Most of the attention on the 2027 EDUCAUSE Top 10 went to item one, "Determining where AI adds real value". We wrote about that in a small way to find where AI pays off. This piece is about two items further down that decide whether that search can happen safely: No. 4, "Designing AI structures for responsible innovation", and No. 7, "Building a strong and adaptive data governance backbone" (full list as reported by Campus Technology).
The EDUCAUSE text for item four is blunter than its title. It says institutions "need safe spaces to experiment" with AI, that "boundaries and safeguards must be in place", and that some teams are building "sandboxes, walled gardens, and other secure structures". Item seven adds that the need for trustworthy data "becomes even more urgent" as AI arrives. Neither item tells you what to put in place on Monday. Below is the checklist we use when a university asks us that question.
Our view first
A lot of university AI policies we read are one document about academic integrity and student use. That matters, but it says nothing about the registrar's office pasting a transcript into a chatbot, or a department buying a summarisation tool on a credit card. Governance for staff and administrative use is usually the gap, and it is also where the personal data sits.
The checklist
Know what is already running
- An inventory of AI tools in use. Name, owner, who pays, what data goes in. Expect to find more than the central IT team approved. Start with procurement and expense records, not a survey.
- A one-page intake form for new uses. What is the task, which data, who uses it, how will you know it worked. If the form takes longer than an hour to fill in, people will route around it.
Data before models
- Data classes mapped to where processing is allowed. For example: public (any approved tool), internal (approved tools under contract), personal data (institution-controlled or contractually restricted hosting only), restricted research or health data (no general AI tools). Four rows in a table beat a forty-page policy.
- An owner and a review date for every document set an assistant reads. This is item seven in practice. An assistant will quote last year's fee schedule with full confidence if nobody retires it.
- Access follows identity. The assistant should only answer from documents the person asking is allowed to open, using the same single sign-on groups as everything else. We explained why this cannot be bolted on later in why a general AI assistant can't just read your organisation's documents.
The sandbox, with rules
- A sandbox with written limits. Which data classes are allowed inside, who can join, how long a pilot may run. Synthetic or public data by default.
- Exit criteria agreed before the pilot starts. What result means "scale it", what means "stop", and the date you decide. Pilots without an end date are how campuses end up with dozens of licences nobody can account for.
Vendors and the tools an assistant can call
- Five contract questions for every AI vendor. Is our data used for training? How long are prompts and outputs kept? Where are they processed? Which subprocessors are involved? Can we export and delete everything at the end?
- A list of the connectors and tools each assistant can use. An assistant that can send email or query a database carries the permissions of whatever it is connected to. Our MCP server audit checklist covers what to look for.
Evidence and people
- Logs you can actually review. Question, answer, documents cited, the user's role, with a retention period agreed with the privacy office. Without these, nobody can say whether item one was achieved.
- A small review group that meets monthly and includes a sceptic. EDUCAUSE item nine talks about listening to "skeptics and resistors rather than dismiss them". Putting one on the group that approves uses is the cheapest way to do that.
- A public register of approved uses. One page staff can check before asking "am I allowed to use this?". It also shows students and faculty what the institution has actually decided.
How this lines up with NIST AI RMF
If your institution wants an external framework to point to, the NIST AI Risk Management Framework (released 26 January 2023, voluntary) groups its guidance into four functions: Govern, Map, Measure and Manage. Roughly, items 1, 2, 11 and 12 above are Govern, 3 to 6 and 8 to 9 are Map, 7 and 10 are Measure, and acting on the exit criteria is Manage. NIST notes on the same page that AI RMF 1.0 is being revised, so check for the newer version before you cite section numbers in a policy.
A sensible order
Do not try all twelve at once. Items 1, 3 and 4 can be done in a month by two people and make every later decision easier. Then pick one office, run a pilot under items 6, 7 and 10, and write the register (item 12) from what you learn. The rest can follow the second pilot. And keep the assistant honest about gaps: one that answers only from cited sources and says "I don't know" otherwise is much easier to govern, as we argued in cite or refuse.
FanMind, our document assistant, was built around several of these items: answers come only from uploaded documents with a citation to the page, access follows single sign-on roles, unanswered questions are logged for the content owner, and it can run on the university's own servers. See how FanMind works.
Frequently asked questions
Which EDUCAUSE 2027 Top 10 item covers AI governance?
Item four, "Designing AI structures for responsible innovation". It calls for guardrails, institutional values and standards alongside safe spaces such as sandboxes for experimentation. Item seven, on data governance, is the closely related one.
What should a university AI governance policy include?
At minimum: an inventory of tools in use, data classes mapped to where processing is allowed, owners for the documents AI tools read, access tied to single sign-on roles, contract terms for vendors, reviewable logs and a public register of approved uses.
What is an AI sandbox in higher education?
A controlled environment where staff and faculty can try AI tools with limits written down in advance: which data is allowed, who can join and how long a pilot runs. EDUCAUSE mentions sandboxes and walled gardens, and consortia for institutions that cannot build their own.
Can universities use the NIST AI RMF?
Yes. It is voluntary and not specific to any sector. Its four functions (Govern, Map, Measure, Manage) give a policy a recognised structure. NIST says version 1.0 is under revision, so check the current text first.
