New Delhi: Uttar Pradesh’s tender for a statewide artificial intelligence platform lets government departments use “frontier models”—meaning commercial AI models run by outside providers—through “their own subscriptions”. The tender allows this as the only exception to its rule that all government data, logs and backups on the platform must stay within India.
The tender limits this use to a “development sandbox”: a separate test setup walled off from the government’s live systems, where software is built and tried out before it reaches the public. The models may then be used only to help write computer code.
No real government data or citizens’ records may be sent to them, only “synthetic data”: made-up data that mimics the format of real records but contains no real information.
Providers must operate under terms that include “zero data retention and no training on customer content,” meaning they cannot keep the data or use it to train their models. Access is to be renewed every quarter, and the models may have no “direct or indirect access” to government data. Each department must get prior approval from the Uttar Pradesh Artificial Intelligence Mission (UP AIM) before using a specific model. The tender does not name the providers.
In January 2025, the Union Finance Ministry asked its staff to avoid AI tools such as ChatGPT and DeepSeek on office devices, citing risks to the confidentiality of government data. The UP tender instead builds paid access to such models into the state’s platform, with departments holding the subscriptions.
On 29 October, the Centre for e-Governance (CeG) issued the 178-page Request for Proposal for UP AIM. It sought a company to build and run the “UP AI Stack,” a shared platform on which every state department will build AI applications, for 36 months, with an option to extend by 24 months. The Internet Freedom Foundation (IFF) first reported the tender.
The tender’s applications cover several government departments, including revenue courts, property registration, urban development, police, education and healthcare.
The platform will draft land dispute orders in the state’s revenue courts. The Revenue Department’s assistant, connected to the Revenue Court Case Management System (RCCMS) portal that tracks revenue court cases, will “prepare draft responses and orders,” summarise pleadings and rank cases for hearing by age, deadlines and stay orders. The tender says the system is designed to handle about 10,225 revenue court officers and staff as users. It says the system will not “autonomously pass judicial or quasi-judicial orders,” and every draft must be reviewed.
The Stamp and Registration Department’s tool will generate risk scores for property transactions, flag suspicious patterns and prioritise high-risk transactions for scrutiny. It will mark revenue-leakage hotspots by district and tehsil and predict future stamp duty and registration revenue from past collections and circle rates. The tender says the tool is designed to handle 7,200 registry transactions a day, with one AI analysis of each.
The tender also seeks a system for the Urban Development Department to monitor “publicly available social media content including online news channels”. The system will “analyse publicly shared complaint images” through “image and object detection”, and extract names and locations from posts, “tags, landmarks, and available metadata”. The stated purpose is to detect civic complaints such as waterlogging and broken streetlights. Departments can “add or modify issue categories and keywords”, and complaints can be routed “without prior human approval”. The tender lists things the police tool in the same document must not do, such as identifying suspects. It sets no such limits for the social media system.
In 2024, the state’s Digital Media Policy prescribed punishments of up to life imprisonment for “anti-national” posts. In December 2025, the UP Anti-Terrorism Squad sought bids on the GeM portal for a “Social Media Monitoring and Analysis Tool”.
Clause 15.5.2 makes Hindi, English and Urdu “mandatory” for all citizen-facing services. In February 2025, Adityanath told the UP Assembly that the Samajwadi Party wanted to teach children Urdu and make them “maulvis”. He was responding to a demand that House proceedings be translated into Urdu, and in the same speech used a term the Opposition called derogatory. Urdu has been the state’s second official language since 1989.
AI agents on the platform, programs that can carry out tasks on their own, are limited to the actions approved before launch. Under the tender, “any increase in an agent’s autonomy after Go-Live” needs only the “prior written approval of UP AIM”.
The tender’s other applications include a police assistant for 15,000 personnel to search FIRs and case histories, and a tool to draft replies to RTI applications and Legislative Assembly questions. They also include a lesson-planning assistant for 6.5 lakh teachers, a voice assistant for ASHA workers, a meetings tool that identifies speakers, and a citizen app called AI Mitra, IFF reported. IFF also reported that the platform’s data layer is to “use Family ID as the linking key” for departmental records, and that the scheme rests on a Government Order rather than a law.
Startups hired by departments may build their own applications on the platform. The company that wins the contract will review each one before it goes live and “may recommend rejection, remediation, conditional approval or suspension”. On 9 October, the Cabinet approved a scheme that lets departments buy directly from startups.
The AI Stack tender, however, requires the lead bidder to have an average annual IT turnover of Rs 500 crore and 1,000 IT staff in India. It must also hold CMMI Level 5 certification, the top grade of a global benchmark for software companies, which usually only large IT firms hold. Startups can join only as partners in a group bid led by such a firm.
The tender states no estimated cost.
Bidders deposit Rs 2 crore as earnest money. Cloud and GPU charges, the cost of the computing power that runs the AI, are paid by actual use, outside the fixed bid amount, and the tender says its infrastructure quantities do not constitute a “ceiling”. The contract runs three years, but bids are scored on one year of cloud rates. The winning company must also set aside 1,000 days of staff work for any changes the mission orders, at no extra cost.
The tender comes eight months after the Puch AI episode. On 23 March, Invest UP signed a Rs 25,000-crore MoU with the Bengaluru startup for AI parks, data centres and an AI university. The company had reported annual revenue of Rs 42.9 lakh. Chief Minister Yogi Adityanath announced the deal on X. After the Opposition questioned the firm’s capacity, the state cancelled the MoU three days later, citing “lack of net worth and credible financial linkages”.
Adityanath had announced the AI Mission in January as a Rs 2,000-crore programme over three years. The 2026-27 budget provides Rs 225 crore. The Cabinet approved the Mission on 9 October, and CeG floated the tender 20 days later.
Bids close on December 7. Implementation runs 12 months from the work order. The Assembly’s term ends on May 22, 2027, and polls are expected in February-March, so the platform will be built and run largely under the next government.
CeG and UP AIM did not respond to ThePrint’s emailed questions.
(Edited by Anupama Yadav)
Also Read: Culture ministry launches new AI search engine. Next target: AI-led tracking of stolen antiquities
