Campus-wide multi-model AI platform
The university replaces 14 unit-level AI licenses with one platform: several commercial models plus one locally hosted model, through chat and an API. Confidential data may only go to the local model.
Illustrative example. Figures are invented to show the method.
Outcome hypothesis
If we provide one multi-model platform in place of scattered unit licenses and personal accounts, we expect lower total cost, fewer unapproved tools, and measured gains in priority workflows because units share one contract, one set of data protections, and a fitting model for each task, provided each model meets data policy and use reaches real workflows, not just logins.
What was measured
- Platform and model usage: $610,000 a year, usage-based
- Platform team: 2 positions
- Security and data review for each new model
- Retired unit licenses: $340,000 a year no longer spent
- Exit: models can be swapped, but API integrations depend on the platform
- Access: 31,000 accounts
- Monthly active users: 38%, each person counted once
- 22 registered use cases, 9 connected through the API
- 70% of requests go to one general-purpose model
- Unapproved tool use (survey): 41% to 18%
- 3 use cases show measured workflow gains, 4 show no change, 15 not yet measured
- Cost per request differs about 30 times between the cheapest and most expensive models
Lens gates
| Lens | Result | Finding |
|---|---|---|
| Quality | Watch | Quality differs by model and task. It is judged per use case, not as one platform score. |
| Capability and readiness | Watch | 60% of staff have not taken training, and many are unsure which model to choose. |
| Equity and accessibility | Watch | Students and staff get the same access, but use is concentrated in a few graduate programs. Accessibility audit passed. |
| Governance and risk | Pass | Each model reviewed for data terms. Confidential data is routed only to the local model. |
| Evidence quality | Watch | Strong data on cost and reach. Outcome data exists for only 3 of 22 use cases. |
What can be claimed
Budget offset of $340,000 from retired licenses, broader and more even access, and less use of unapproved tools.
That the platform improved teaching, learning, or research overall. Outcomes are known only for the three measured use cases.
Fund the platform for two more years. Standardize the three proven use cases, restrict the least accurate model to drafting, and require a review band and outcome measure for every new use case.
Portfolio review each year. Each use case is reviewed on its own schedule.
A platform is not a use case. Judge it by the portfolio running on it, and fund the measurement that makes that possible.