Make accountability unambiguous, who decides, who builds, who answers.
Aligns withNIST · GovernISO/IEC 42001RACI
The idea
This pillar defines the human operating model: who owns AI decisions, who implements them, who reviews them, and how those roles coordinate across a decentralized institution. It turns Pillar 4's governance structures into named responsibilities and closes the accountability gap, where an AI system affects outcomes but no one is responsible for the result.
It's made concrete through a RACI-style map for key AI decisions; named owners for each system across its lifecycle (sponsor, technical owner, data steward, reviewer); and coordination across IT, academics, legal, and the units. The result: for any consequential AI use, you can name the accountable person.
Why it matters
When responsibility is diffuse, two things happen: nothing gets owned in good times, and everything gets disowned when something breaks ('the algorithm did it'). Clear roles are what let an institution move quickly and answer cleanly (to a student, a provost, or a regulator) about who approved a system and who stands behind it.
Responsibility operates at four layers, not one list of names. Most accountability failures happen in the gaps between them, especially when the person who owns a system is mistaken for the person accountable for the decisions it informs. Name all four explicitly, then let your RACI hang off them.
Program layer
Institution-wide
Who answers for AI at the institution?
Who
Cabinet sponsor, AI governance body, AI lead
Owns
Strategy, policy, the operating model, and the program's overall health.
System layer
Per AI system
Who owns this tool across its lifecycle?
Who
System sponsor, technical owner, data steward, reviewer
Owns
A specific system from procurement through operation, monitoring, and retirement.
Decision layer
Per consequential decisionMost often skipped
Who is accountable for acting on what the AI produces?
Who
The human decision-maker in the domain: advisor, admissions officer, manager
Owns
The decision and its consequences. Often a different person than the system owner, and the layer most often skipped.
Individual layer
Every user
What does each person owe when they use AI?
Who
All faculty, staff, and students
Owns
Baseline duties of disclosure, verification, and acceptable use in daily work.
Cutting across all four is coordination: roles that span IT, academics, legal, and the units, and the “orphaned ownership” problem when something falls between them. It's a connecting concern, not a fifth layer, and it's why the layers must be named together.
How you can tell
When it's working
Key AI decisions and processes have a RACI mapping that people actually use.
Every consequential AI system has a named sponsor, technical owner, data steward, and reviewer.
Coordination mechanisms connect roles across IT, academics, legal, and the units.
For any AI use, you can name the accountable human without a meeting.
When it's missing
Responsibility is assumed to be 'IT's job' or 'the committee's,' so it's no one's.
Systems run with no named owner; when they drift, nobody is watching.
Roles exist on paper but don't connect across silos.
After an incident, the institution can't say who approved the system.
Each practice is something you can build, assign, and evidence, with a first move, an owner, and what proof looks like. Move one practice up one level at a time.
The five levels: 1 Initial · 2 Emerging · 3 Defined · 4 Managed · 5 Optimizing. Each practice describes levels 1, 3, and 5. Level 2, Emerging: some level-3 behavior exists in pockets, but informally. Level 4, Managed: level-3 behavior is applied consistently, measured, and reviewed on a set cadence.
1
Decision-rights & RACI matrix
For each class of AI decision, who is responsible, accountable, consulted, and informed.
1InitialGovernance work is volunteer overtime on top of day jobs.
2EmergingGood practice exists in pockets, but it is informal and has not been institutionalized.
3DefinedKey operational roles are named in position descriptions with allocated time; changes to job duties are worked through labor relations and, where applicable, collective bargaining.
4ManagedThe practice is applied consistently, measured, and reviewed on a defined cadence.
5OptimizingRoles are budgeted, backfilled, and reviewed for adequacy as the portfolio grows.
Next moves
Write AI responsibilities into position descriptions, not side-of-desk assignments.
Name data stewards for the data domains AI systems touch most.
Review staffing adequacy annually against portfolio size.
Bring labor relations in early: where AI changes duties, workload, or monitoring, review job descriptions and meet any bargaining obligations before rollout.
Who owns it
CHRO · Labor relations · CIO · Unit leaders
Evidence it exists
Updated position descriptions
Steward roster
Staffing adequacy review
Labor-relations consultation record
4
Accountability for AI-assisted decisions
When an AI-assisted decision harms someone, it is clear who answers for it and how people appeal.
1InitialResponsibility diffuses to 'the algorithm'; affected people have no recourse.
2EmergingGood practice exists in pockets, but it is informal and has not been institutionalized.
3DefinedAn accountability framework assigns responsibility for AI-assisted decisions, with human review rights and notification standards.
4ManagedThe practice is applied consistently, measured, and reviewed on a defined cadence.
5OptimizingAppeals work in practice, exercised, tracked, and feeding improvements back to system owners.
Next moves
Adopt an accountability framework: a named human answers for every consequential AI-assisted decision.
Guarantee a human-review pathway and tell people when AI was involved in decisions about them.
Build an appeals process with communication templates, and track its use.
Who owns it
Legal · Student affairs · HR
Evidence it exists
Accountability framework
Notification standards
Appeals records
In practice
Worked example
Naming the owner
An early-alert AI flagging at-risk students runs for two years; when its flags are questioned as biased, no one can say who owns it, IT built it, an advising dean requested it, institutional research feeds it. Under this pillar a RACI is drawn: the advising dean is accountable, IR is the data steward, IT the technical owner, and a governance reviewer signs off annually. The model is audited, adjusted, and (most importantly) has a human who answers for it the next time a student asks why they were flagged.
Watch for
Writing a RACI no one consults, it has to live in daily workflows.
Assigning ownership without authority or time to actually exercise it.
Forgetting the lifecycle: someone must own a system in operation, not just at launch.
Readiness & maturity questions
Readiness asks: are we prepared? Use these to surface blind spots before you build. Every “no” shows where the work is. Representative prompts for reflection, not a scored test. For the scored version, use the Maturity Assessment.
Strategic Leadership
Do executives understand the AI landscape?
Is there a clear AI strategic leader?
Are role expectations documented?
Is there a cabinet-level discussion on AI?
Are trustees or regents informed of AI plans?
Are shared-governance structures included?
Are DEI leaders involved in AI planning?
Are leaders trained in AI ethics and use?
Are communications about AI visible and proactive?
Are executives modeling responsible AI behavior?
Operational Execution
Are roles for implementing AI clearly assigned?
Are IT/data teams staffed for AI delivery?
Are project managers familiar with AI systems?
Do we have role clarity across units?
Are there liaisons between tech and functional teams?
Are operational leaders included in design phases?
Is there budget alignment with execution plans?
Are vendor and internal responsibilities clear?
Are AI projects scoped with defined ownership?
Are failures traced back to gaps in roles?
Enabling & Engagement
Are HR and professional development units AI-aware?
Are onboarding processes including AI basics?
Are mentorships or peer programs available?
Are communication teams equipped for AI messaging?
Is IT support trained on AI troubleshooting?
Are faculty development teams engaged in AI pedagogy?
Are students engaged in user research or pilots?
Are community engagement offices aligned with AI plans?
Are marketing and recruitment staff AI-literate?
Are enablers (e.g., librarians, advisors) part of AI decisions?