Reverse Auctions and Why AI Shouldn't Be in One

Opinion, so labelled. A GeM reverse auction is an honest machine for a homogeneous good. Sellers underbid each other in public. The State pays less for the same chair. GFR lists electronic reverse auction as a tender type. GeM made it easy. None of that is the scandal.

The scandal is using that machine on an article that is not the same chair. An on-prem agent with a training ban, an egress proxy and a deletion certificate is not the same article as a hosted copilot with a cheerful India-region slide. If you technically qualify both and then RA the price, you have asked the market to compete on the only attribute that does not measure sovereignty.

This is not legal advice. If your manual forces RA above a threshold, write a specification that makes the remaining bidders actually comparable, or record why RA is a misfit and use the method your competent authority will defend.

What RA optimises

It optimises the last number. It assumes the earlier packets already equalised quality. That assumption is true for A4 paper. It is false for most first-generation agent buys, because evaluation committees are still learning which quality tests matter. They pass anyone with a demo and a GSTIN. Then they RA.

The winner is often the firm that planned to recover margin in change requests, or the firm that planned to host cheaply abroad, or the firm that will staff the account with people who leave. You did not discover value. You discovered who would flinch.

When RA is a tool versus when it is a category error

Article	RA fit	Why
Standard laptops on a published spec	High	Homogeneous after TEC
On-site data-entry seats	Medium	Quality is staffing; still risky
A second-year support renewal of an installed stack	Medium-high	The article is now known
First on-prem agent with a training ban	Low	Quality is the perimeter
PAC-shaped unique connector plus RA on the whole stack	Very low	You mixed a monopoly limb with a price race

What an RA does to a file

QCBS, where your manual allows it, is the grown-up alternative. Score the perimeter. Score the training ban. Score the exit. Then let price have a weight that cannot drown those scores. If your organisation will not do QCBS for software, a two-packet L1 on a fail-able spec is still better than an RA on a brochure.

The opinion, in one page

Do not use reverse auction as the default for first-time agent platforms. Use it, if you must, only after a technical evaluation that can fail a hosted model, a silent training clause, and an exit that does not return weights and logs. If your tests cannot fail those things, you are not ready to price.

If rule or habit still forces RA, publish the quality bar in brutal English and keep a floor in your own bid/no-bid note. Buyers should expect capable vendors to walk. Walking is data.

A note on MSE preference and RA together: if you RA first and then try to apply a price-band match, you have mixed two machines. Read the live bid. Some texts run preference before RA, some after. Software quantity shares after an RA are how you end up with two winners and one stack. If you cannot write the sequence in one sentence, do not tick both.

Objections from people who like the countdown

RA is transparent. So is a scored quality table. Transparency about the wrong variable is still a bad buy.

We saved 22 percent. You saved 22 percent of a number that may have been fiction, and you may have bought a transfer you cannot see.

GFR wants L1. GFR wants a responsive L1 for the required goods. If the goods are not the same, L1 is a typing error.

Technical qualification that only asks for a GSTIN, a turnover, and a brochure is not qualification. It is a guest list. An RA after a guest list is a price race among unlike objects. Write tests that can fail. Then, and only then, talk about a countdown.

If you still must RA

Sometimes the manual, the financial power, or a standing instruction leaves you no graceful exit. Then the work is to make the remaining bidders actually the same article. Fail hosted inference. Fail training on customer data. Fail an exit that cannot return logs and adapters. Fail a missing on-prem runbook. Do this in technical evaluation, in writing, before anyone sees a rupee.

Publish a start price that is not a fantasy. An absurdly high start invites a theatre of cuts that still lands above a reasonable rate. An absurdly low start invites a race into insolvency. Reasonableness does not retire when the countdown begins. It just gets harder to write.

After the RA, do not accept a variation that undoes a test. The firm that went lowest will be first to ask for a hosted fallback or a thinner support roster. If you say yes, you auctioned a specification you did not mean. The file should treat that request as a new article or a refusal.

Vendors: put the floor in a bid/no-bid note before the event. Include GeM charges under the dated revenue policy. When you hit the floor, stop. The chat will tell you to go one more step. One more step is how support dies in month four. Walking is an allowed outcome. It is also a signal the buyer should record.

Playbook for the next AI bid design

• Write fail-able tests: residency of inference, training ban, log export, deletion certificate.
• Decide in the note: QCBS / scored quality / tight L1 / RA.
• If RA remains, record why the remaining bidders are homogeneous.
• Refuse to RA a PAC article as theatre.
• After award, do not accept a variation that undoes the test you RA'd past.

File note, labelled opinion for the competent authority

Electronic reverse auction is available on GeM and recognised in GFR. For this first agent platform, quality attributes (perimeter, training, exit) are not homogeneous across likely bidders. It is proposed not to use RA, or to use it only after tests that fail hosted and training-on-customer-data designs. This is a method choice, not a vendor preference.

Prcept AI will join an RA only to a written floor. We will not quietly cheapen an air-gap to win a countdown. If your committee wants sovereignty, do not auction it.

How a buyer or seller should act on this

Treat “Reverse Auctions and Why AI Shouldn't Be in One” as an operating problem, not a thought piece. A P2 Procurement who searches “reverse auction software procurement” is usually one bid, one CRAC, or one rejection away from a cash event. The file that wins is the one with dates, document names and a named officer — not a paragraph that restates GeM’s homepage.

Opinion: a reverse auction is a race to a number. An agent platform is a race to a perimeter. Using RA as the default for novel AI is how you buy a cheap leak. That is why this opinion ends in artefacts: screenshots of the live portal term, the clause you invoked, and the date you last checked it. GeM, GFR notes and state portals move. A citation without a date is folklore.

Confirm the live GeM FAQ, the current revenue policy and the bid text before you copy any number from a blog — including this one. If the portal and this article disagree, the portal wins. Put the printout in the file.

• Write the purpose of the buy in one sentence a DDO will sign.
• Name the route: catalogue, custom bid, bid, RA, CPPP, or state portal.
• Attach the exemption or preference documents you will actually upload (Udyam, DPIIT, MII, OEM).
• Record who can accept the consignee receipt and who raises the bill.
• Do not invent a category, a PAC, or a price-reasonableness story after L1 is public.

Close this loop before the next CAB

Put “Reverse Auctions and Why AI Shouldn't Be in One” 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 P2 Procurement, not “the vendor.”

Revisit the item when the model, the GeM term, the region, or the SI changes. “reverse auction software procurement” is not a one-time workshop. It is a watch item. Date the last check. Unsigned watch items are souvenirs.