Can your AI actually explain why it gave you that answer, using your institution’s own definitions, not its best guess at them?
A few years ago, I asked ChatGPT to help me research a segment of the Australian higher education sector. What came back was detailed, specific, and completely wrong. It invented a technical university network — member institutions, a head office address, even a mission statement — and presented all of it with total confidence.
Nothing about the answer looked fabricated. That was the problem.
I’ve been fascinated by AI since I was introduced to it at university back in 1986. When ChatGPT arrived, I jumped in immediately — for creative work, it was genuinely stunning. But the moment I pointed it at something that required real institutional knowledge, it fell over, and it didn’t fall over quietly. It fell over confidently.
That distinction matters a lot more when the question isn’t “write me a story” but “which programs are running at a loss, and why.”
The LLM Misconception
Every vendor conversation about AI in university finance right now tends to skip straight to the exciting part: ask a question in plain English, get an instant answer, no multi-day or multi-week analysis required. And that future is real. But there’s a step before it that most of these conversations quietly skip over.
A large language model is extraordinarily good at language. It is not, by default, good at knowing what your institution means by “course” whether your FTE calculation includes casual academic staff, or how your central overhead gets split across teaching and research. Point an LLM at a data warehouse without that context, and it won’t refuse to answer. It will answer, using its best guess at what your numbers mean, and it will sound just as confident doing that as it did inventing a university network that never existed.
For a chatbot, that’s a mildly embarrassing failure mode. For a financial decision, it could lead to multi-million-dollar mistakes.
Two Foundations, Not One Tool
We’ve spent 27 years building causal cost models for a range of large, complex organizations including military, government, financial services and for the last 20 years higher education in Australia, US and Canada. For the last few years, ever since the release of ChatGPT, we’ve been researching how best to maximize the benefits of AI and minimize the risks. This requires two pieces of groundwork first.
A causal model – an explicit map of how money and effort actually move through the institution. Not broad averages or arbitrary splits, but real cause-and-effect: which resources support which activities, and which activities support which programs, research, and community outputs. Without this, an AI system has no foundation to reason from, it’s working from a spreadsheet’s-worth of disconnected numbers rather than an understanding of how the institution operates.
A semantic model – the institutional dictionary that tells the AI what your terms actually mean. What counts as a valid FTE. Which organisational hierarchy is the authoritative one. What is a “program”. Without this layer, the AI has to infer relationships statistically or linguistically and different campuses, different revenue rules, and multiple definitions of “student load” mean there usually isn’t one obvious right answer for it to infer.
Skip either one, and you get an AI that answers fast and wrong instead of slow and right. Neither is actually useful to a CFO.
What This Looks Like in Practice
The good news is that most of this groundwork isn’t exotic. It’s the unglamorous, entirely solvable stuff: are your department codes consistent between Finance and Student Systems? Can you trace a specific line in the General Ledger to the teaching activity it actually supports? Is salary data sitting against the right organisational unit, or parked in a central holding account?
In our experience, most institutions have real gaps here and that’s normal, not a red flag. It’s also exactly the kind of gap that’s straightforward to close once you know where to look.
Where to Start
We’ve written up the full picture, the two-layer problem above, the most common data issues we see across institutions (and how to fix them), and a self-assessment checklist you can run against your own systems before you spend a dollar on AI tooling, in a new whitepaper: Preparing for Enterprise AI for University Financial Decision Management.
It’s written for CFOs and finance leaders who want the “will this survive contact with our General Ledger” version of this conversation, not the demo-day version.