Compute & Cost
A Reusable TCO Calculator for Departments
· 9 minute read
Price the unit, the people, the meter, the logs, the eval and the exit. If a cell is blank, the bid is a story. This is the spreadsheet we wish every AI file already contained.
A finance controller in a line ministry asked three bidders for total cost of ownership. She received three slide titles and one 'TCO dashboard' that added licence to professional services and called the rest optional. Nobody had priced the two people who would sit the night watch on the node. Nobody had priced the 180-day log store. Nobody had priced the day the vendor's India region changed its meter.
TCO is not a vendor genre. It is an arithmetic of the estate you will still be running when the implementation partner has gone to the next state. If you cannot put that arithmetic on one sheet the department owns, you cannot compare bids. You can only compare adjectives.
This calculator is a field tool for CIOs and finance, dated 17 August 2026. It is not a GFR schedule and not legal advice. Copy the rows. Change the numbers. Refuse any bid that cannot fill them. Prcept AI publishes this because on-prem agents look expensive until you fill the hosted rows honestly, and hosted APIs look cheap until the meter meets a public URL.
Reuse it. The same sheet should work for a scholarship cell, a campus helpdesk and a PSU vendor desk. If a row does not apply, write zero and a reason. Blank is how optional costs become next year's supplementary.
The seven blocks that make a TCO
Block A is the unit. How many closed pieces of work per month, and what is the as-is loaded cost of one. Without A, every later rupee is a vibe.
Block B is the model path: hosted tokens or reserved capacity, or on-prem cards, power, and a depreciation line. Write both if you are comparing. Do not let a bidder fill only their favourite block.
Block C is people: prompt owner, evaluator, node operator, reviewer who still signs, and the CISO hours you will spend on tools and egress. People dwarf cards in most Indian departmental estates.
Block D is the surrounding estate: retrieval index, object store, identity, SIEM, and the CERT-In 180-day log retention in Indian jurisdiction. Logs are not free. They are also not optional for a government ICT system that will be asked for them.
Block E is assurance: red team before public rollout, hold-out eval after every change, and the annual internal audit sample. Skip E and you will pay it as an incident.
Block F is the meter risk: a denial-of-wallet cap, overage rate, and who gets the SMS when 70 percent of the monthly ceiling is gone. Hosted paths need this in rupees. On-prem paths need the equivalent in GPU hours so a runaway loop cannot cook a rack unattended.
Block G is exit: export of prompts, policies, traces, embeddings and the four-sentence decision packet, plus the hours to stand the same unit up on another stack. If G is 'call the CSM', you do not have TCO. You have a hostage.
| Cell | What you write | Fail if |
|---|---|---|
| A1 | Units per month and as-is ₹ / unit | The unit is 'engagement' |
| B1 | Token ₹ or card+power ₹ | Only year-1 licence is shown |
| C1 | FTE by role, loaded | Implementation is 'included' |
| D1 | Log store and SIEM ₹ | Logs are 'in the platform' |
| E1 | Eval + red-team ₹ / year | Assurance is a certificate |
| F1 | Hard monthly cap and owner | Overage is 'rare' |
| G1 | Exit drill hours and artefacts | Exit is a clause with no drill |
Hosted versus on-prem on the same sheet
Do not run two cultures. Run one sheet with two columns. The hosted column must include egress, residency caveats, abuse-review paths, and the overage. The on-prem column must include power, spare parts, the person with root, and the model-update media path.
A Mumbai region is not a zero in the sovereignty row. Telemetry and third-party model APIs can still leave the perimeter. Price the control you actually have, not the region name. Those arguments live in companion pieces; the calculator only needs a row that says 'undeclared egress: fail' or 'declared and accepted: residual ₹'.
Taxes, AMC, and GeM price rules belong in the finance column, not in a vendor footnote. If a bidder will not put GST and AMC on the same sheet as tokens, they are not ready for a government file.
How to fill it without lying
Use last quarter's real tickets, not a forecast from a deck. If you do not have tickets, you do not have a unit. Do the two-week time-and-motion first.
Use loaded people cost from your own establishment, not a national average. A NIC deputation and a contract engineer are different numbers. Write which.
Use the vendor's worst public overage, not their introductory credit. Credits expire. Caps do not.
Write assumptions as dated sentences. 'We assume 40,000 citizen queries a month at a 12 percent agent-handled rate, review on all benefit decisions, logs retained 180 days in the SDC.' A later officer should be able to change one sentence and recompute.
- No cell named miscellaneous above two percent of the year-1 total.
