INDUSTRIES · GOVERNMENT AND PUBLIC SECTOR
AI for the public sector, on infrastructure you control.
Files read and replies drafted for an official to sign. Citizen identifiers masked before a model reads them, a deployment on your own tenant or hardware, and a recorded trail an auditor can open.
IN SHORT
For government departments and public-sector bodies, AgentAnywhere is designed to run on your own tenant or your own hardware, in your facility. Aavaran, our data-masking layer, masks citizen identifiers before a model reads them, Manthan reads long files and drafts replies for an official to review, AgentAnywhere Vaani is designed to take voice notes in Indian languages on Indian infrastructure, and every step is recorded. Top secret and secret workloads are not permitted on cloud under the MeitY memorandum of 20 March 2026, and we would discuss those only as an on-premises deployment.
01 · Citizen records are the data
Grievances, case files and records carry names, identifiers and benefit details. They are the data a department is least able to send to a foreign service.
02 · Classification decides the road
MeitY's March 2026 memorandum does not permit top secret and secret workloads on cloud. For other data, the cloud route depends on the category of the data.
03 · An official signs, and the file must show it
The AI's part has to be visible in the file, so that a successor or an auditor can see what was drafted, what was changed and who approved it.
WHY IT FITS
Sovereign, our models, and the Trust Platform.
Three things, in the terms of government and the public sector: where it runs and who controls it, which models do the work, and the controls and the record around every call.
01 · Sovereignty
It runs on your tenant or your hardware, in your facility, under your authority.
- Designed to deploy on your own tenant or hardware, in your facility, region-pinned, and designed to deploy air-gapped.
- For top secret and secret workloads, which MeitY's memorandum does not permit on cloud, we would discuss only an on-premises deployment.
- Swaraj, the sovereign deployment, is designed to pin a deployment to the region you choose.
02 · Our models
The line we select, measure and serve, for the work of a department.
- Manthan reads long documents, synthesises them and drafts replies and options for a person to decide on, in English and Indian languages.
- AgentAnywhere Vaani is designed to take voice notes in Indian languages on Indian infrastructure. Its masking, routing and tally are ours; speech recognition and synthesis are open-licensed external components used for now.
- Taksha builds software on your own network. Manthan and Taksha are built on open foundation models, measured and served by us. You can bring your own open model.
03 · The Trust Platform
Controls around every call, and a record the department keeps.
- Aavaran masks citizen identifiers before a model reads them, with a signed Trust Receipt for each masking call.
- Kavach is designed to screen each request for prompt injection and unsafe output. The Agent Gateway applies your policy at the point of the call.
- Every step is recorded, in a trail you can query and export as CSV.
INSIDE YOUR PERIMETER
What happens to one citizen's request.
Follow one request from the moment it arrives to the receipt an auditor can check. Raw values are designed not to cross the boundary.
One request · start to receipt
Illustration
Your perimeter · your cloud, your region, or designed to deploy air-gapped
Step 1 of 5
A request arrives
A question comes in. It carries a name, an account, a card number: the things that must not travel.
THE RULES
The rules that bind you.
What each instrument requires in plain words, where it stands today, and what it means for AI in government and the public sector. Dates that have not arrived are written as dates.
MeitY Office Memorandum F. No. 9(3)/2025-EG-II, 20 March 2026: cloud selection framework
Indicative, not mandatory
What it requires
Indicative: it does not impose any mandatory obligations. Top secret and secret workloads are not permitted to be hosted on cloud. Category A data can use NIC and State Data Centre services, or sovereign cloud providers as and when notified.
What it means for AI
Where the work runs on cloud, the provider follows the data category. For top secret and secret work there is no cloud route, and we would discuss only an on-premises deployment.
What it requires
The exemption covers only instrumentalities the Central Government notifies. It is not a blanket exemption for government: for other State processing, most of the Act still applies.
What it means for AI
Plan for citizen data as personal data under the Act unless your body has been notified. Whether it has is for your counsel.
