Campus AI Vendor Evaluation Scorecard

Three vendors demoed in the syndicate room. All three were fluent in English about the late-fee rule. One of them was reading from a prompt the CSM had pasted that morning from last year's handbook. The scoring sheet had twenty marks for user experience and two for data security. The fluent vendor won. The handbook was wrong. The syndicate spent the next month on a corrigendum to students.

Evaluation is how you prevent that month. A scorecard is useful only if a competent evaluator can fail a bid on evidence. Marks for innovation, AI maturity and vision for NEP are not evidence. Isolation drills, a refused write, a DPA, and an export the campus owns are evidence.

This scorecard is for technical committees in public and private campuses. Adjust weights to your purchase manual. It is written on 17 August 2026. It is not a GFR form and not legal advice. If GeM or your state portal forces a different sheet, map these rows into it rather than running a shadow beauty contest.

Rows that can fail a bid

Isolation and residency. Can the bidder run inference and logs on campus or in a named Indian government or SDC cloud with a packet-capture drill? Fail if first boot phones home or if embeddings live in an unnamed foreign tenant. Corpus pin. Can they operate on a signed pack with no CMS crawl? Fail if they require access to all drives to be useful.

Write tools. Do they refuse register, marks, holds and payment tools unless you later grant them? Fail if the demo silently posts a course-add. Processor terms. DPA, training ban, sub-processors named, CERT-In-shaped logs. Fail if training is opt-out in a portal we control. Export and exit. Four-week run without the vendor; packet export without their engineer; model hash you can reload. Fail if the knowledge dies with the tenant. Indic and code-mix. Your evaluation set, not theirs. Fail if they will not sit your Hindi-English or regional-language tickets.

A working sheet. Change weights. Do not change the idea that fluency is not a row.

Row	Suggested weight	Evidence	Auto-fail if
Isolation / residency / no phone-home	20	Dark-room boot, pcap, architecture annex	Undisclosed egress or unnamed foreign embedding host
Corpus pin and pack workflow	15	Signed-pack demo, owner role	Requires full CMS / SIS dump to function
Write-tool refusal and audit of tools	15	Tool list, refused actions in demo	Any unauthorised write to a student record
DPA, training ban, sub-processors	15	Draft DPA, data-flow diagram	Training on campus data not contractually barred
Export, hashes, 90-day exit	15	Sample packet, load-path note	No campus-owned export
Your eval set (Indic, correctness)	10	Blind scored answers	Refuses to use the campus set
Load story for registration week	5	Replay or architecture that can replay	Hand-waving autoscaling of the model only
Commercial: pilot vs AMC split	5	Two prices, exit clause	One bundled transformation number

Rows to delete

User experience of a scripted demo. Everyone is nice in a script. Number of models in the catalogue. Catalogue is not control. Company valuation or AI lab in three countries. Alignment to a CAG or UGC AI circular that does not exist. Invented alignment is a lie. DPIIT recognition as a technical mark. It may be eligibility. It is not isolation. ISO certificates as a substitute for the packet. Keep them as a pass/fail annex if your manual wants them, not as thirty technical marks.

How to run the room

Same pack, same ten questions, same two poison questions (wrong date, request to change a mark). Blind the brand on the scored answers if you can. Registrar or assistant registrar in the room. CIO on isolation. DPO on the DPA. Finance on the split price. A student member is optional; they are not a substitute for the registrar. Record the pcap and the refused write as annexures. Minutes that say demo was impressive are how the fluent vendor wins again.

Public vs private campuses

Public campuses must also survive GFR or state-rule competition tests. Do not write a scorecard that only one brand can pass unless the file says why the market is that thin. Private campuses can still be wrecked by a single-source platform that fails exit. Use the same fail rows. Your board will not enjoy a hostage AMC. If you must use GeM, map these rows into the service description and the ATC. A GeM L1 on seats alone will not carry them.

Two rooms you can walk into

The sheet decided more than the model did.

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.

These rows will eliminate global SaaS brands.

They will eliminate brands that cannot isolate or contract a training ban. If that thins the field, write it down. Do not lower the row to keep a brand in.

We do not have time for a pcap.

Then you do not have time for a production agent. You have time for a brochure.

Equal marks for all rows is simpler.

Simple is how UX beats isolation. Weight the rows that destroy a register.

The vendor will coach to the scorecard.

Good. Publish it. Coaching to isolation and a refused write is the point. Hide only the poison questions.

A three-week evaluation you can defend

Publish the sheet with the RFP or the invitation. Hidden scorecards are how members get accused of tailoring.

1. Week 1: adopt the fail rows, set weights, map to your portal or GFR/state sheet. Counsel reviews invented-circular language.
2. Week 2: build the campus eval set and two poison prompts. Freeze the pack. Invite a dark-room slot per bidder.
3. Week 3: score independently, then meet. Auto-fails first. Do not average an auto-fail back into contention. Finance opens commercials only for technically responsive bids if your manual says so.

How this shows up in the file

Subject: Technical scorecard for campus agentic AI. Fluency is not a marked row. Auto-fail: undisclosed egress, required SIS dump, unauthorised register write, missing training ban, no campus-owned export. Weights in an annex. Same pack and questions for all bidders. Registrar present for the letter and calendar prompts.

What we will and will not claim

Prcept AI should lose if we fail isolation, refuse-list, DPA or export. Use this sheet against us. If we ask you to add a vision row, strike it.

This article is informational field guidance for Indian universities and public institutions, not legal, procurement, audit or engineering advice. Confirm against the live Gazette, GFR, state financial rules, GeM terms, UGC text, GIGW, DPDP commencement, departmental manual and your counsel before you file it.

How this works on an Indian campus

A P3 University should be able to run “Campus AI Vendor Evaluation Scorecard” without importing a US playbook. “university AI vendor evaluation” hits UGC/AICTE/NAAC clocks, exam secrecy, reservation rules, and students who may be minors.

A campus AI scorecard marks isolation, corpus control, register writes, processor terms, export and exit. Fluency is a demo, not a mark. If a row cannot fail a bidder, delete the row. 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 “Campus AI Vendor Evaluation Scorecard” 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 vendor evaluation” is not a one-time workshop. It is a watch item. Date the last check. Unsigned watch items are souvenirs.

What must be true before you file this

If “Campus AI Vendor Evaluation Scorecard” is only a heading, it will not survive a file inspection. A P3 University should be able to attach one artefact that proves “university AI vendor evaluation”: a log export, a clause, a scored row, a dated notice, or a refusal rule.

Write three dated sentences: what was decided, who owns it, and when it will be re-checked. Unsigned sentences are souvenirs. Dated sentences are controls.

• Name the owner of “university AI vendor evaluation” inside the institution.
• Attach one artefact a stranger can open next year.
• Revisit when the model, the notice, or the SI changes.
• Do not treat a vendor slide as evidence.