Pricing Models for Government AI Contracts

Finance asked for the cost of AI. The indentor sent a vendor one-pager: rupees per user per month, first year discounted, GPU extra, integration extra, outcome-based optional. The one-pager was not a price. It was a menu that became whatever the change-request log needed it to become.

Government AI contracts fail commercially in two opposite ways. One is a pure L1 on a thin licence line while people, GPUs and SI months hide in a later indent. The other is an outcome price so vague that nobody can say when the rupee was earned. Both look modern in a steering-committee slide.

This guide is for DDOs, finance members and procurement officers who must put a comparable number in the financial envelope. It is order-of-magnitude honesty, not a costing study. It is not legal advice. GFR still wants a reasonable price. Reasonableness for an agent platform is a three-year stack, not a seat coupon.

Five shapes you will actually see

Perpetual licence plus AMC. Familiar from earlier software. Honest if the entitlement is offline and the AMC states what model refresh is included. Dishonest if year-two assurance is the only way to keep the licence server green.

Term subscription. Common on GeM catalogues. Honest if the unit is a named environment plus named users or named workflows, and if exit artefacts are included. Dishonest if every new tool is a new SKU.

Token or call metering. Honest if you can count, cap and reconcile. Dishonest if retrieval-heavy drafts explode the bill and the only meter is the vendor dashboard.

SI wrap. A system integrator bids a lump sum and subcontracts the platform. Honest if the platform, people and GPU are unbundled enough to audit. Dishonest if the SI margin hides a phone-home stack you cannot see.

Outcome or per-file prices. Tempting for grievance disposal or draft notings. Honest only if the outcome is defined, sampled, and not gameable by refusing hard cases. Most agent work is not a clean outcome unit yet. Do not pretend it is.

What to put in the financial bid so prices compare.

Shape	Make bidders quote	Reject or recast if
Perpetual + AMC	Licence, year-2/3 AMC, refresh inclusions, offline entitlement cost if any	AMC is the real kill switch
Subscription	Environment price, user or workflow bands, overage rule, exit export included or not	Every integration is a new SKU
Token / call	Unit definition, department counter, cap, alert, blended estimate for the published workload	Only vendor meter exists
SI wrap	Platform vs people vs GPU vs AMC as schedules, even if one lump sum is paid	A single line called AI solution
Outcome	Definition, sample, exclusions, who judges, what happens to refused cases	Vendor can pick easy tickets

Price the stack, not the SKU

An agent that looks cheap and needs two GPU nodes, a vector store administrator, an SI for every department API, and a 24×7 prompt engineer is not cheap. Publish a pricing schedule that forces those lines into the envelope, even as estimates, so L1 is not a licence-only fiction.

At least name: platform software, model-hosting hardware or hosting (if any), implementation SI, departmental effort you will not pay the vendor for, AMC or subscription years two and three, and exit assistance. You do not need a consultancy-grade costing model. You need the same skeleton on every bid.

Do not ask bidders to quote a citizen-facing SLA at five nines and then compare only the licence cell. The person who will sit on the human gate is a cost. If you omit that person, you have omitted the control.

What finance can actually compare

Give finance a three-year total for a published workload: number of users, number of workflows, expected documents per day, whether write-back is in scope, whether a second site exists. If bidders may quote different shapes, require a conversion table into that workload. A token bid and a subscription bid that cannot be converted are not competitors. They are two conversations.

Price reasonableness still needs a method when competition is thin. Last purchase price will often not exist. Use the independent stack estimate you wrote before floating. A vendor quote is not an estimate.

Taxes, GeM fees where they apply, and exchange-rate clauses for any imported GPU or licence should be explicit. Silent extras are how a reasonable L1 becomes an unreasonable bill.

Discounts that are traps

Year-one free is not free if years two and three are unquoted. PoC-to-production credit is not a credit if it requires production personal data in the PoC. Academic or government special prices that expire at the first renewal are a lock. Write the renewal formula now.

Volume discounts that assume all-department adoption in month one will punish you when only one directorate goes live. Band the price on actual commissioned workflows, not on a hope in the inauguration speech.

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.

GFR wants L1. Stop complicating the envelope.

L1 among comparable stacks is a method. L1 among a licence, a token cloud and a turnkey mystery is a category error. Make them comparable, then apply the method your manual allows.

Outcome pricing will align incentives.

Only if you can define the outcome without letting the vendor refuse the hard file. Most noting and grievance work is not there yet. Use outcomes as a small, sampled bonus at most.

We cannot estimate GPU need, so we will leave it out.

Then you are not estimating the system. Put a band in the bid — CPU-only allowed, or one class of GPU — and make bidders state assumptions. Silence is how the indent returns in November.

Startups cannot quote three years.

Then they cannot take a three-year production contract. A shorter term with a tested exit is more honest than a renewable mystery.

Build a comparable envelope in three weeks

1. Week 1: publish the workload — users, workflows, documents per day, write-back or draft-only, sites, term.
2. Week 2: issue a pricing schedule with the five shapes mapped into that workload, plus people, GPU, SI and exit lines.
3. Week 3: write your own order-of-magnitude stack number for reasonableness. Then freeze the schedule so bidders cannot invent private units.

How this shows up in the file

Subject: Pricing schedule for the agent-platform bid — for finance concurrence.

Bidders will quote against a published three-year workload. Platform, implementation, hardware or hosting assumptions, AMC or subscription, and exit assistance appear as separate cells even if payment is lumped. Token or outcome units, if offered, must include a department-side counter, a cap, and a conversion into the published workload.

The attached departmental estimate is an order-of-magnitude reasonableness aid, not a costing study. This note is not legal advice.

How to read an abnormally low stack

A bid that undercuts your departmental band by half is not automatically corrupt and not automatically clever. Open the cells. If SI months are near zero, someone expects a change request. If GPU is missing, someone expects a second indent. If exit is free and undefined, someone expects you to renew. Write those three sentences before you celebrate L1.

Your rules may already have an abnormally low bid process. Use it. Do not invent a vibe rejection. Ask the bidder to explain the stack against the published workload. A hollow explanation is a technical or commercial fail, depending on where you placed the unbundling. A real explanation updates your band for the next file.

The same discipline applies to a bid that is far above the band. Extra isolation drills, a second site, or a heavier eval can be reasons. A luxury model brand is not a reason unless the bid asked for that brand — and it should not have.

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 “government AI pricing model” is usually drafting or scoring a bid. “Pricing Models for Government AI Contracts” belongs in eligibility, the evaluation matrix, or a numbered annexure. If it only lives in the covering note, L1 will ignore it.

The cheapest line item is rarely the cheapest three years. Compare agent prices as a stack — platform, people, GPU, SI, AMC — and refuse units only the vendor can count. 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 “Pricing Models for Government AI Contracts” 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. “government AI pricing model” is not a one-time workshop. It is a watch item. Date the last check. Unsigned watch items are souvenirs.