a traceable delegation chain.
If an AI agent performs an action, leadership should be able to answer:
Who authorised this type of action?
Under which policy?
Within what limits?
Who owns the outcome?
Who monitors it?
Who can intervene?
That is far stronger than simply saying:
“The AI did it.”
RACI also struggles with dynamic AI
Another challenge is that AI systems can change.
Models are updated.
Prompts change.
Data changes.
Permissions expand.
Workflows evolve.
Agents receive new tools.
Usage scales.
The original accountability model may therefore become outdated.
Material changes should trigger an accountability review.
For example:
New data
Does the existing data owner remain appropriate?
Greater autonomy
Does decision authority need to change?
Larger scale
Does oversight need strengthening?
New customer impact
Does the risk owner change?
New business outcome
Does value ownership need reassignment?
Accountability should therefore be reviewed when the AI changes materially, not simply documented once.
Committees do not own outcomes
This deserves explicit attention.
AI governance committees can be valuable.
They bring together expertise.
They can challenge higher-risk use cases.
They can coordinate standards.
They can resolve cross-functional issues.
But a committee should not become the answer to:
Who owns this AI?
Committees provide governance.
Named leaders own outcomes.
Otherwise responsibility becomes diffused across a group.
Everyone attended the meeting.
Everyone contributed.
Nobody is clearly accountable.
This creates another Stratify principle:
Committees govern. Individuals own.
For more on where governance ownership should sit, read Who Should Own AI Governance? A Practical Model for Boards and Leadership Teams.
The Board does not own operational AI
The same distinction applies to Boards.
Boards provide:
oversight;
challenge;
assurance;
strategic scrutiny.
They should not become operational owners of individual AI systems.
Management remains responsible for execution.
This matters because organisations can respond to AI risk by escalating too much activity upwards.
Strong governance is not:
more senior approval.
It is:
clearer authority at the appropriate level.
What should the CEO own?
The CEO should not personally own every AI initiative either.
The CEO's role is to ensure the organisational system has:
direction;
priorities;
accountability;
investment discipline;
appropriate governance;
workforce change;
value measurement.
Business outcomes should still sit with the relevant business leaders.
For more on this distinction, read The CEO's Role in AI: What Should the Chief Executive Actually Own?
The Accountability Evidence Pack
For significant AI use cases, I would eventually expect the organisation to maintain a small evidence set.
This might include:
Business owner
Who owns the outcome?
Use-case owner
Who manages the application?
Decision rights
Who can approve/change/stop?
Risk classification
What level of governance applies?
Human oversight
Where does human judgement remain?
Value owner
Who owns the expected benefit?
Controls
What boundaries apply?
Monitoring
How is performance observed?
Review
When was accountability last confirmed?
This does not need to become a 50-page document.
The purpose is to make accountability visible and auditable.
The AI Accountability Review
A Fractional CAIO or internal AI leader could periodically review:
OWNERSHIP
Does every material use case have appropriate owners?
AUTHORITY
Are decision rights still appropriate?
RISK
Has the risk profile changed?
AUTONOMY
Has AI authority increased?
VALUE
Is benefits ownership clear?
OVERSIGHT
Are controls operating?
EVIDENCE
Can leadership demonstrate all of the above?
This creates a recurring accountability discipline rather than a one-time RACI workshop.
Ten questions to test AI accountability
For any material AI use case, ask:
What business outcome are we accountable for?
Who ultimately owns that outcome?
Who owns the AI capability?
Who owns the specific use case?
Who owns the material risk?
Who owns the expected value?
Does each accountable owner have sufficient authority?
Who independently oversees or challenges the activity?
What evidence demonstrates responsibilities are being fulfilled?
What changes would trigger an accountability review?
If leadership cannot answer these questions clearly, completing a RACI spreadsheet alone probably will not solve the problem.
RACI is useful. AI accountability needs more.
RACI remains a valuable tool.
It helps organisations clarify:
who performs work;
who is accountable;
who contributes;
who needs visibility.
But AI requires another layer.
Leadership needs to understand:
what outcome matters
who owns it
what authority they have
who performs the work
who provides oversight
what evidence demonstrates control
and increasingly:
what authority has been delegated to AI itself.
That is why the question is not:
Should we stop using RACI for AI?
It is:
What do we need to add to RACI so accountability remains clear as AI becomes more embedded, consequential and autonomous?
Because when AI makes a recommendation, influences a decision or takes an action, the organisation still needs to know:
who ultimately answers for the outcome.
Frequently Asked Questions About RACI and AI Accountability
What is RACI in AI governance?
RACI is a responsibility-assignment framework that identifies who is Responsible, Accountable, Consulted and Informed for activities or decisions relating to AI.
Is RACI enough for AI governance?
RACI can clarify responsibilities, but AI may also require explicit business ownership, decision rights, risk ownership, value ownership, human oversight, evidence and authority boundaries.
Who should be accountable for AI?
Accountability depends on the outcome. Business leaders should normally own business outcomes, Technology may own AI capability, specialist functions may own particular risks, and executive leadership should ensure the overall accountability system is clear.
Can an AI system be accountable?
AI can perform tasks and receive delegated authority, but ultimate organisational accountability remains with people and the organisation deploying it.
What is the difference between AI ownership and AI responsibility?
Ownership identifies who is ultimately answerable for an outcome. Responsibility identifies who performs the work required to achieve or manage that outcome.
What is the difference between accountability and decision rights?
Accountability defines who answers for an outcome. Decision rights define who has authority to make specific decisions about the AI.
Should an AI committee own AI?
An AI committee can provide governance, challenge and coordination, but material business outcomes should normally have named individual or role-based ownership.
How often should AI accountability be reviewed?
Accountability should be reviewed periodically and whenever material changes occur in the AI's purpose, data, risk, autonomy, scale, permissions or business impact.
How Clear Is AI Accountability in Your Organisation?
AI accountability becomes difficult when ownership, authority and responsibility are spread across multiple functions.
Stratify helps leadership teams identify where accountability is clear — and where gaps could prevent AI from scaling safely or creating measurable value.
Use the Executive AI Accountability Assessment to examine:
Executive accountability
Business ownership
Decision rights
Governance responsibilities
Human oversight
Value ownership
Escalation
Test Your AI Accountability
Because an organisation does not become accountable for AI simply by completing a RACI.
Accountability works when ownership, authority and evidence align.
Comments