Pre-Bid Queries That Expose a Weak AI Vendor

The pre-bid was booked for forty-five minutes. Twenty minutes vanished in introductions. Ten vanished in a vendor explaining transformer architecture to a committee that had not asked. The remaining time was spent on bid-security format and whether the turnover figure was average or best of three. Nobody asked where a tool call goes when the agent writes back to the district MIS.

Pre-bid exists so ambiguity dies in public, on a corrigendum, before envelopes are opened. GFR practice and the procurement manuals treat the conference and the written clarification as part of the bid. A question asked in the corridor after tea is not a clarification. A question that never becomes a written reply is a gift to the bidder who already knows the chair.

This checklist is for the officer who will actually type the queries — indentor, CISO nominee, DPO nominee, or the young DS who has to minute the meeting. It is not a script for humiliating startups. It is a script for making claims falsifiable. Weak vendors hate falsifiable questions. Strong vendors prefer them, because the alternative is losing later to a cheaper slide.

How to run the hour so it produces paper

Issue a query format with the bid: clause number, question, why it matters. Require written submission before the meeting. Use the meeting to hear answers and to watch who looks at whom when isolation is mentioned. Then publish written replies to all bidders. Do not let a spoken assurance amend the bid.

Seat CISO, DPO and the owner of the system of record in the room, even if they only speak twice. Procurement cannot invent a tool-privilege answer. If those people are too busy for pre-bid, they will be too busy for the incident as well.

Record who drafted the specification. If a bidder's earlier PoC deck is visible in the statement of work, say so and invite others to propose equivalents. Vendor-shaped specs that are not disclosed are how a pre-bid becomes a formality.

Questions that change the bid

Ask fewer, sharper questions. Ten that produce corrigenda beat forty that produce philosophy. Each question below is written so a yes without an artefact is a fail.

Pre-bid questions to put on the query sheet. Adapt clause numbers to your draft.

Theme	Ask this	A weak answer sounds like	What you then write in the corrigendum
Egress	List every destination the runtime will contact in 24 isolated hours, including licence, DNS, time, registry, APM and support.	We can disable telemetry if required.	Empty or named schedule is a scored or mandatory row. If required is not a schedule.
Training	Does any affiliate, model host or safety pipeline retain prompts, files, embeddings or traces to improve a model?	We follow all applicable laws.	No training / no improvement on customer content, in the contract and in the architecture.
Tools	For a write-back to our MIS, which identity is used, what is the blast radius, and is there a dry-run?	We support all major APIs.	Named tool contract or the row is unmet.
Human gate	Which actions cannot fire without a named department role, and can that gate be switched off in config?	Human-in-the-loop is available.	Gate is a control with a log, not a feature flag.
Accuracy	What task, dataset and metric will you accept as the acceptance test?	We typically see 95 percent accuracy.	No percentage without a task. Publish the method.
SLA	Besides uptime, will you take a target on tool-call success, gate latency and log-export freshness?	We offer 99.99 percent uptime.	Uptime alone is insufficient. Do not invent a magic availability number.
Exit	Will you price a restore drill of prompts, evals, adapters and traces onto our hardware before final payment?	We will assist in good faith.	Priced drill or the exit row fails.
Subprocessors	Name the model host and every subprocessor who can see content or traces.	Best-of-breed cloud partners.	Named list. Change requires notice and a right to refuse.

Questions you should not ask

Do not ask which minister has blessed the product. Do not ask for a free production PoC on live citizen files. Do not ask bidders to write your evaluation matrix in the meeting — that is how one firm authors the exam. Do not ask for a price in the pre-bid if you are running two envelopes; you will contaminate the technical sitting.

Do not ask questions that only the incumbent PoC vendor can answer because they already have your schemas. Publish a redacted schema pack to all, or accept that you have already chosen.

• No live personal data as a quick check during pre-bid.
• No private clarifications after the meeting.
• No new mandatory brand names introduced verbally.
• No promise that we will relax that row if nobody can meet it without a corrigendum.

After the meeting

Publish answers by clause number. If two bidders asked the same thing, publish one answer. If an answer changes a row, change the row. If you realise the specification was vendor-shaped, widen it and extend the bid due date. A short delay is cheaper than a representation.

If a bidder refuses to answer in writing and offers a demo instead, minute that. Demos are not replies. The committee that later marks clarified in pre-bid with no PDF is inventing a document.

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.

Tough questions will scare away MSMEs.

Unclear questions scare honest MSMEs. Precise questions scare brochure resellers. Isolation and a training ban are not turnover tests. Keep EMD and prior-experience relaxations where the current orders put them, and keep the technical questions factual.

We should not educate vendors in public.

You are not educating them. You are binding yourself. A public answer stops the committee from inventing a private meaning later.

Legal can ask these after L1 is known.

After L1 the winner has leverage and the losers have a representation. Ask while the bid is still a competition.

The vendor said these are trade secrets.

Destinations, subprocessors and training rights are not trade secrets in a public contract. Model weights might be. Do not accept a secrecy claim that hides an egress path.

Ten days around the pre-bid

1. Day −10: circulate this checklist to CISO, DPO and the MIS owner. Each must add one question in their noun, not in adjectives.
2. Day −7: publish the query format and the last date for written questions.
3. Day 0: hold the meeting. Minute attendees and refusals to put answers on paper.
4. Day +3: issue written replies and any corrigendum. Extend the due date if a row got harder or clearer.
5. Day +7: freeze private mail with bidders on technical meaning. Further questions go on the portal.

How this shows up in the file

Subject: Pre-bid queries for the agent-platform tender — draft for portal upload.

Queries in the annexure ask for a destination list, training and improvement rights, tool identity and dry-run, human-gate behaviour, acceptance-test method, SLA objects beyond uptime, priced exit drill, and named subprocessors. Spoken answers will not amend the bid. Replies will be published to all.

No live personal data will be used in any demonstration associated with the pre-bid. This note is not legal advice.

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.

How to put this in the RFP, not the preamble

A P2 Procurement who searches “pre-bid queries AI vendor” is usually drafting or scoring a bid. “Pre-Bid Queries That Expose a Weak AI Vendor” belongs in eligibility, the evaluation matrix, or a numbered annexure. If it only lives in the covering note, L1 will ignore it.

A weak vendor survives a polite pre-bid. These questions force a written answer on egress, training rights, tool privileges, accuracy method, and who pays for the exit drill. QCBS weights are a choice you must publish before opening. Accuracy is a task plus a dataset, not a slogan. SLAs for agents must name tool-calls, human gates and log export — uptime alone is a hosting metric.

Do not let a vendor write the specification and then bid on it. Record unsolicited proposals. Pay for pilots that touch personal data. Write exit before you write go-live.

1. Move the control from the preamble into a scored or eligibility row.
2. Attach a one-page definition (accuracy, SLA, language, data handling).
3. Require an artefact in the technical bid, not a slide.
4. Extend the bid date if a corrigendum is material.
5. Minute the demo on your data, offline if you claimed air-gap.

Close this loop before the next CAB

Put “Pre-Bid Queries That Expose a Weak AI Vendor” 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. “pre-bid queries AI vendor” is not a one-time workshop. It is a watch item. Date the last check. Unsigned watch items are souvenirs.