When a PAC Is Justified for AI Software

The PAC committee of a large teaching hospital met on a Thursday to bless an agent that would draft discharge summaries. The vendor had sent a six-page justification. Every paragraph began with only we. The medical superintendent asked one question: have we asked anyone else. Silence. The PAC was not signed. The bid went advertised eight days later. Three technically responsive offers arrived.

That is the normal case. This guide is about the abnormal case: when a Proprietary Article Certificate for AI software is actually justified, how to write it under GFR 2017 Rule 166 rather than the Rule 161 mix-up, and how to keep the certificate from swallowing the next three years of buying.

We sell this category of software. We have an interest. Read the tests as if you will have to explain them to CVC, not to us. This is not legal advice.

The default answer is no

Agent platforms are tools. Retrieval, planning, tool-calling, human approval and logging are capabilities several Indian suppliers can assemble on-prem or in an air-gapped rack. A PAC that says only this firm can run an agent in a state data centre is almost always false in 2026.

GFR Rule 166 is the single-tender rule. The PAC form in the note applies when you rely on 166(i) or 166(iii). 166(i) needs knowledge that only one firm manufactures the required goods. 166(iii) needs a standardisation or compatibility story, on expert advice, approved by the competent authority. 166(ii) is emergency, which is a different sentence and a different kind of honesty.

If you cannot pick a limb, you do not have a PAC. You have a preference. Preferences go into a scored advertised bid, not into a certificate.

Five justifications that can be true

These are the only stories we have seen survive a careful finance wing. Each one is narrower than the software brochure.

PAC limbs that sometimes hold for software-shaped buys

Claim	GFR hook	Evidence that works	Evidence that fails
Only this OEM ships a certified adapter to our existing core system	166(iii) compatibility	OEM letter, interface control document, failed third-party tests	Our team already likes their UI
A statutory or regulator-notified system of record is sold by one firm	166(i) if truly sole manufacturer	Notification, licence, absence of any other authorised seller	They were first to demo
A court or statutory auditor has named a specific system already in place	Often 166(iii), sometimes neither	Order or audit para that names the system	We would rather not migrate
Emergency restoration after a breach, time-boxed	166(ii), not the PAC note	Incident report, time limit, later competitive regularisation	The secretary is travelling next week
Spares or version-locked licences for an already purchased stack	166(iii)	Original contract, compatibility matrix, expert note	We want the same vendor for a new department

Notice what is missing. Indian ownership. DPIIT recognition. A private benchmark. A claim that the weights are open. A claim that the model is sovereign. Those may be bid criteria. They are not proprietary facts.

What a founder should never put in a PAC pack

How to write the reasons box

The GFR form asks why no other make or model is acceptable. Write in nouns. Name the protocol, the certificate, the equipment, the order. Dates help. Names of failed alternatives help. Superlatives do not.

1. State the required article in one sentence that a non-technical joint secretary can repeat.
2. Name every substitute you actually tested or surveyed, including GeM listings and CPPP awards.
3. Say what each substitute failed: interface, certification, delivery inside the air-gap, or a legal constraint.
4. Limit the PAC to the article that failed substitutes, not to a family of future modules.
5. Put an expiry or a review date. Proprietary facts rot.

If step two produces two viable substitutes, stop. You are now in advertised or limited tender territory. Grief about timelines is not a limb of Rule 166.

Price still has to be reasonable

A justified PAC removes competition. It does not remove the duty to show that the public is not being charged a monopoly rent. GeM buyer guidance has for years told PAC buyers to do extra diligence on rates, commonly read with Rule 149. Use last purchase price, a broken-down rate card, a comparable private invoice with numbers redacted if needed, or an independent technical estimate. Do not use the vendor's own uniqueness slide as a price benchmark.

For novel agent software there is often no last purchase price. That is a reason to demand a modular rate card: licence, on-prem support, integration, training, exit assistance. Compare the modules you can compare. Cap the modules you cannot. Write the method. The method is the defence.

Objections, answered without theatre

If we compete, we will lose six months. Then start the bid now and run a time-boxed 166(ii) only if a true emergency exists. A late indent is not an emergency the GFR wrote down.

The other vendors will not sign our security schedule. Put the security schedule in the advertised bid as mandatory. PAC is not how you enforce security. Specifications are.

Our existing vendor will walk if we test the market. Let them walk in writing. A threat to walk is evidence you already have lock-in. Lock-in is a reason to compete the next layer, not to certify it as proprietary.

Legal said PAC is safer than a messy bid. Legal is wrong if the limb is false. A messy advertised bid with a clean file beats a clean-looking PAC with a false sentence.

Scope discipline after a justified PAC

The failure mode of an honest PAC is not the first six months. It is month fourteen, when someone adds a second directorate, a new workflow, and a hosted fallback, all on the original certificate. The original fact — a certified connector, a trial clearance, a locked protocol — does not travel automatically. Each add-on needs its own limb or its own bid.

Write a negative list into the PAC note. The certificate does not cover training on department data. It does not cover a foreign model host. It does not cover a second legal entity. It does not cover year-two modules that were only a slide. Negative lists feel unfriendly in the room. They are how you still have a defensible file when the vendor's account manager changes.

Pair the negative list with a regularisation date. If the PAC is a bridge, the advertised bid for the residual scope should already have a drafting owner and a week. A sunset without a named officer is a wish. Wishes renew themselves.

If you are the founder on the other side of a justified PAC, your job is to make the limitation easy to write. Offer a connector SKU and an agent SKU. Offer a six-month trial and a separately competed production. Offer an exit package priced in the first rate card. The committee that can split your offer is the committee that can sign. The committee that is handed one lump will either refuse or over-certify. Over-certification is how you inherit a vigilance query you did not need.

Committee playbook, ten working days

• Day 1: refuse any pack that cites Rule 161 as the PAC rule. Recast to 166.
• Days 2–4: independent market scan by someone who is not the indentor. GeM, CPPP, two state portals, two peer institutions.
• Day 5: decide the limb or reject PAC.
• Days 6–7: draft reasons, expert note, finance concurrence, rate method.
• Days 8–9: limit scope and write the sunset.
• Day 10: competent authority sits with the opposite page visible: the list of substitutes.

File note the committee can adopt

The committee considered a request for Single Tender Enquiry for an agentic software article. Rule 166 GFR 2017 was read. A market scan is placed opposite. Substitutes exist at the capability layer. The only fact that survives is [compatibility / statutory certification / emergency], limited to [named article] for [duration].

PAC, if issued, will not cover [named adjacent modules]. Price reasonableness is recorded on the next folio by [method]. A competitive process for the residual scope will be initiated by [date]. Members note that vendor-authored uniqueness language was struck.

Prcept AI will supply architecture evidence for an open bid. We will not draft your PAC. If your constraint is on-prem inference and a training ban, write those as specifications and compete them. That is the honest buy.