What it requires
Commenced by G.S.R. 843(E) of 13 November 2025. Section 8(5): protect personal data, including data with a processor, by reasonable security safeguards. Rule 7(2)(b): a fuller breach report within 72 hours. Rule 6(1)(e): logs and personal data kept one year.
What it means for AI
A hosted model service can be a data processor. Masking before the model reads, and a trail you can query after an incident, help you show the safeguards and rebuild what happened.
What it requires
Government can declare a computer resource that affects critical information infrastructure a protected system.
What it means for AI
Any AI added to a system that is declared protected sits inside that status. Plan it as part of that system.
What it requires
Report listed incidents within 6 hours of noticing. Keep ICT logs for a rolling 180 days within Indian jurisdiction.
What it means for AI
An AI system produces ICT logs. A recorded trail that stays in your deployment is the material for the log duty. Reporting stays with you.
What it requires
Guidance. No separate AI law is needed at this stage; existing laws govern AI.
What it means for AI
There is no AI statute to wait for. The rules above apply to AI systems as they apply to any other system.
This is what the rules say, not legal advice. Your compliance function owns the reading and the filing.
USE CASES · BY FAMILY
What runs inside a department.
Each card opens the page for that family or machine. Tell us your setting and we will show you the closest demonstration.
WORK SCENES
What it looks like on a working day.
The problem, what runs, and what the person at the desk gets. Each scene is an illustration of how the work is designed to go, not a customer account.
Scene 01
A grievance queue
The problem
- Grievances arrive as plain text carrying names, phone numbers and ID numbers.
What runs
- Aavaran masks the identifiers, and Manthan summarises and groups the grievances on the masked text.
What they get
- The officer reads a summary and a category for each one; where a name is needed for the reply, unmasking is policy-gated and recorded.
Scene 02
A long file to answer
The problem
- An officer must read a thick file before drafting a reply.
What runs
- Manthan reads across the file and drafts a reply with the options and their trade-offs.
What they get
- The officer edits, decides and signs, and the drafting step is recorded.
Scene 03
A helpline in the citizen's language
The problem
- Citizens call or send voice notes in many languages, and the recordings are sensitive.
What runs
- AgentAnywhere Vaani is designed to understand, mask, classify and route each note on Indian infrastructure.
What they get
- The right desk receives a masked, categorised request, urgent ones go to a person, and only an anonymised tally is kept.
Scene 04
A portal built in-house
The problem
- The department's code cannot be sent to an outside tool.
What runs
- Taksha works on your own network, and an independent referee checks each claimed success.
What they get
- Your reviewers receive each change with the referee's verdict, and your release process decides what ships.
Scene 05
The audit asks what the AI did
The problem
- An auditor wants to see what the AI system read, drafted and who approved.
What runs
- Every step is recorded, with a signed Trust Receipt for each masking call.
What they get
- The department exports the trail as CSV and hands it over.
THE SOVEREIGNTY TEST
The sovereignty test, for government and the public sector.
If an AI system cannot answer these from its own records, you are renting it. Beside each question: what a rented AI service usually answers, and how we answer. Read the test.
01Where does the data run?
What a rented AI service usually answers
“Hosted in India.” The vendor decides when the model changes.
How we answer
On your own tenant or hardware, in your facility, region-pinned and designed to deploy air-gapped. For top secret and secret work, an on-premises deployment only.
02Which model made that decision?
What a rented AI service usually answers
“The AI.”
How we answer
Model Hub is designed to register each model with an owner and lineage. Manthan and Taksha are built on open foundation models, measured and served by us.
03What did the model actually see?
What a rented AI service usually answers
The citizen's full record, identifiers included.
How we answer
Aavaran masks identifiers before the model reads them, with a signed Trust Receipt for each masking call.
04What stops the agent doing harm?
What a rented AI service usually answers
A prompt that tells the model to be careful.
How we answer
Your policy is applied outside the model. Kavach is designed to screen each request for prompt injection and unsafe output, and the Agent Gateway applies the policy at the point of the call.
05Who approved the consequential step?
What a rented AI service usually answers
“The system.”
