AI for Universities (2026): How to Roll It Out to Staff, With ChatGPT Edu, Claude, Gemini and Copilot Side by Side

How should a university roll out AI?

Roll AI out to staff the way you'd roll out any system that touches student records: one approved assistant on an institutional contract, a written line on what data never goes in, and a pilot in two or three administrative offices before anyone signs a campus-wide deal. Start where the writing volume is highest, usually admissions, advancement and the registrar. Train those teams on their own work for an hour, measure a month, then widen. As of 2026, the tool choice matters less than the order you do things in.

Most "AI in higher education" writing is about students and cheating. That's a real fight, and it's not this page. This one is for the provost, the CIO and the office heads who have to decide what the people running the institution are allowed to use on Monday.

And they are already using something. If there's no approved tool, the enrollment counselor with 400 unanswered emails is pasting them into a personal account. That's the risk you're actually managing.

What do the four vendor education programs include?

Every vendor has an education program now. Nobody lines them up, partly because two of them won't tell you a price. Here's what each vendor's own page said when I read them on October 7, 2026:

A few things worth noticing. ChatGPT Edu leads on admin controls: SSO, SCIM, custom data retention windows and a usage analytics dashboard, plus SOC 2 Type 2 and four ISO certifications listed right on the page. Anthropic's page names administrative work directly ("drafting policy and onboarding communications to reviewing transcripts and building budgets"). Google and Microsoft both hand staff a basic assistant for free if you already run their suite, and charge for the version that lives inside email and documents.

One trap for small colleges: if you were hoping for a nonprofit discount, Anthropic's help center lists "higher education institutions" as not eligible for its nonprofit pricing. The general nonprofit playbook in the AI for nonprofits guide still applies to how you roll out. The price route just runs through the education program instead.

What does it cost to give 150 administrative staff an assistant?

Here's the worked math for a 150-person administrative staff (not students, not faculty), using only the numbers each vendor publishes:

Two footnotes. Claude Team is the self-serve fallback if you don't want a sales process, and it stops at exactly 150 seats, so a 150-person staff fills it to the brim with no room for the 151st hire. And Education Plus has its own $6 per user per year cost, which I've left out because your Google rep will tell you how many users it has to cover.

The honest reading of that table: the gap between $0 and $36,000 a year is the gap between "an assistant in a browser tab" and "an assistant inside Outlook or Gmail that can read your files." Start staff on the free tier you already own, find out which offices actually live in it, and buy the paid seats for them. Not for all 150 on day one.

What does FERPA mean for staff using AI?

Here's the uncomfortable finding. I searched all four vendor education pages for the word FERPA on October 7, 2026. Zero hits on every one. What they promise is "we don't train on your data." That's a good promise. It is not a FERPA answer.

FERPA lets an institution share education records with an outside vendor without consent only when that vendor qualifies as a "school official." The regulation, 34 CFR 99.31(a)(1)(i)(B), sets three conditions. The vendor has to perform a service you'd otherwise use employees for. It has to be "under the direct control of the agency or institution with respect to the use and maintenance of education records." And it has to follow the rules on redisclosure.

That "direct control" phrase is a contract question. So, in plain terms:

• A personal account is never under your direct control. Student records in a free consumer login are the problem, whatever the vendor's training policy says.
• An institutional contract can be, if its data protection terms say so. Get the vendor's education or enterprise terms in writing and have counsel read them against the school-official test before any student record goes in.
• Until counsel signs off, keep identifiable student records out entirely. Most of the useful admin work in the next section doesn't need them.

Write that line down. A one-page AI acceptable use policy with a "student records" row covers most of it. The bigger questions (who approves new tools, who owns the vendor review, what gets audited) belong in an AI governance framework, and on most campuses that's a joint job for the CIO and general counsel.

Which three offices should start first?

Pick offices with high writing volume, repeatable questions and a director who wants it. Here's where I'd start, with one thing each office can try this week.

1. Admissions and enrollment: the inbox

Enrollment teams answer the same twenty questions about deadlines, transcripts and deposits all cycle long. Micro-action: collect the office's ten most-sent replies into one document, load it into the approved assistant, and have counselors draft from it. The counselor still reads and sends every reply. Use first names and the question only, no application details, until the FERPA review is done.

2. Advancement: donor letters and call prep

Gift officers write thank-you letters, stewardship reports and briefing notes before every visit. Micro-action: take one anonymized stewardship letter that worked and ask the assistant for three versions for different gift levels in your president's voice. Then time how long the next ten letters take against the last ten. Donor files stay in the CRM. The assistant gets the template and the talking points, not the giving history.

3. The registrar: policy Q&A

The registrar's office is a policy library with a phone attached. Students and staff ask the same questions about add/drop dates, transcript holds and grade appeals, and the answers live in a catalog nobody reads. Micro-action: load the public academic catalog and the published calendar into a shared project and have front-desk staff check their answers against it, with the page citation, before replying. This is public material, so it's the safest place on campus to start.

Notice what's missing: grading, admissions decisions and anything that judges a person. Those stay human. AI drafts, a person approves.

How do you get staff to actually use it?

A license and a login link get you a handful of enthusiasts and an awkward renewal meeting. What works is an hour per office on that office's real work: their emails, their letters, their policy questions, with the data line explained by someone who can answer "can I paste this?" on the spot. Then a check-in after two weeks to see what stuck.

That's the same shape as any good AI training for employees program. Universities just have more offices and more committees between the decision and the first login.

Who can help a university roll out AI to staff?

I work with leadership teams at organizations of roughly 10 to 200 staff, nonprofits and associations included, on exactly this: picking the tool, writing the data line, and getting the first offices using it. For a college, that's the administrative side, the president's cabinet and the office heads who report to it. The paid path starts with my AI Readiness Assessment: $2,500 flat, two weeks, a 15 to 25 page written roadmap and a one-hour walkthrough. A larger institution with many departments gets a written custom scope after the Assessment, priced to the work. If you'd rather talk first, book a free 30-minute strategy call. It's the same approach I describe in AI for associations, which have the same mix of small staff, many stakeholders and member data to protect.

What I'd do this week

Ask IT one question: which AI assistant are staff already licensed for? Odds are good it's Copilot Chat or Gemini at no extra cost. Then ask general counsel to read that vendor's education terms against the school-official test, write the one-page data line, and pick the three offices. The paid seats can wait a month. The policy can't.

Sources

• OpenAI, ChatGPT Edu: features, data privacy, certifications, contact sales (read October 7, 2026)
• OpenAI, Introducing ChatGPT Edu: training policy, no published price (read October 7, 2026)
• Anthropic, Claude for higher education: university-wide plan, data use, administrative uses (read October 7, 2026)
• Claude pricing: Education plan "discounted rates," Team for 2 to 150 seats at $20 annual / $25 monthly (read October 7, 2026)
• Claude Help Center, Getting started with Claude for Nonprofits: higher education not eligible (read October 7, 2026)
• Google for Education, Compare editions: Fundamentals no cost, Education Plus $6 per user per year, Gemini data terms (read October 7, 2026)
• Google for Education, Google AI Pro for Education: $20 / $24 / $15 pricing and data terms (read October 7, 2026)
• Google for Education, Gemini for Education: no cost for qualifying institutions (read October 7, 2026)
• Microsoft, Copilot in education: $18 per user per month academic offering, Copilot Chat on A1/A3/A5 (read October 7, 2026)
• 34 CFR 99.31, FERPA regulations (eCFR): the school-official outsourcing conditions (read October 7, 2026)