Where campus data becomes evidence
Institutional research turns campus data into the numbers that leaders, accreditors, and the public rely on. AI opens new analysis, such as joining scattered data that was never worth the analyst time before. It also adds risk: conclusions drawn from data nobody has defined or checked.
A practitioner orientation to adapt to your office, not a mandate. IR, assessment, and institutional effectiveness are organized differently across campuses; use this to frame IR's role in the institution's AI work.
IR's distinct AI mandate
Six places where AI depends on what IR does, and where IR is best placed to lead.
Data definitions & one version of the truth
An AI tool that counts “retained students” differently from the official definition produces a second, conflicting number. IR owns the definitions AI tools should use.
Analytics governance
Who may run which models on student data, for what purpose, and with what review. IR brings the methodological judgment that access controls alone don't provide.
Survey & outcomes data
Survey responses come with consent terms and confidentiality promises. Putting them into AI tools has to honor both, including small-cell suppression.
Predictive models & fairness
Retention-risk scores and similar models can encode bias. IR can test them for disparate impact before they shape who gets outreach.
Measuring AI's own impact
Someone has to show whether an AI initiative worked. IR is the natural evaluator, and the Compass's efficacy measures give it a place in the scoring.
Reporting & accreditation evidence
Figures reported externally must be reproducible. If AI helped produce them, document the method, the data version, and the checks.
What AI opens up, and what keeps it honest
AI makes analysis possible that used to cost too much analyst time. The checks that make IR's numbers trusted still apply.
- Joining data across systems to answer longitudinal questions, such as what happens to students after they join a course waitlist
- Coding thousands of open-ended survey responses
- Drafting data dictionaries and documentation
- Turning around routine ad hoc requests faster
- Check AI-produced figures against known, published numbers before release
- Apply small-cell suppression and watch for re-identification
- Record the model, prompt, and data version so results can be reproduced
- Keep identifiable data out of tools that haven't been vetted
AI can speed up the analysis. IR still decides whether a result is valid enough to publish.
Questions for analytics and AI vendors
Analytics platforms increasingly ship with AI built in. Ask these before relying on their output.
A score or forecast you can't explain is hard to defend to a dean, an accreditor, or a student. Ask for documentation of features, training data, and validation results.
Student and survey data should not improve a vendor's product without explicit agreement. Get a contractual answer.
Dashboards that slice by many dimensions can expose individuals. Confirm suppression rules match your own.
If the model changes underneath you, last year's figures may not match. Ask how versions are tracked.
IR and IT: different layers of the same stack
Both work with institutional data, but they are responsible for different things. AI makes that division matter more.
Infrastructure, integration, identity and access, security, and whether a tool can connect to the data warehouse safely.
Definitions, analytic methods, interpretation, and official reporting, including whether an AI tool's output is valid enough to use.
Data classification, the data governance council, and sign-off when an AI tool touches the warehouse or official figures.