Cost of Adding One More Language, Modelled

A finance committee asked how much it would cost to 'add Tamil' to an agent that had been accepted on Hindi and English. The vendor said a number that looked like a fine-tune on a rented GPU week. The CIO added 18 percent. The file moved. Six months later the department was still paying two translators to rewrite every failure message, the kiosk font was missing a glyph, and the Chennai desk had quietly gone back to the old IVRS. The fine-tune invoice was the smallest true line.

This is a costing model, not a price list. We will not print a rupee figure and pretend it is a 2026 market survey. Hardware rates, annotator wages and IndiaAI hours move. What does not move is the shape of the pile. If your note only has the model line, your note is fiction.

Use the model to compare languages, not to decorate a cabinet slide. A language with a mature Unicode stack, a staffed desk, and last year's tickets is cheaper than a language with a new official script, no font on thin clients, and twenty volunteers.

Not financial advice. Your finance department, GFR object heads and any IndiaAI eligibility letter still govern how you pay.

Cost the pile, not the checkpoint

A language is a service. Services have discovery, build, operate and refresh. The checkpoint is one build input. Sometimes you do not fine-tune at all. You write prompts, you constrain tools, and you spend the money on eval and fallback. Pretending that GPU hours are the cost is how English-first vendors win change-requests later.

Discovery: ticket count, scripts, code-mix, speech or text, whether official instruments must be emitted, whether RTL or complex shaping applies. A week of honest discovery saves a quarter of fake build.

Build: eval set, annotators, IME and fonts, UI chrome, agent behaviour, PDF path, numeral contract, and a human desk that can actually speak the language. Operate: triage of LANG tags, refresh of scheme names, font licences, and the desk roster. Refresh: a yearly re-eval because schemes rename themselves.

The drivers that actually move the number

Script engineering. Perso-Arabic Urdu and some Northeast orthographies are not 'plus one string file'. See the RTL guide. If the SI estimate does not mention fonts and PDFs, it is not an estimate.

Register. Secretariat noting is cheaper than citizen SMS code-mix, which is cheaper than legal instruments that must be bilingual under a live official-language duty. Do not cost a chat and then emit licences.

Speech. A text language plus IVRS is two languages in this model. Acoustic data is its own pile.

Desk substitution. If you already employ competent speakers on a helpdesk, year-one agent quality can be lower and the desk takes the tail. If you do not, you are buying a desk whether the file says so or not.

Relative weights for a planning conversation. These are order-of-magnitude shares, not a study and not a quote.

Pile item	Often small when	Often large when	Planning share (honest band)
Eval design + annotation	You already have tickets and markers	You must collect speech or a new script	Often the largest year-one line after people
Model adapt / prompt / tools	You only add a constrained FAQ	You want open generation on legal text	Sometimes modest; do not assume it dominates
Client engineering (font, IME, PDF, RTL)	Latin / Devanagari already in the stack	RTL, Mayek, missing thin-client fonts	Can exceed the model line
Human desk (year)	A roster already speaks the language	You must hire or translate through English	Usually larger than GPU increment
Yearly refresh + LANG triage	Schemes are stable	Mission-mode renames every quarter	Budget as operate, not as surprise

How to model without inventing rupees

Count person-weeks: linguist or language officer, two community markers, one SI engineer, one tester, one desk lead. Price those weeks at your actual salary-plus-overhead, not at a consultant day-rate from another State. Add font licences as actual quotes. Add GPU hours only if you will really train or heavily evaluate, and take those hours from the live IndiaAI or GeM rate card on the day you file — not from this article.

Show three scopes. Scope A: UI chrome plus a human desk, no agent generation. Scope B: agent on FAQ and status, eval of 200 items, no instruments. Scope C: instruments, PDFs, speech. Committees like one number. Give them three. The jump from B to C is where projects die.

Do not annualise a one-week fine-tune and call it TCO. The desk and the refresh are the TCO.

When the model says do not add it this year

When you cannot find twenty real strings. When you have no desk and no hire plan. When the official script is still in political motion and the tickets are in another script. When the only pressure is a map colour. Write the refusal. Fund the corpus. Add the language when the pile is real.

A delayed language with a desk is more honourable than a live language that silently gates to English.

Two cost notes, one of them complete

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.

Just give us a rupee per language so we can multiply by twenty-two.

That multiplication is how you write a false budget. Languages are not identical SKUs. Use the pile and three scopes.

IndiaAI subsidy will make the incremental language free.

Subsidy, if you are eligible, may cut some compute. It will not hire markers or fix PDFs. See the IndiaAI guide. Do not paste a percentage into this model.

Open weights mean the language is already included.

Included in a crawl is not included in your scheme names. Eval still exists. Desks still exist.

We will know the cost after the pilot.

Then the pilot must be scoped as discovery, not as implied go-live for five more languages. Discovery without a cost model becomes a free option on the department.

Three weeks to a language cost note finance will not laugh at

Sit finance in week one, not at the end for a rubber stamp.

1. Week 1: pick one candidate language. Pull ticket share. Write whether speech, paper and official-language instruments are in scope.
2. Week 2: collect actuals — salary bands, font quotes, existing desk roster. Leave GPU as 'live rate card, attached printout', not a number from memory.
3. Week 3: publish Scope A/B/C. Recommend one. If you recommend delay, attach the corpus plan so it is not a political vacuum.

File note you can paste

Subject: Incremental cost model for adding [language] to the citizen agent.

This note does not quote a market survey. It costs discovery, eval, client engineering, desk, refresh and any compute at live rate cards attached. Three scopes are presented. The recommended scope is [A/B/C]. GPU unit rates and any IndiaAI subsidy are taken from printouts dated [insert], not from vendor memory.

A single rupee-per-language figure is not adopted. This is not financial advice.

What we will put in the estimate

Prcept AI will show the pile. We will refuse a one-line 'add Tamil' quote if paper or speech is hiding. We will price a desk-only year when that is the honourable scope.

If a competitor's language is cheaper because they omitted fonts, PDFs and markers, they are not cheaper. They are later.

This article is informational field guidance for Indian public institutions, not legal, language-policy, procurement, finance or engineering advice. Confirm against the live Gazette, Official Languages Act and Rules, your State's official-language law, MeitY / IndiaAI notices, GFR, GeM terms, DPDP text, departmental manuals and your counsel before you file it.

How to test this with real speech, not staff English

“Cost of Adding One More Language, Modelled” fails in the field if you only tested officers. A P1 CIO/CTO should hear a first-generation student, a rural caller, or a Hinglish grievance before claiming “cost adding language AI”.

Adding Odia after Hindi is not a line item called 'model'. It is a pile of eval design, annotators, fonts, RTL or shaping, a human desk, and a yearly refresh. Model the pile. Do not buy a percentage off a slide. Twenty-two scheduled languages is a Constitution fact, not a model fact. Script support is not language support. Official language rules may require bilingual output even when the model prefers one script.

• Name the languages and scripts in the eval set.
• Include code-mix and scheme-name tests.
• Measure comprehension, not BLEU alone.
• Design a human fallback when language fails.

Close this loop before the next CAB

Put “Cost of Adding One More Language, Modelled” 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. “cost adding language AI” is not a one-time workshop. It is a watch item. Date the last check. Unsigned watch items are souvenirs.