Agents by 2028? The AI skills India's public servants will need

At least 80% of governments will deploy AI agents to automate routine decision-making by 2028, Gartner predicts (Gartner, 2026). An earlier wave of digital government moved forms and payments online. This one changes how decisions get made. For India's public servants, from a district office to a ministry, the skill that matters is knowing how to work alongside these systems and when to overrule them.

Where India already is

India has been building the base for this. Through the IndiaAI Mission, 38,000 GPUs are available at ₹65 per GPU-hour, and the Bhashini language platform supports 20 languages with more than 350 models (PIB, Dec 2025). By March 2026, 190 IndiaAI Mission projects had been approved, 78 of them with government entities (PIB, Mar 2026).

AI is already at work in public services:

• Courts. The SUVAS tool translates Supreme Court judgments into 18 Indian languages, and the e-Courts project has an outlay of ₹7,210 crore. The government's own wording is that AI "assists but never replaces judicial decision-making" (PIB, Feb 2026).
• Farmers. The Kisan e-Mitra chatbot has answered more than 93 lakh queries, over 8,000 a day, in 11 languages (PIB, Feb 2026). AI-based local monsoon forecasts were sent by SMS to 3,88,45,214 farmers in 13 states (PIB, Mar 2026).
• Schools. AI and computational thinking are being introduced from Grade 3 from the 2026–27 session, with teacher training through NISHTHA (PIB, Oct 2025).

NITI Aayog has recommended making AI literacy a foundational skill (PIB, Oct 2025). That recommendation applies to the people running these systems as much as to students.

What the day-to-day could look like

These are illustrations of possible uses, not descriptions of current government practice.

• A block development officer asks an approved assistant to summarise fifty grievance letters by theme, then reads the ones flagged as urgent herself.
• A court clerk uses a translation tool to prepare a first draft of an order in a regional language, then has it checked by someone fluent before it is issued.
• An agriculture extension worker uses a chatbot to answer a farmer's question on pest control, and checks the advice against the department's own guidance before passing it on.
• A policy officer uses AI to compare three state schemes and list their differences, then verifies each point against the official notifications.

In each case the system speeds up reading, sorting and drafting. The officer remains responsible for the decision.

1. 1.Use an approved tool

   Only the data allowed goes in.

2. 2.AI reads, sorts, drafts

   Summaries, translations, comparisons.

3. 3.Verify against originals

   Every citation, figure and rule quoted.

4. 4.The officer decides

   And can explain the decision to a citizen.

Skills public servants will need

Explaining a decision. Gartner expects that by 2029, 70% of government agencies will require explainable AI and a human in the loop (Gartner, 2026). If a citizen asks why their application was rejected, "the system said so" is not an answer. Officers need to understand enough about how a tool reached its output to explain it or challenge it.

Verifying before acting. A Bengaluru bench of the Income Tax Appellate Tribunal withdrew an order that cited judgments that did not exist (Taxscan, 2025). In February 2026, the Supreme Court of India said relying on fake AI-generated judgments "would be a misconduct" (MediaNama, 2026). Every citation, figure and rule quoted by an AI tool must be checked against the original.

Handling citizen data carefully. Government holds some of the most sensitive data there is. Staff need a clear sense of what may go into which tool, and a habit of using only approved systems.

Working across languages. With Bhashini and SUVAS in use, many officers will work with machine translation. Knowing its limits, and when a human translator must check the text, will matter.

Knowing where the human stays. Deciding which decisions an agent may make on its own, and which must always come to a person, is a design choice. Officers who understand the tools will be better placed to make it.

What stays human, and the risks

Accountability stays human. A citizen can appeal to an officer; they cannot appeal to a model. The government's own line on courts, that AI assists but never replaces judicial decisions, is a good rule well beyond the judiciary.

The risks are real. When Deloitte Australia delivered a government report with a fabricated court quote and non-existent references, it refunded more than A$97,000 (CFO Dive, 2025). Unchecked AI output in public work damages trust in institutions, not only in one report. Bias in data, errors in translation and over-reliance on confident answers are all reasons to keep a trained person in the loop.

What to do next

1. Find out which AI tools your department has approved, and what data may go into them.
2. Do the "try this week" exercise with a public document you know well.
3. Write down three decisions in your work that must always stay with a person, and why.
4. Before using any AI-translated text with citizens, agree with your team who checks it and how.
5. Share one verified, useful way of using AI with your team, along with the checks you applied.

Thinqmesh Academy teaches AI Fluency live in small cohorts, including judgement on what must stay with a person, and each month ends with a project you keep. See the programme.