Cost of a Failed Pilot, Honestly Counted

The industry is fond of saying that failed pilots are tuition. Tuition is a fee you plan to pay for a curriculum. A failed government pilot is often a sanction you cannot repeat, a team that will not volunteer again, and a PAC question about a GPU that now warms a store. Count it like a project, not like a TED talk.

This is a data-study method. We will not quote a percentage of Indian government AI pilots that fail. We have not run a census, and we will not borrow a global consultancy figure and relabel it Bharat. What we can do is show the cost stack of a failure, the residual-value questions GFR already knows how to ask, and a worked illustration with labelled assumptions.

A pilot that you can kill cheaply is a well-designed pilot. A pilot that can only succeed is not a pilot. It is a purchase with a nickname.

What counts as failure, for this file

Failure here means you stop. The agent does not go to production. The sanction is not renewed. The hardware is reassigned or condemned. Partial success — a gold set you keep, a workflow you redesign, a smaller model you retain — is not failure. It is a result. Do not call a result a failure to make a speech, and do not call a stop a pivot to avoid counting.

Time-box the count. Cash from first sanction to the stop note. Officer days from the named team. Residual value of assets on the day of the stop, using a labelled assumption, not a fake GFR rate. Opportunity: the other IT file that waited. That last line is narrative unless you can name the file. If you can name it, put it in.

Do not count lost citizen trust as a rupee. Count the RTI, the newspaper, the reopened grievances if they exist. Trust is a purpose of the kill criteria, not a cell in the spreadsheet.

The stack you will actually find

Cash: vendor invoices, GeM licences, IndiaAI hours, SI milestones you cannot claw back, annotators, travel, a leased line you opened for a temporary cloud. Open the PAO's report. Do not estimate from memory.

People: the product owner, the DPO's hours, the NIC / SDC engineer, the officers who sat in workshops instead of clearing files. Use their loaded cost only if your finance manual already does that for other projects. Otherwise count days, not rupees, and let finance convert if they wish.

Assets: GPUs, servers, labelled disks, a rack PDU. Existence, utilisation, disposal or reassignment. An idle card after a failed pilot is the finding CAG knows how to write. Reassign it in the stop note. Contracts: AMC that auto-renews, a cloud account without a cap, a professional-services bag with a remaining milestone that pays on time not on outcome. Read the kill clauses before you start, which is the only time they are cheap.

Illustration — a mid-sized departmental pilot stopped at month seven. Figures are labelled fiction for shape, not a survey result.

Line	Illustrative cash / days	What decided it
Software / SaaS invoices	₹18–40 lakh (range)	Whether year-one was prepaid
SI milestones paid	₹20–60 lakh	Whether go-live was the first big milestone
IndiaAI / cloud hours	₹2–15 lakh	Whether an account cap existed
Gold-set / cleanup labour	40–120 officer-equivalent days	Whether the set is reusable
Hardware not reassigned	₹0 if reassigned; else a write-off file	The stop note's asset paragraph
AMC already committed	₹0–20 lakh	Whether you signed year-two in year-one

Make failure cheap on purpose

Design the pilot so the expensive things are reversible. Rent the burst. Own only the small card you would keep for a narrower workflow. Pay SI on evidence — gold set accepted, packet export demonstrated, MIS write in staging — not on calendar. Put a kill date in the sanction. Put a kill owner in the RACI.

Keep the artefacts that survive a stop: the gold set, the purpose map, the data-flow, the packet spec. Those are capital in the intellectual sense and cheap compared to a second discovery. Prcept would rather lose a production order than run a pilot that cannot be stopped. We will put kill criteria in the SOW. If a competitor tells you failure is impossible because their model is frontier, they are selling you a purchase.

What should survive the stop

A cheap failure leaves four artefacts: the gold set with its licence to reuse, the purpose and role map, the packet specification, and the exploded Ledger A of what was actually spent. Those four make the next attempt shorter. A failure that leaves only a slide deck is a failure you will pay for twice.

Close the cloud account on the stop day. Unused credits are not a reason to keep a hose open. Reassign hardware on the stop day. Unused GPUs are not a consolation prize. They are the finding.

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.

Talking about failure will kill the sanction

A sanction that cannot survive a kill paragraph will not survive contact with the MIS. Write it. Finance officers prefer a cheap stop to a quiet bleed.

We cannot publish a cost of failure; it will become a headline

Then write it as an internal annex. The headline you are trying to avoid is the one where you cannot say what you spent.

Industry failure rates say this is normal

Industry rates are not your file. Do not comfort yourself with a global percentage you cannot source to a public Indian dataset.

The hardware is an investment even if the pilot dies

Only if you reassign it in the stop note with a workload. A crate is not an investment.

Write the stop before you start

1. Before kickoff: list kill criteria, kill date, remaining SI payments if you stop, hardware reassignment owner, account cap.
2. Day 30: count cash and days so far. If the gold set is not started, you are already failing expensive.
3. Kill week, if it comes: issue the stop note, reassign assets, export the gold set, close the cloud account, tell finance the cash total.
4. The following month: file the reusable artefacts. Do not hold a lessons-learnt meeting without the cash annex.

How this shows up in the file

Subject: Pilot [name] — stop / continue at [date].

Kill criteria were [list]. Status: [met / not]. Cash to date by head: [table]. Officer days: [n]. Assets: [reassign / retain / dispose]. Reusable artefacts: gold set [ref], packet spec [ref]. Remaining contractual exposure: [amount]. This is a stop / a redesign / a go-live recommendation.

Illustrative figures in any annex are labelled. This is not a national failure statistic and not legal advice. Design the stop before you start.

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 P1 CIO/CTO searching “failed AI pilot cost” needs a number a CFO can defend, not a GPU brand. “Cost of a Failed Pilot, Honestly Counted” belongs in a cost model with people, power, idle time, AMC and the cost of a failed pilot.

A failed pilot is not free because the licence was discounted. Count cash, officers, leftover hardware, and the year you cannot spend again. This is a method, not a national failure census. 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 “Cost of a Failed Pilot, Honestly Counted” 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 P1 CIO/CTO, not “the vendor.”

Revisit the item when the model, the GeM term, the region, or the SI changes. “failed AI pilot cost” is not a one-time workshop. It is a watch item. Date the last check. Unsigned watch items are souvenirs.