How Universities Can Use AI to Connect People With Institutional Knowledge

Every university runs on knowledge that never made it into a system of record. The associate registrar who knows which exception codes the state auditor actually cares about. The lab manager who remembers why the 2019 grant closeout took eleven months. The advisor who can tell a transfer student, from memory, which three courses will and won’t articulate. This is institutional knowledge, and on most campuses it lives in people’s heads, in shared drives nobody has opened since a reorg, and in email threads that vanished when someone retired.

The cost is not abstract. It shows up as the same question asked forty times a semester, as a six-month onboarding ramp for a new financial aid counselor, as a faculty member rebuilding a syllabus template that already exists in three other departments. For IT leaders, it shows up as ticket volume that never declines no matter how many knowledge base articles the team writes.

AI is the first technology in a while that can plausibly do something about this – not because it is smart, but because it is finally good enough at retrieval to make scattered, badly organized, permission-controlled content usable. The interesting work here is architectural, and it belongs to central IT.

Why Campus Knowledge Is Harder Than Enterprise Knowledge

Application sprawl is not unique to campuses – the average company now manages 305 SaaS applications, according to Zylo’s 2026 SaaS Management Index. What is different in higher education is that the sprawl is structural rather than accidental. 

Sarah Visser of Calvin University described her institution’s starting point to EDUCAUSE Review as “128 different applications,” and Calvin is a small institution; add a research enterprise and a medical center and the count climbs steeply. 

The SIS, the LMS, the ERP, the grants management platform, the IRB system, the housing system, the ITSM tool, a dozen departmental SharePoint sites, a library discovery layer, a research data repository, and whatever the medical school bought independently in 2017.

Three characteristics make this genuinely different from a corporate deployment:

• Permissions are non-negotiable and deeply layered. FERPA governs student records. HIPAA governs anything touching the health system. Export control rules govern some research data. Grant terms govern other research data. Any retrieval system that flattens permissions to make search easier has created a compliance incident rather than a productivity tool.
• Governance is federated by design. Colleges and departments have real autonomy. Central IT often cannot mandate that the School of Nursing migrate its documentation anywhere. A system that requires content consolidation before it delivers value will stall at the first dean who says no.
• The knowledge has seasons. Add/drop deadlines, financial aid disbursement, commencement, and grant submission cycles. Content that is authoritative in September is actively wrong in January, and the failure mode of a confident AI assistant citing last year’s deadline is worse than no assistant at all.

The Architecture That Actually Works

The pattern that has emerged across successful deployments is consistent, and it looks less like a chatbot project than like an integration project with a conversational surface.

Connect rather than migrate. Rather than asking departments to move content, index it where it lives through connectors to existing systems. This sidesteps the political problem entirely – nobody gives up ownership – and it means the index stays current because the source stays current. Platforms purpose-built for this pattern, such as AI platforms designed for higher education environments, ship with several hundred prebuilt connectors precisely because the alternative, writing and maintaining integrations for every campus system, is where these projects go to die.

• Enforce permissions at query time, per user. The index should carry the access control lists from the source system and evaluate them against the person asking. A student and a dean asking the identical question should get different answers, and in both cases the answer should contain nothing the asker could not have found on their own with enough patience. This is the single design decision that determines whether your general counsel signs off.
• Ground every answer in a citation. Retrieval-augmented generation is not just an accuracy technique here, it is an accountability one. When the assistant tells a staff member the tuition refund deadline, it should link to the policy page it read. That link is what lets the user verify, and it is what lets your team notice that the authoritative page is three years stale.
• Treat freshness as a first-class signal. Weight recency, mark superseded documents, and give content owners a way to flag authority. A retrieval system trained to prefer whatever matches keywords best will happily surface a 2021 policy PDF over the current one.
• Nothing in this stack is exotic. It is connectors, an index, an access control layer, a retrieval pipeline, and a model – and the hard parts are the middle three, which is why this is an infrastructure decision rather than a procurement of a chat interface.

Where the Value Shows Up First

The temptation is to launch with a student-facing assistant, because that is the visible win. The better sequence is usually the opposite.

• Start with staff service desks. IT, HR, financial aid, and the registrar’s office answer high volumes of repetitive questions against documentation that already exists and is already reasonably well governed. Deflection is measurable, the content owners are identifiable, and mistakes are caught by professionals rather than by an eighteen-year-old who then acts on bad information.
• Move to onboarding and cross-departmental work. A new hire who can ask “how do we process a late add” and get a cited answer reaches productivity materially faster. The same pattern helps anyone working across unit boundaries – which, in a university, is nearly everyone doing anything interesting.
• Then research administration. Pre-award and post-award staff spend enormous time reconciling sponsor terms, internal policy, and prior submissions. Surfacing prior successful proposals, prior budget justifications, and current sponsor guidance in one place is a concrete time saving for a population that is chronically under-resourced.
• Student-facing last, and narrowly scoped. Once the retrieval layer is proven and the content underneath it has been cleaned up by the earlier phases, extend it to students for a bounded domain – course logistics, campus services, deadlines – with clear escalation to a human.

Governance Is the Project, Not a Phase of It

An AI layer over institutional knowledge does not create governance problems so much as it makes existing ones legible. The moment staff can ask questions in natural language, every contradiction between two departments’ documentation becomes visible, loudly and immediately.

Plan for that. Assign content ownership before launch, not after. Build a feedback loop where a wrong answer routes to the owner of the source document. Log queries and review the ones that fail, because unanswered questions are the most honest documentation gap analysis your institution will ever get.

You also need a risk posture that survives a board conversation. Model behavior monitoring, data residency, audit trails, and a documented position on what the system will and will not be used for – admissions decisions and student conduct being two obvious exclusions. Teams formalizing this often start from established frameworks and tooling; DevOpsSchool’s rundown of AI risk management tools and frameworks is a reasonable starting point for mapping NIST AI RMF and EU AI Act obligations onto an actual control set.

The Real Measure

Success here is not chat sessions or adoption percentages. It is whether a new financial aid counselor is competent in six weeks instead of six months, whether the registrar’s office answers the same question fewer times, and whether the knowledge that used to leave with a retiring staff member now stays behind in a form someone else can find.

That is an unglamorous outcome. It is also the one that compounds, and it is squarely an IT deliverable – which makes it worth building deliberately rather than buying as a demo.

Maria