All insights

Compute & Cost

Hidden Costs in AI Vendor Quotes

· 9 minute read

The quote is the part the vendor numbered. The file is the part you will pay. Tick SI, cleanup, AMC, power, idle GPU and evals before you score L1.

An AI quote that fits in one cell is incomplete. Procurement officers who have bought servers for fifteen years already know this. The new trick is that the missing cells have new names — eval sets, embeddings, idle GPU, prompt-version labour — and the vendor's commercial format was designed not to have rows for them.

This checklist is for the officer who must score two bids and does not want to explain a supplementary sanction in month eight. It is not an accusation that every SI is dishonest. It is an admission that a legal quote can still be a bad map of the cash.

Tick the list. Put the ticks in the comparative statement. If a bidder will not explode a lump, mark the row unpriced risk and decide whether that is acceptable. Will be absorbed is not a price. Prcept will put these rows in our quote even when it makes us look more expensive than a lump.

The six you will pay anyway

System integration. The licence or the model is the smaller invoice. Connecting the agent to the grievance MIS, the HR pay file, the campus IAM, the HSM, the syslog, and the officer's actual login is the larger one. If the quote says integration extra or standard connectors included without naming your systems, you do not have a quote for go-live.

Data cleanup. Production scans are stamped, skewed, bilingual and incomplete. Demo PDFs are not. Cleanup is either your staff, a contractor, or a silent delay. Price it.

AMC and support. Warranty ends. The reasonable band for government IT AMC is a market conversation, not a law — see the companion data-study. If the quote is silent after year one, you have a cliff.

Power. On-prem GPUs draw watts whether the slide mentioned PUE or not. Use the SERC method. Do not accept negligible. Idle GPU. A card you bought for a peak and leave powered is a standing bill. If the architecture cannot sleep or share, write the idle rupees as a feature of the bid, not as your incompetence later.

Eval sets. Someone must build and refresh the gold cases. If the vendor's accuracy is a public leaderboard, you will build the real set after go-live, which is to say you will pay twice.

Checklist for the comparative statement. A blank is a risk, not a saving.
Hidden lineAsk in writingAcceptable evidenceReject as priced
SI / integrationNamed systems, environments, RACI after exitSOW with systems and hoursConnectors included
Data cleanupWho labels, who scans, who rejectsHours or a discovery milestone with a capCustomer to provide clean data
AMCYears 2–5 rupee, scope, response, exclusionsA commercial annexSilence, or as per GFR 10 percent
PowerWatts, PUE assumption, who pays the DISCOMA method sheetGreen AI, no watts
Idle GPUPower-cap, share, scheduleA utilisation designOne node per workflow by default
Eval setWho owns cases, cadence, accept criteriaA gold-set plan in the SOWLeaderboard screenshot

The quieter seven

Storage: embeddings, traces, conversation logs, CERT-In 180-day floors and departmental retention. Priced in the storage article. Egress and snapshot fees on any cloud path, including empanelled GPUs. Key management and HSM professional services. Privilege-access and jump-box work for air-gap, including the people — see the ops-headcount piece. Language packs and Indic eval that were included only in English. Change requests when a circular changes the form. If the contract treats every circular as out of scope, you have bought a statue. Exit: export, sanitisation, knowledge transfer. The last invoice is not the last cost.

How to score this without poisoning L1

If the tender is a pure L1 on a badly specified BoQ, hidden costs will not save you; the specification will have to carry them as line items. Add provisional items for cleanup, eval and year-2 AMC so the bid you open is the bill you will pay. CVC and GFR both prefer a complete specification to a clever scoring trick after the fact.

If you have QCBS or a consultancy-like SI envelope, put unpriced risk in the technical narrative and make completeness of the cost map a marked item. Do not invent a points scheme that is a disguised brand preference.

If a competitor's lump wins, we would rather lose the bid than win the para. A complete L2 that finance can live with is better public money than an incomplete L1 that needs a supplementary.

Language and change are not nice-to-haves

If the quote's eval and support are English-only, the Indic queue is an unpriced risk. Write the language mix of last year's tickets on the comparative statement. Ask each bidder what is included for those languages. A connector that cannot tokenise your script is not a connector.

