Why L1 Pricing Fails for AI Procurement

The financial envelopes went down the table. Three technically qualified bids. The chair read the totals. The cheapest was a hosted copilot with a rupee-per-seat line and an integration to be scoped. The dearest was an air-gapped install with a local licence, a human-gate module, and a priced exit drill. The room did the arithmetic Indian public procurement is built to do. L1 won.

Nobody in that room was corrupt. L1 — lowest eligible price — is a fair rule when the remaining goods are alike. A ream of paper that met the spec is a ream of paper. An agent platform that met a four-row brochure spec is not alike to another. The failure happened before the envelopes. Eligibility had not carried sovereignty, tool discipline, or exit. Price then did the only job it was given: it finished a comparison that should never have been a comparison.

This is an opinionated field note, not a direction to ignore GFR. For goods, L1 remains common and often right. QCBS is discussed in the Manual for Procurement of Consultancy and Other Services and in later DoE writing on certain works and non-consultancy services. Weights are not a universal mandate you should invent. The argument here is narrower: do not run pure L1 against an eligibility sieve that an adjective can pass.

What L1 actually is

L1 is the lowest price among bids that survived the published filter. The filter can be a specification, a qualification, a two-envelope technical responsiveness test, or a scored threshold. If the filter is real, L1 is a discipline against favouritism. If the filter is a demo and a turnover figure, L1 is a discipline against quality.

Reverse auctions make the same assumption faster: the remaining items are commodities. An agent that may write to a system of record is not a commodity. A reverse auction on a thin spec is how you pay less for more risk.

Two-envelope L1 is still L1. Opening technical first is hygiene. It does not create quality marks. It only stops the chair seeing price while pretending to judge architecture. Useful. Insufficient.

Where agent platforms are not alike

They differ on who can see the prompt. They differ on whether a licence server can halt inference. They differ on whether a write-back has a dry-run. They differ on whether logs exist as files you can keep. They differ on whether an adapter will load on your GPU after exit. Those differences are not premium features. They are the difference between a processor you control and a service you rent with extra adjectives.

A hosted product can be the right buy when the workflow is non-sensitive and the file says so. It is not the same good as an isolated stack. Putting both in one L1 bucket because both said AI platform — yes is a category error. Split the object or split the lots. Do not average them with a price.

Eligibility missing versus quality missing — both make L1 lie.

Difference	If it is only a scored nice-to-have	If it is missing from eligibility entirely	What to do
No undeclared egress / offline licence	Cheap hosted bid still qualifies	L1 will pick the phone-home stack	Make isolation a pass/fail or a heavy QCBS block if your method allows
No training on customer content	Improvement rights hide in the MSA	You fund someone else's model	Pass/fail in the processing schedule
Human gate on write-back	Vendor will sell it as CR	The agent writes unsupervised	Mandatory for write-back workflows
Exportable logs and evals	You will discuss later	CERT-In and exit both fail	Mandatory artefact at acceptance
Priced, tested exit	Renewal becomes the only path	Same, only later	Put the drill in the envelope

When L1 is still honest

L1 is honest for a named GPU SKU that several OEMs can supply against the same warranty. It is honest for a well-specified workstation. It is honest for a second-year AMC on a platform you already accepted, if the work is the same. It is honest for an agent only after the remaining bids have passed isolation, training-rights, tool, gate, log and exit tests that a committee can fail.

That last sentence is just L1 after a hard filter. Many organisations will remain on that method because their software is still treated as goods, or because QCBS needs a competent-authority step they have not taken. Fine. Spend your courage on the filter, not on a speech against L1.

If you cannot write a fail-able filter, you are not ready to float. Delay is cheaper than an L1 hostage.

How to argue the note without inventing a mandate

Do not write DoE has banned L1 for AI. It has not. Do not write QCBS 80:20 is compulsory for all software. It is not. Write: the object is not a homogeneous good; quality differences affect residual risk and total cost; therefore we will either (a) apply a method our manual allows that scores quality, or (b) keep L1 but move the listed controls into eligibility and acceptance.

Cite the live manual you actually use — goods, consultancy, other services, or a state code. Cite CVC's long-standing expectation that specifications should not be theatre. Cite your DPO on processing risk. Then stop. A note that over-claims the law is how finance sends the file back.

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.

Audit will object if we do not take L1.

Audit objects to an unexplained departure, and also to a wasteful buy. A file that shows why remaining bids were alike — or why they were made alike by a hard filter — is the defence. A file that says AI is special without rows is not.

QCBS is for consultants, not software.

Treat that as a live reading of your manual, not as folklore. If your method truly cannot score quality, harden eligibility. Do not use the folklore as an excuse for a brochure filter.

The minister wants speed. L1 on GeM is speed.

Speed to the wrong award is a PAC meeting next year. A two-week delay to write fail-able rows is speed of a different kind.

If we tighten eligibility only one firm will pass. That is PAC by stealth.

Then widen to functional controls any capable Indian on-prem firm can meet, or admit single source under the correct GFR limb with reasons. Do not use a fake competition to avoid a real PAC.

A week to decide whether L1 can be honest

1. Day 1: list controls that, if missing, make the good a different good.
2. Day 2: ask counsel which method your manual actually allows for this object.
3. Day 3: if L1 stays, move those controls into eligibility and acceptance with artefacts.
4. Day 4: mock two bids — hosted thin and isolated thick — through the filter. If both pass, start again.
5. Day 5: write the note in the modest voice. No invented mandates.

How this shows up in the file

Subject: Method of selection for the agent platform — L1 only after a fail-able filter.

The undersigned does not propose a departure from the organisation's procurement method without competent-authority approval. The proposal is that, if L1 is used, eligibility and acceptance will include isolation or named egress, a training-rights ban, tool and human-gate contracts for write-back, exportable logs, and a tested exit. Bids that cannot show those artefacts will not be treated as like goods.

If a scored quality-cum-cost method available under our manual is preferred, a separate note will propose weights for this procurement only. This note is not legal advice.

What to tell the purchase committee in one minute

Say this and sit down: we are not asking to abolish L1. We are asking that only like goods reach L1. Isolation, training rights, gates, logs and exit are how these platforms differ. If those stay outside eligibility, the cheapest brochure will win and we will float again next year.

If someone asks for the legal mandate, do not invent one. Hand them the live manual page and the residual-risk sentence. A committee that hears AI is special without rows will either rubber-stamp or send the file back. Rows are what make the minute defensible.

If the committee still wants speed, offer the two-week filter rewrite, not a speech. Speed that produces a second tender is not speed. The GeM L1 that bought a second project is the example to keep on the table, not a sermon about modernity.

This article is a field guide for Indian public buyers, not legal, procurement, financial or audit advice. Confirm every citation against the live GFR compilation on doe.gov.in, the relevant DoE procurement manual, GeM terms, CVC guidance and your own counsel before a sentence enters a tender file.