Universities and AI: Managing Costs Without Limiting Usage

With the rise of generative AI and the advent of consumption-based pricing, several campuses are reevaluating their financial management. Fordham, Notre Dame, and UNLV are combining sanctioned tools, safeguards, contractual clauses, and training to prevent overspending while allowing for experimentation.

Gaining Visibility on Consumption-Based Spending

At Fordham University, a three-tier model allows the IT department to track about 90% of AI spending, according to Anand Padmanabhan. At Notre Dame, the renewal of ChatGPT Edu has led to a fixed price while monitoring consumption in the new consumption-based model, thus avoiding direct exposure of users to token billing. The institution offers ChatGPT Edu through backcharging, provides Google Gemini, allows Microsoft Copilot for productivity suite users, and supports GitHub Copilot. Requests for higher quotas are generally accepted while establishing a realistic usage baseline. Brandon Rich indicates a desire to understand actual usage once subsidies are removed. Notre Dame is also considering distinguishing advanced users from ordinary usage through identity groups, in order to apply limits and potentially differentiated pricing.

Establishing Safeguards Based on Profiles and Loads

Universities are implementing mechanisms to avoid unexpected bills while taking into account the differing needs between courses, research, and administrative functions. A professor teaching agentic coding may require capabilities beyond those of an agent dedicated to document synthesis, while researchers may seek high-end models or large context windows, whereas routine administrative tasks may suffice with more economical models. The system established at Fordham takes these distinctions into account among teachers, researchers, students, and administrators, imposing strict criteria regarding data security, privacy, compliance, significant financial commitments, and decision-making, while maintaining space for experimentation. Students have access to the tools required for their curriculum, and Anand Padmanabhan summarizes the philosophy: “centralize the safeguards, not the ideas.” At Notre Dame, Brandon Rich explains the focus on sanctioned platforms that are verified for data protection and security.

Negotiating Contracts that Make Consumption Manageable

Institutions are seeking contractual clauses that make AI consumption visible and controllable. Anand Padmanabhan believes that CIOs must understand precisely what triggers billing, including differences in token or credit consumption rates and how multiple agent actions can be billed behind a single request. Providers are expected to deliver reports detailing consumption by application, by organizational unit, and by workload where applicable. Contracts should include thresholds, alerts, bandwidth limits, spending caps, restrictions on models, and the ability to halt a workload before costs escalate.

Anchoring Model Efficiency through Training and Usage

Reducing costs involves aligning each task with the chosen model: the latest and most expensive solution is not always required for synthesis, brainstorming, or routine coding assistance. Fordham mandates training before granting a professional or enterprise license, emphasizing the use of economical models for routine workloads and reserving high-performance models when the additional cost is justified. For Anand Padmanabhan, choosing an AI model is a financial architecture decision, and it is essential to move away from the notion that every request requires the most powerful model.

Decentralizing Certain Decisions and Measuring Created Value

Cost controls are placed at the relevant level. Fordham entrusts the management of non-shared AI tools and services to schools or departments, which generally have a better understanding of their specific needs than a centralized IT department. Institutions are encouraged to assess the academic or operational value of investments, relying on administrative indicators such as cycle time, service quality, error reduction, team capacity, and user satisfaction. For student-facing solutions, the focus is on engagement, accessibility, and outcomes, while for research, speed of analysis, computational efficiency, and scientific productivity take precedence. Anand Padmanabhan emphasizes that time savings do not automatically translate into budget reductions, and decisions must be made regarding the use of capacity created by AI.

Diversifying Campus Offerings to Frame the Rise of AI

As access to generative AI expands and pricing evolves towards consumption, universities are seeking to preserve pedagogical autonomy, support intensive research, and open tools to students while maintaining budget visibility. The proliferation of use cases exposes institutions to risks of unpredictable spending due to prompt practices, redundant processes, and the systematic use of premium models. Institutions are combining validated tools, access restrictions, and coordinated organization at the campus level. UNLV provides free or department-funded services with varying levels of protection, Notre Dame organizes its initiatives through AI@ND, and Fordham structures access according to a three-tier model. Anand Padmanabhan notes that the goal is to enable experimentation while clarifying spending and access pathways to tools.