How we answer
A named official approves what goes out, and the approval is recorded.
06What will you show the regulator?
What a rented AI service usually answers
A presentation to the audit committee.
How we answer
A recorded trail, queryable and exportable as CSV, with a signed Trust Receipt for each masking call. Your department owns the reading of the rules.
STATED PLAINLY
What this is not.
- Top secret and secret workloads are not permitted to be hosted on cloud under the MeitY memorandum of 20 March 2026. We would discuss such work only as an on-premises deployment.
- That memorandum is indicative and imposes no mandatory obligations. Your department decides how it applies.
- AgentAnywhere is software you run on your own tenant or hardware.
- A model's draft supports an official. The official decides and signs.
- AgentAnywhere Vaani is shown without accuracy figures, its first pilot deployment is being built, and its speech recognition and synthesis are open-licensed external components used for now.
- Masking reduces what a model reads. It does not take data outside any law, and your counsel owns that view.
THE OFFER · THE SOVEREIGN PILOT
Bring your hardest workload. We come to your data.
Pick the one workload your risk committee keeps deferring because the data cannot go anywhere. We run the pilot where the data already sits, with our architects beside your team, and you decide at the end whether it goes live.
For departments and public-sector bodies: bring the grievance queue or the file backlog that cannot go to a foreign service. We come to your tenant or your hardware.
What the pilot is
- One workload, one team, six to ten weeks.
- A sovereign environment by default.
- Solutions Architecture paired with your team for the duration.
- It ends with one agent in production under your governance, or a documented decision not to proceed.
What we add
- The Trust Platform is on from the first call: data masking, policy screening, the Agent Gateway and a recorded trail.
- The record stays with you and is exportable as CSV.
- You can bring your own open model.
- A reply within one business day.
A risk committee has no reason to refuse, because the data does not have to go anywhere to find out. The pilot ends one of two ways: one agent in production under your governance, or a documented decision not to proceed, with the record behind it.
OTHER INDUSTRIES
Other industries
FAQ
Frequently asked questions.
- Can a government department use AI without sending citizen data abroad?
- Yes. AgentAnywhere is designed to run on your own tenant or hardware, in your facility, region-pinned and designed to deploy air-gapped. Aavaran masks citizen identifiers before a model reads them, and every step is recorded.
- Can top secret or secret work use cloud AI?
- No. MeitY's cloud selection memorandum of 20 March 2026 says top secret and secret workloads are not permitted to be hosted on cloud. We would discuss such work only as an on-premises deployment.
- Which cloud can a department use for which data?
- MeitY's memorandum of 20 March 2026 is indicative and does not impose mandatory obligations. It says Category A data can use NIC and State Data Centre services, or sovereign cloud providers as and when notified. Your department decides how to apply it.
- Does the DPDP Act exempt government?
- Not as a blanket. From May 2027, sections 17(2)(a) and 17(4) exempt only instrumentalities the Central Government notifies. For other State processing, most of the Act still applies. Your counsel decides which applies to your body.
- Does India have a law for AI?
- Not a separate one. MeitY's India AI Governance Guidelines of November 2025 are guidance, and they say no separate AI law is needed at this stage because existing laws govern AI.
- What can Manthan do for a department?
- Manthan reads long documents, synthesises them and drafts replies with the options laid out, in English and Indian languages, for an official to review. It runs inside your perimeter and is built on open foundation models, measured and served by us.
- How does a department show an auditor what the AI did?
- Every step is recorded, each masking call carries a signed Trust Receipt, and the trail is queryable and exportable as CSV. Your department owns the reading of the rules.
- What does the Sovereign Pilot involve for a department?
- One workload, one team, six to ten weeks, in a sovereign environment, with Solutions Architecture paired with your team. The Trust Platform is on from the first call, the record stays with you and is exportable as CSV, and it ends with one agent in production under your governance or a documented decision not to proceed.
Bring your hardest workload.
We come to your data. One workload, one team, six to ten weeks, in a sovereign environment. Tell us the workload and a solutions architect will reach out within one business day.