In Higher Ed, AI Success Hinges On The Data Foundation Beneath It

Higher education has spent years attempting to license its way out of friction. A process gets painful, a new SaaS product gets tacked on to fix that one part, and the data starts moving between systems on connections nobody designed for reliability. We end up with an institution running on brittle integrations and competing definitions of facts, and that is the ground we want an AI agent to stand on. Automating on top of it doesn't clear the underlying mess. It runs the mess faster, and at a higher bill.

Patricia Clay is the Associate Vice President for Information Technology and CIO at Hudson County Community College, where she sits on the president's cabinet and leads infrastructure, information security, and enterprise systems for the institution. She started the college's information security program and has spent more than two decades in higher-education IT, including years administering ERP systems and leading technology at DeSales University. Her read on what AI can and can't do for a college starts well before the model.

"All of it goes back to fundamentals," Clay said. "If we don't know where our data lives, how we're protecting it, and who has access to it, we're always just setting ourselves up for failure. And with AI systems, we can get to that failure a lot faster and more expensively." An agent is only as good as the data and the infrastructure that power it, and whether an automation pays off is often decided long before anyone picks a model. Clay's point is that most institutions haven't done that groundwork, and that rushing the automation layer only makes the gap more expensive to discover.

• The rickety middle: Clay described the pattern that produced the problem. "This particular part of the process is painful, so let's license the SaaS product that handles it best," she said. "But in the background, I'm pulling data from here and pushing it to there, and the bridges between them are rickety. And it does break down." The individual tools work fine on their own. It's the connective tissue between them, assembled one urgent purchase at a time, that fails under load. A pile of point-to-point bridges leaves an institution without the control plane an agent has to act through. Before an institution commits to new agents, Clay wants the same scrutiny applied to tools it already owns, and to the build-versus-buy calls that keep adding to the pile. "Why do I need three different CRMs?" she asked. "Or did it just seem like a good idea at the time?"

• The number is not the number: The deeper issue sits in the definitions, which is why the Hudson team is building a data warehouse. What counts as a first-time full-time student, or an employee, turns out to be several different answers depending on who's asking and why. Grant-funded versus operationally funded staff, a part-time student counted as one head or as a fraction, continuing-education students counted or not, these decisions change the story. The technical definition is generally easy. The collaboration to agree on it is not, and it needs to happen before agents start surfacing numbers that nobody can reconcile with the standard reporting definitions. "Anyone who's ever written a report can tell you there's no such thing as 'the number is the number,'" Clay said. "It depends on what you asked me for and what the purpose was. That doesn't make the number bad. It means we have to understand the definition."

The same fundamental gap appears in both what these systems hand back and in the trust of the output. A clean-looking report or a fluent paragraph reads as authoritative, and the human instinct is to accept it. Clay has watched that instinct turn into demands for updated institutional policy in the one area where higher ed reacted fastest: cheating. The initial response leaned on tools that promised certainty, but couldn't deliver it, which is its own argument for verifying what a model returns rather than trusting the score on the screen.

• Judgment over detection: Hudson's administration has told faculty plainly not to lean on AI-detection scores. "Do not accuse people only based on that number," she said. "You have to use your professional judgment, because using an AI tool to detect an AI tool is problematic." The scores look precise and are anything but, a problem compounded by research showing detectors produce false positives on human writing, especially from non-native English speakers. And even a correct detection tells you almost nothing. As Clay put it, catching AI text doesn't reveal whether it was used to polish after extensive drafting or to generate the whole thing five minutes before a deadline. She is looking toward a skill older than the technology: critical thinking. Where did the artifact come from, and how do you know it's correct?

• Augment, don't substitute: Where AI earns its place, in Clay's view, is in clearing the routine so people can do the human part. She leans on the 80/20 rule. "If we can make the 80% of transactions that just generally happen go smoothly, then our people can spend the time they need on the 20%," she said. The students who need the 20% need a person. Clay's example was a student who walks in ready to drop out because they think they can't pay, when the real issue is they never finished filing the FAFSA, or didn't know they could spread the term out instead of dropping below full-time. "Especially our younger students are really craving human support," she said. Automation that guides someone through a form in the moment, she added, and doesn't pretend to be human while doing so, is a game-changer.

• What can we actually afford: Cost discipline is at the crux of it, and Clay sees the bill climbing as the systems get more capable. "Before we rush to 'we need this thing,'" she said, "let's think about what we already have that might do it." The question she keeps returning to is whether an institution is holding itself to a reasonable number of tools that its staff and budget can actually manage. This discipline matters more as the price of running these systems rises. Agents can't solve that math on their own. If anything, she expects them to surface new problems that then need their own attention and their own spend. "AI isn't going to magically fix everything," she said, "especially not as we're eventually going to pay for all this very expensive compute."

All of this is not AI resistance to Clay. She is co-leading a responsible-AI governance council and has her own system instructions tuned to push back on her rather than flatter her. The order of operations is critical. Lock down the data, understand the definitions, consolidate what you run it on into a single control plane, keep human judgment in the loop, and then the automation has something solid to stand on.

She closed on where technology cannot reach. "The people you remember from college aren't the ones who handed you a transcript or a bill," Clay said. "You remember the person who took the extra time with you when you were struggling. I don't think an AI is ever going to replace human connection."