The number nobody could defend
At a EDUCAUSE 2026 session called Top 10 Topics in Data Governance, four university data officers described an incident that has become a cautionary tale in higher ed IT circles this fall. An AI chatbot at Rowan University was asked a simple question, show me enrollment, and it answered with a number near 600,000 students. Rowan is a mid-sized New Jersey university with nowhere close to that many students enrolled in any given year, and the answer was obviously, immediately wrong to anyone who looked at it, which is itself part of why the story traveled so far at the conference.
The chatbot had not malfunctioned in any technical sense. It had simply counted enrollment across every academic year in its reach and added the totals together, because nothing in the system told it that enrollment meant current, active, this-term students rather than a historical sum stretching back years. The word enrollment, it turned out, carried different meanings in the registrar's office, financial aid, and institutional research, and the AI inherited all of those conflicting meanings at once, with no mechanism for flagging the ambiguity before producing a confident answer.
The survey behind the anecdote
The Rowan example landed with the EDUCAUSE audience because it matched what the organization's own data governance community group had just spent a year documenting across its membership. A survey of 65 member institutions found that 27 respondents said data roles were unclear at their institution, and 18 said no one was explicitly in charge of monitoring data quality at all. Sixty five percent said they had restructured their data governance function in the past three years, which reads less as a sign of steady progress than as a sign that the first attempt at a fix rarely holds for long.
Those numbers describe an unglamorous but familiar condition inside large institutions. Most have data, dashboards, and increasingly AI tools layered on top of a governance structure nobody fully owns end to end. The gap between having data and actually governing it has existed for years, quietly, inside spreadsheets and static reports. What has changed is that AI assistants now expose that exact same gap directly to end users, in the form of a confidently wrong answer delivered instantly, rather than a quietly inconsistent report that an analyst might catch before anyone outside the office sees it.
Three institutions, three different starting points
Kelli Van Wasshenova of Denison University described forming a cross-functional data governance team spanning operations, finance, IT, and institutional research, essentially building the org chart that should have existed before anyone asked an AI tool to query enrollment in the first place. Edlira Stefani of Bentley University took a different route, embedding governance policies directly into systems and workflows so staff do not have to recall a rulebook before touching a record, closing the gap between written policy and daily practice at the point where errors actually get introduced.
Ivana Fredotovic of Bowling Green State University framed the whole exercise as a change management challenge rather than a purely technical one, and her open question to the room, how do you get buy-in from the entire institution to do that, is the one every CIO eventually has to answer. Data governance policy that lives in a PDF nobody reads does not stop an AI assistant from aggregating across terms it was never told to separate, no matter how well designed the policy document itself turns out to be.
Why AI turns a known problem into an urgent one
Data governance debt has always been present in large institutions, but it used to surface slowly, in a mismatched report here or a reconciliation headache there, usually caught by someone with enough context to notice the number looked off. An AI assistant removes that friction entirely from the process. It answers instantly, it answers with total confidence regardless of the underlying ambiguity, and it answers at a scale no single analyst could reach in a week, which means an undefined term now propagates to every user who asks a related question rather than staying contained to one bad query from one confused employee.
The EDUCAUSE presenters were explicit that governance guidance now has to extend to AI tools specifically: defining acceptable error thresholds before deployment, deciding clearly who has authority to fix a record when it turns out to be wrong, and setting access rules for exactly which systems an AI assistant is even allowed to query in the first place. None of that is technically AI-specific work. It is the same ownership and shared-vocabulary work institutions have deferred for a decade, except the deadline now arrives whenever the next AI chatbot quietly goes live on campus.
The maturity curve EDUCAUSE mapped
The community group organized the top data governance challenges it surfaced into four stages: foundation, guidance, operations, and impact. Foundation means building a shared vocabulary and formally assigning responsibility for data, which is exactly where Rowan's enrollment problem originated and where most institutions in the survey admitted they were still stuck. Guidance means setting concrete data quality standards, defining who is authorized to fix errors once found, and establishing access rules that explicitly cover AI tools alongside human users, not as an afterthought bolted on once a chatbot is already live.
Operations means embedding those policies into daily workflows so compliance does not depend on any individual remembering a rule under time pressure, which is Stefani's point at Bentley in practice. Impact means an institution can demonstrate that employees find trusted data quickly, that measurable data risk has actually dropped over time, and that leaders trust the numbers behind their decisions enough to act on them without demanding a second, manual verification step before every meaningful decision gets made.
What this means for the CTO's roadmap
The practical lesson for any CTO or CIO evaluating an AI assistant for institutional data, in higher ed or any other regulated sector, is sequencing. Buying or building the AI layer before resolving who owns the underlying data and what its core terms actually mean is building directly on sand, and the Rowan incident is simply what it looks like when that sand shifts in front of a real end user instead of staying hidden inside a back-office report that nobody outside the department ever sees.
That argues for a specific discipline before any AI rollout that touches institutional records: inventory the handful of terms that matter most to decision-making, assign explicit, named ownership for each one, and write down the acceptable error threshold before granting any AI tool query access to production data. It is slower than simply flipping on a chatbot and announcing a pilot. It is also the difference between an assistant that genuinely saves time and one that has to be walked back publicly after it tells a provost the school somehow has 600,000 students enrolled.