Circulars change. A grievance scheme that is stable for eighteen months is a blessing, not a plan. If the contract treats every new GR as a change request, year-two cash is a lottery. Either include a defined number of circular updates in the annex, or budget a change bag. Silence is how L1 becomes a second sanction.

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.

If we add provisional items we will inflate the estimate and lose approval

An estimate that omits go-live work is not smaller. It is late. Sanction the honest BoQ or cut scope in public.

The SI said those items are in the professional-services bag

Ask for the bag to be unzipped. Hours, systems, exclusions. A bag is a lump with a nicer name.

Power is not an IT cost

Then put it in the file as a facilities cost attributable to this purchase. Hiding it does not move it.

Eval sets are research; we are an operational department

Then you are buying a tool you cannot test. Operational departments test bridges. They can test a model.

A one-week quote autopsy

  1. Day 1: paste the six rows and the quieter seven into the comparative statement as empty columns.
  2. Day 2–3: send written questions. Accept only named systems, rupees or explicit not included.
  3. Day 4: mark unpriced risk. Decide whether to add provisional BoQ items or to reject incompleteness.
  4. Day 5: re-score only after the rows are filled or formally accepted as buyer-side cost. Initial the sheet.

How this shows up in the file

Subject: Completeness check of AI commercial quotes — [tender].

Quotes were exploded against SI, data cleanup, AMC, power, idle GPU, eval sets, storage, egress, KMS, air-gap ops, language, change and exit. [Bidder] left [rows] unpriced. Those rows are treated as [buyer-side estimate / unpriced risk / reason to seek a revised commercial]. L1 will not be recommended on an incomplete map.

This checklist is not a CVC circular and not legal advice. Unpriced risk is not a saving.

This article is informational field guidance for Indian public institutions, not legal, procurement, tax, accounting, tariff or engineering advice. Confirm against the current Gazette, GFR, GeM term, SERC tariff order, IndiaAI portal rule, CAG mandate, DPDP text, departmental finance manual and your counsel before you file it. Figures are methods and order-of-magnitude illustrations, not a dataset of real deployments and not a substitute for a live quote.

How to put this in the finance note

A P2 Procurement searching “hidden cost AI quote” needs a number a CFO can defend, not a GPU brand. “Hidden Costs in AI Vendor Quotes” belongs in a cost model with people, power, idle time, AMC and the cost of a failed pilot.

The quote is the part the vendor numbered. The file is the part you will pay. Tick SI, cleanup, AMC, power, idle GPU and evals before you score L1. IndiaAI subsidy, if you use it, is a live notice — not a permanent discount. On-prem TCO includes ops headcount. Do not invent Rs/hour. Cite the source of every rupee.

  • Separate capex, opex, and one-time cleanup.
  • Show utilisation, not just peak GPUs.
  • Price the human fallback, not only inference.
  • Date every tariff and subsidy assumption.

Close this loop before the next CAB

Put “Hidden Costs in AI Vendor Quotes” 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. “hidden cost AI quote” is not a one-time workshop. It is a watch item. Date the last check. Unsigned watch items are souvenirs.

Questions this usually raises

What hidden costs appear most often in government AI quotes?
In our experience of files, SI integration, data cleanup, post-warranty AMC, power, idle GPU and gold-set labour. Storage and exit are the next quiet pair.
Is a lump-sum SI quote illegal?
Not by itself. It is a poor map. GFR wants a fair and complete specification. Ask for an explosion even if the contract form stays lump-sum.
Can we hold a retention against hidden costs?
Retention secures performance of what was specified. It does not magically fund work that was never in the BoQ. Specify the work.
Should AMC be a percentage?
It can be a percentage of a defined hardware value or a rupee schedule. What it cannot be is a silent year-two surprise or a fake statutory rate.
Will Prcept's quote include these rows?
Yes, or it will say not included in those words. We would rather be L2 on a complete sheet than L1 on a poster.
Is this a CVC circular?
No. It is a procurement checklist. Confirm against GFR, CVC instructions and your tender conditions before you score.

Sources