Building an AI Cell Inside a University

The vice-chancellor announced an AI cell on the foundation-day dais. By Monday it had a logo, a WhatsApp group, and three vendors in the inbox. It did not have a charter, a list of systems of record, or the authority to say no. Six months later the cell had sponsored a chatbot that sat on the homepage and a research GPU that accounts thought was the same project. They were not the same project. One was a fiduciary problem. The other was a lab.

This is a playbook for vice-chancellors, registrars and syndicates who want a cell that can survive a CAG paragraph and a student complaint. It is not a research-centre design. Faculty working on machine learning are a department. An AI cell is operations: which agents exist, on which data, with which human verbs, under which processor contract.

UGC, AICTE and NAAC have not, that we will invent, prescribed the org chart of a campus AI cell. MeitY's India AI Governance Guidelines of November 2025 are a national conversation, not a university regulation. Use them as orientation. Do not cite them as if they appointed your deputy registrar.

Charter before logo

The charter is one page. Purpose: inventory, approve, refuse, and review agent workflows that process university data. Out of scope: marking student homework, running a commercial start-up, and replacing the computer centre. Authority: can halt a pilot that lacks a data map. Reporting: to the registrar or a pro-vice-chancellor, not to a vendor steering committee.

Membership is functional, not ornamental. Registrar or delegate (fiduciary and file). Computer-centre head (custody and identity). Controller or delegate (exam store is sacred). Finance (money workflows). Counsel or DPO (DPDP). Student-welfare (minors, hostels). IQAC (so NAAC evidence does not become a second census). One faculty member who actually builds models, to translate, not to veto operations with a paper.

A student representative is useful for design reviews. They are not a substitute for the equal-opportunity cell on first-generation and disability questions.

If a chair is missing, the cell will approve whatever the loudest vendor brought.

Chair	Owns	Can stop
Registrar / PVC	Charter, syndicate notes, purchase file	Any pilot without a purpose map
Computer centre	Identity, logs, isolation, exit	Anything that phones home or shares a god credential
Counsel / DPO	Basis, minors, processors, notices	Training on customer data, casual transfers
Controller	Marks and unfair means	Helpdesk agents that can see tabulation
Finance	Fee, grant, procurement money	Write-off verbs and unofficial APIs into banks
Student welfare	Hostels, discipline, first-gen design	Risk scores and ragging mediation

Inventory is the first deliverable

Week one is not a hackathon. Week one is a list: every chatbot, every Copilot, every 'temporary' OpenAI key a faculty member put on a lab server that now has student essays. Shadow AI is the estate. The cell that does not look will be surprised in a breach.

Each row: owner, vendor, data classes, training yes/no, where it runs, who can halt it. Rows without an owner are halted. That single rule creates the cell's authority faster than a brochure.

Split research compute from operational agents in the inventory. A GPU for a CS department paper is not a licence to run a placement ranker on the same box with the same accounts.

How the cell says no without becoming a parking lot

Publish a two-page intake: purpose, data classes, verbs, human owner, success metric that is not 'engagement'. Meet fortnightly in the first term. Decisions are approve-with-conditions, refuse, or send to syndicate if the risk is novel.

Time-box. A cell that takes three months to approve a hostel ticket matcher will be bypassed. A cell that approves a proctoring gaze model in a day will be in the newspaper. SLA the dull yes. Slow the accusation tools.

NAAC and NIRF will tempt the cell to become a metric factory. Refuse. Quality cells already exist. The AI cell should not harvest personal data to polish a rank.

What to buy the cell — and what not to

Buy time: a full-time coordinator who can write notes. Buy a room in the data centre, not a SaaS 'AI OS for campuses' that wants every store. Buy training for clerks, not a foreign study tour for the logo.

Prefer on-prem processors who will accept a training ban and DPIIT-aware procurement where your rules care. Prcept is one such B2G option. We are not the cell. If a vendor offers to chair your cell, they are not a vendor. They are a capture.

How the cell talks to IQAC and the syndicate

Quarterly, table three objects: the inventory, the list of refusals, and one sampled packet from a live workflow. Syndicates understand refusals. They should see that the cell said no to a ranker. If the only papers they see are logos and hackathon photos, they will fund the next poet-bot.

IQAC may ask for a paragraph on 'innovation'. Give them the hostel SLA graph or the fee orphan count. Do not give them a student-level extract. Quality language is not a licence to widen a purpose. If NAAC's live template asks for a specific indicator, the custodian supplies that indicator. The cell does not become a warehouse.

Minutes of the cell are official files. Write who recused when their department's tool was discussed. A CS professor should not mark their own lab's shadow key as out of scope. Recusal is how the cell stays an office.

Objections you will hear — and what to do with them

These are the lines that stall the file. Answer them in the room, then put the answer in the note. A spoken answer without paper will be forgotten by the next officer.

Faculty freedom means we cannot inventory lab tools.

Faculty freedom is for scholarship. A key that processes other people's personal data is operations. Research on public models with public text is a different row. Student essays are not public text.

A cell will slow innovation.

It will slow the kind that leaks a mark sheet. It should same-week the hostel matcher. If everything takes a term, fix the SLA, do not abolish the no.

IQAC should run this. It is quality.

IQAC is a purpose. If they run the cell they will be tempted to harvest. Keep them in the room. Do not give them the estate.

We will hire a Chief AI Officer from industry.

You can. They still need the registrar and counsel. A CAIO without the file is another logo. Prefer a coordinator who can write a note the syndicate will sign.

A 30-day stand-up

No logo until the inventory exists. The logo is week five, if you must.

1. Days 1–7: syndicate approves the one-page charter and the halt power. Name the coordinator.
2. Days 8–18: inventory every tool, including shadow keys. Halt ownerless rows.
3. Days 19–24: publish the intake form and the fortnightly SLA.
4. Days 25–30: pick one dull pilot (hostel tickets or fee match) and refuse one glamorous one (ranker or gaze). The pair teaches the campus what the cell is.

How this shows up in the file

Syndicate resolution: charter attached, halt power named, inventory to be tabled every quarter, vendors not members, research compute listed separately from operational agents.

Without the resolution the cell is a WhatsApp group. WhatsApp groups do not survive show-cause letters.

This article is informational field guidance for Indian public universities and colleges, not legal, regulatory, academic, labour or procurement advice. Confirm against the live AICTE / UGC / NAAC notice, the DPDP Act and Rules as commenced, your Act and statutes, reservation orders, labour counsel and the syndicate before you file it.

How this works on an Indian campus

A P3 University should be able to run “Building an AI Cell Inside a University” without importing a US playbook. “university AI cell structure” hits UGC/AICTE/NAAC clocks, exam secrecy, reservation rules, and students who may be minors.

If the AI cell is three enthusiastic faculty and a vendor, you have a club. If it can refuse a workflow, you have an office. DPDP applies to student personal data. Chatbots are not a strategy. Exam and admissions writes stay with officers. Affiliated colleges need isolation, not one shared index.

• No production student data in a vendor SaaS sandbox.
• Write the academic integrity policy before the tool.
• Consent and purpose tags on student-facing agents.
• Budget for staff training, not only licences.

Close this loop before the next CAB

Put “Building an AI Cell Inside a University” on the next change-advisory or bid-opening agenda as a single line item with an owner. If it cannot earn a line item, it will not earn a control. The owner should be a P3 University, not “the vendor.”

Revisit the item when the model, the GeM term, the region, or the SI changes. “university AI cell structure” is not a one-time workshop. It is a watch item. Date the last check. Unsigned watch items are souvenirs.