State AI Mission Structures Compared

A secretary asked for a one-page comparison of state AI missions. The consultant delivered a league table: green ticks for states that had issued a policy PDF, amber for states that had announced a mission from a podium, red for states that had only a line in the budget speech. The table was already stale. One cell pointed at a society that had been merged. Another pointed at a CM cell that now sat under the IT department. A third treated a university centre as the state's production stack.

That is how bad advice is written. Missions are political and administrative instruments. They are renamed, merged, parked under a new principal secretary, and sometimes left as a letterhead while the actual work happens in the state data centre. Comparing them as if they were listed companies with stable articles is not analysis. It is scrapbooking.

This field guide is a method. Use it when a minister wants to know how other states did it, when a finance department wants to know who will own the servers, or when a vendor slide claims alignment with a mission that no longer exists in that form. We name structural types, not a current scoreboard. If you need today's chart, open the latest government order. Do not quote this page as the chart.

Prcept AI builds sovereign agentic systems for Indian public buyers. We will not sell you a fake map of twenty-eight missions. We will help you put decision rights, data classes and human gates on one page before the first GPU is indented.

Why league tables of state missions lie

A policy PDF is not a production system. A mission logo is not a lawful basis. A centre of excellence in a campus is not a citizen-facing workflow. League tables collapse those four facts into a tick. The tick then travels into a cabinet note.

Structures also change faster than blogs. A 2024 society can become a 2026 directorate. A special secretary can be given additional charge and the weekly meeting dies. A budget head can move from IT to planning. If your comparison cites a tweet as the current design, you are briefing fiction.

The honest comparison is therefore of types and of tests. Type answers: where does the work sit. Test answers: can that seat actually approve money, data and a speaking order. Two states can share a type and fail different tests.

Five structural types you will actually meet

Treat the names below as types, not as compliments. A state may run two types at once. Write the type you are walking into on the cover of the file.

Structural types — verify against the live government order, not this table.

Type	Typical seat	What it can do well	What usually breaks
Secretariat cell	IT / planning / CM office additional charge	Coordination notes, inter-departmental letters	No servers, no DPO time, no DFPR for hardware
Society or mission SPV	Registered society under a governing council	Hiring, workshops, some procurement agility	Unclear data fiduciary; council that never meets
SDC-led programme	State data centre / SeMT / e-governance agency	Hosting, network, backup, identity already exist	Treated as a hosting ticket; no policy owner
Line-department programme	Revenue, health, education, social welfare	Real files, real charters, real officers	One-off bots; no shared register; vendor sprawl
Campus or CoE overlay	State university / IIT / IIIIT partnership	Eval, language work, talent	Research MoU sold as a production SLA

Five tests that beat a logo

Score each type, or each live arrangement, on five tests. Write a sentence per test. If you cannot, the mission is a brand.

1. Decision rights. Who may approve a workflow that touches a citizen or money. Name the post, the instrument of delegation, and the meeting that can withdraw it.
2. Budget. Which head pays for metal, power, people and AMC. A mission that can hire a consultant but cannot buy a UPS is not a platform owner.
3. Data. Who is the fiduciary for each corpus. Where embeddings and traces live. Whether a department can refuse to donate a register.
4. Procurement. GeM, open tender, rate contract, or campus purchase. Who signs the indent. How CVC-facing fairness is recorded.
5. Human gate. Who signs the speaking order when the agent drafts. If the answer is the chatbot, stop.

What is worth copying, and what is theatre

Copy a published citizen-charter map. Copy a departmental AI register. Copy a rule that the SDC will not host a workflow without a lawful-basis sentence. Copy a monthly restore drill. Those travel across states.

Do not copy another state's vendor. Do not copy a named model. Do not copy a rupee figure from a press note. Do not copy an organogram dated last year without opening this year's order. Do not copy a CoE brochure into a production SLA.

IndiaAI and MeitY guidance can inform a state note. They do not appoint your fiduciary. They do not replace DPDP. They do not replace your charter. Cite them as context. Then write the five tests in your own file.

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.

Other states already have a mission, we will look late

Late is cheaper than a letterhead that processes personal data without a register. Announce after the five tests have owners. A quiet SDC landing with two reconstructable workflows beats a logo week.

The CM wants a single state model

A single inference cluster can be a hosting fact. A single model is not a lawful basis for every department. Split purpose, access and speaking-order owners even if the GPUs are shared.

The university will run it, they have GPUs

Campus metal is useful for eval. Citizen production on a research VLAN with student admins is a different risk class. If the CoE hosts, write processor terms, access, exit and who faces RTI.

We will update the comparison after the next reshuffle

Then do not put the old comparison in a sanction. Put the method and today's dated government order. Reshuffles are certain. League tables are not.

A four-week comparison you can actually file

1. Week 1: collect only primary papers — latest GO, budget head, SDC policy, DPO nomination, any AI register. Ignore vendor one-pagers and news explainers.
2. Week 2: classify the live arrangement into one or two types. Write the five tests in sentences with post names. Mark any test without a name as red.
3. Week 3: pick two departments that already have a citizen charter. Ask whether any proposed agent SLA would beat or break that charter. Write the answer into the comparison.
4. Week 4: circulate a two-page note: type, five tests, two red items, and what you will not copy from other states. Put the dated GOs in the annex. Do not attach a league table.

How this shows up in the file

Subject: Method for comparing state AI mission structures — not a ranking.

This department will not treat press notes or undated organograms as the current design of any state's programme. Comparison will be by structural type and by five tests: decision rights, budget, data fiduciary, procurement route, and human gate. Structures change; the tests do not.

No workflow will be hosted solely because another state announced a similar mission. Each workflow will carry its own lawful-basis sentence and speaking-order owner. This note is an internal aid. It is not legal advice.

What the comparison note must refuse to do

A comparison that ranks states by number of announced use cases will be quoted in a budget speech. It will then be used to justify a hosted tenant because 'we are behind'. That is not a method. That is peer pressure with a spreadsheet. Refuse rank. Refuse use-case counts. Refuse any cell that cannot name a government order dated this year.

When a vendor offers a ready-made heat map of state missions, ask for the date of each underlying GO. If they cannot produce the GOs, they produced a collage. Prcept AI will not publish such a collage either. We would rather lose a slide contest than put a merged society back on a live map.

District missions, municipal AI cells, and smart-city SPVs are not automatically children of the state mission. If a city wants an agent, it needs its own five tests. Borrowing the state logo does not borrow the state's fiduciary. Write that in the covering note so a municipal commissioner cannot claim the SDC already approved them.

Language policy is part of structure. A mission that sits in the IT department and publishes only in English is a different instrument from a mission that must serve Eighth Schedule languages through CSCs. Do not score those as the same type because both have a GPU. The citizen-facing duty is a structural fact.

• Attach the latest GO, not a news clipping, for every type you claim exists in your state.
• Name the DPO or equivalent even if the mission society wants to 'handle privacy operationally'.
• Write whether SeMT, NIC, or a private SI will hold root on the cluster.
• Record which departments have refused to donate a register, and honour the refusal until a lawful basis exists.
• If two types coexist, say which type is allowed to spend and which type is allowed to speak.

This article is informational field guidance for Indian public institutions, not legal, procurement, security-accreditation, engineering or board advice. Confirm against the live Gazette, GFR, GeM term, CVC instruction, CERT-In direction, UIDAI regulation, DPE guideline, DPDP text, departmental charter, plant safety manual and your counsel before you file it.