The approaches illustrate the central challenge for higher education CIOs: controlling costs and protecting institutional data without imposing a centralized approval process that discourages experimentation.
Consumption Pricing Changes the Budget Equation
Brandon Rich, director of AI enablement at the University of Notre Dame, says the university centrally provides Google Gemini and permits the use of Microsoft Copilot for users of the company’s productivity suite. It also offers ChatGPT Edu through a chargeback model and supports GitHub Copilot.
“Our goal is to direct people to sanctioned platforms because we know that we’ve vetted them for data privacy and security,” he says.
Fordham University, meanwhile, has adopted a three-tiered model that gives the entire university access to a general-purpose option, with more detailed review and governance as users request more advanced models.
The approach gives Fordham visibility into approximately 90% of its AI spending, according to Anand Padmanabhan, vice president and CIO at the university.
READ MORE: Token economics should shape artificial intelligence strategy.
“Our goal is to make experimentation easy, provide a clear idea of how much the university is spending on AI and clarify for the community how they can get what they need,” Padmanabhan says.
Notre Dame encountered that transition when renewing its ChatGPT Edu agreement. Rather than expose users directly to token-based billing, the university retained a flat, outward-facing price while closely monitoring consumption under the new model. It is also generally approving requests for higher quotas while it establishes a realistic usage baseline.
“We want to understand what the real usage looks like once the subsidies are out of the way,” Rich says.
Build Guardrails Around Users and Workloads
One of the key capabilities universities need is a mechanism that prevents unexpected bills while accommodating legitimate differences in demand. An instructor teaching agentic coding may require more capacity than an employee summarizing documents, while researchers may need premium models or large context windows. Routine administrative tasks can often use less expensive models.
Fordham’s governance framework reflects those differences among faculty, researchers, students and administrators.
The university establishes non-negotiable requirements around data security, privacy, compliance, major financial commitments and decision-making, but leaves room for experimentation within those boundaries. It also provides students with an AI tool when a course requires it.
“Centralize the guardrails, not the ideas,” Padmanabhan advises. “For students, if a course requires an AI tool, we provide it for them.”
Institutions also need contract terms that make consumption visible and controllable. Padmanabhan says CIOs should determine exactly what triggers a charge, including whether models consume tokens or credits at different rates and how multiple agent actions are billed behind a single request.
Vendor reporting should show consumption by application, organizational unit and workload where appropriate. Contracts should also give institutions access to thresholds, alerts, rate limits, spending caps, model restrictions and the ability to pause a workload before costs escalate.
Rich says Notre Dame is considering identity management groups to distinguish power users from those with ordinary requirements, allowing it to assign different limits and potentially different prices.
LEARN MORE: Establishing governance over artificial intelligence systems keeps higher ed secure.
Make Cost Awareness Part of AI Literacy
Universities can reduce spending by matching each task to an appropriately priced model. The newest and most expensive option is not automatically necessary for summarization, brainstorming or routine coding assistance.
Fordham requires training before allocating a pro or enterprise license. Those sessions emphasize using lower-cost models for routine workloads and reserving more capable models for tasks that warrant their added expense.
“AI model selection is a financial architecture decision, not just a technology preference,” Padmanabhan says. “We should move away from the assumption that every AI request needs the most powerful model available.”
UP NEXT: Data literacy is key to artificial intelligence ROI in higher education.
Cost controls should also reflect where decisions are best made. Fordham shifts responsibility for AI services and tools that are not intended for universitywide use to individual schools or departments, which often have a clearer understanding of their own use cases than a central IT organization.
Institutions must then assess whether that spending produces meaningful academic or operational value. Administrative measures could include cycle time, service quality, error reduction, staff capacity and user satisfaction.
Padmanabhan explains that student-facing tools call for attention to engagement, accessibility and outcomes, while research workloads require measures such as speed of analysis, computational efficiency and research productivity.
“Time savings matter, but universities are not factories, and not every hour saved translates into a budget reduction,” he says. “Sometimes, the more important question is what we do with the capacity that AI creates.”