- No negative line called efficiency gain unless the measurement method is attached.
- No 'AI will free 8 FTEs' without a posting-order theory. Freed time that stays in the same chair is not a saving.
- Three-year view required. Year-1 discounts are how bad TCO hides.
What the calculator is not
It is not a scoring formula that picks a bidder. Eligibility and quality still come first. Price last. A cheap TCO on an undeclared write tool is a threat model with a discount.
It is not a licence to invent savings. If you cannot measure the unit after go-live, strike the saving from the sheet and from the minister's briefing.
It is not legal advice about how to structure a GeM bid. It is the arithmetic you should attach so the bid has a spine.
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.
Vendors will not fill a departmental sheet.
Then they are not bidding on your estate. They are bidding on their brochure. Ineligible.
We cannot estimate tokens yet.
Then you are not ready to buy a metered path. Buy a time-boxed discovery with a rupee ceiling, or go on-prem with a known card count. Guessing a meter is how denial of wallet starts.
On-prem looks worse because we have to show power.
Good. Show it. Then show the hosted abuse review, the overage, and the exit. Honesty is the comparison. Hidden power is how on-prem gets a reputation it has not earned; hidden meters are how APIs get a reputation they have not earned.
Three years is too long; the models will change.
That is why block G exists. Models will change. Your traces, prompts and unit must still be yours. A three-year sheet with an exit drill is how you survive the change.
Ten days to a signed sheet
Do this once. Reuse the file. Change the unit row for the next directorate.
- Day 1: copy the seven blocks into a departmental spreadsheet. Protect the header row.
- Day 2–3: fill block A from twenty timed cases. No vendor in the room.
- Day 4–5: send the blank B–G rows to every bidder with a deadline. Same sheet, both columns if they offer both paths.
- Day 6: CISO marks any undeclared tool, egress or missing log row as a fail, not a deduction.
- Day 7: finance loads people and tax. Strike efficiency gains without a measurement method.
- Day 8: compute year 1 and years 1–3. Print.
- Day 9: tabletop the denial-of-wallet SMS and the exit drill. If either is theatre, raise F and G.
- Day 10: programme owner, finance and CISO initial. File. Only then open commercial envelopes against a number you understand.
How this shows up in the file
Subject: TCO sheet for [workflow] — assumptions dated [date].
Attached is the departmental seven-block calculator. Unit: [sentence]. Monthly volume: [n]. Paths compared: [hosted / on-prem / both]. Hard spend or GPU cap: [₹ or hours], owner [post]. Log retention: 180 days in [location], plus departmental record schedule. Exit artefacts listed in G1. Efficiency gains: none / [method attached]. This sheet is an internal pricing aid. It is not a tender and not legal advice.
Any bid that cannot populate B–G on this sheet is incomplete. Blank cells are not zeros unless a reason is written.
This article is informational field guidance for Indian public institutions, not legal, procurement, security-accreditation or engineering advice. Confirm against the current Gazette, GFR, GeM term, CVC instruction, CERT-In direction, DPDP text, departmental manual and your counsel before you file it.
Questions this usually raises
- Can we put this sheet in an RFP?
- Yes, as an annexure the bidder must return. Keep the unit and people costs in the department's column if those numbers are sensitive. Require the vendor cells either way.
- Is a vendor TCO tool acceptable instead?
- No. You will be comparing different religions. One departmental sheet, two columns, dated assumptions.
- Where do DPDP costs sit?
- Mostly in C (DPO time), D (stores you must be able to erase), and E (rights-request drills). Do not create a vague 'compliance' lump. Map the duty to a cell.
- What about CERT-In log cost?
- Block D. Specified ICT logs for a rolling 180 days in Indian jurisdiction are a live duty from the 28 April 2022 directions for covered entities. Price the store. Do not invent a new CERT-In 'AI log' product unless you also keep the statutory floor.
- How often should we recompute?
- At go-live plus six weeks, then quarterly, and after any model or price change. A sheet that is not recomputed is a souvenir.
- Will Prcept fill the sheet?
- We will fill our cells. We will not fill your unit or your people. If our on-prem column loses on an honest sheet, it should lose.
Sources
- Department of Expenditure — General Financial Rules, 2017
- Department of Expenditure — Delegation of Financial Powers Rules, 2024
- CERT-In Directions under Section 70B, 28 April 2022 (PDF)
- Digital Personal Data Protection Act, 2023 (India Code)
- Prcept AI — on-prem / air-gapped agents
- India AI Governance Guidelines (PIB document, November 2025)