While India's ambitious AI initiatives successfully build advanced models for governance and public services, their real-world success depends heavily on human behavior, institutional incentives, workflow integration, and the careful balance between algorithmic automation and local human judgment.

While India's ambitious AI initiatives successfully build advanced models for governance and public services, their real-world success depends heavily on human behavior, institutional incentives, workflow integration, and the careful balance between algorithmic automation and local human judgment.

While India's ambitious AI initiatives successfully build advanced models for governance and public services, their real-world success depends heavily on human behavior, institutional incentives, workflow integration, and the careful balance between algorithmic automation and local human judgment.

India's artificial intelligence story is, by now, familiar in its arc. A pilot project dazzles with its precision: an AI model flags tuberculosis on chest X-rays; a chatbot delivers personalised agricultural advice; a grievance-triage system promises to sort citizen complaints before they reach a district collector. The IndiaAI Mission, approved in 2024, has a budget outlay of ₹10,371.92 crore over five years.

Yet, the difficult part of AI in government may begin after the model has been built. A technically successful system still has to survive fragmented workflows, limited time, uneven infrastructure, institutional incentives and the judgement of the people expected to use it.

The usual autopsies of technology projects blame data quality, patchy connectivity or creaky legacy systems.

These are real constraints, but not the whole story. What they miss is the human being expected to use the tool. We have grown skilled at building algorithms, but far less skilled at understanding the bureaucrat, nurse or extension officer who has to fold them into an already complex working environment.

Consider the district collector, monitoring dozens of schemes, managing disaster preparedness and fielding competing demands through multiple digital platforms that rarely talk to each other. For them, a new AI system can feel like one more interface to learn, one more password to remember and one more metric to report.

This is not simply a matter of attitude. Behavioural research shows that when attention and cognitive resources are constrained, people rely more heavily on familiar routines. India's frontline health workforce bears this out. ASHA workers already juggle a growing stack of government reporting applications, often in difficult field conditions. Research on Haryana's ASHA workforce has found that a proliferation of survey apps, combined with poor connectivity and ageing smartphones, has left many workers maintaining paper registers alongside digital ones, effectively doubling their workload for no added compensation. Similar patterns have been documented among community health workers elsewhere in the country: tools meant to lighten the load instead add administrative demands, forcing workers to reconcile records across systems that don't talk to one another. The lesson applies as much to AI-based decision-support tools as to reporting apps: technology should reduce workload, not redistribute it.

Then there is the question of incentives. Bureaucracy is never neutral towards risk, and its structure tilts towards the status quo. An official who follows an AI recommendation that departs from standard procedure, and that recommendation fails, must explain that decision to superiors, auditors and perhaps the courts. An official who ignores the recommendation faces no such scrutiny. The rewards for innovation are faint compared with the penalties for failure, so the safest algorithm is the one that confirms what was already being done.

Trust introduces another challenge. Some users fall into automation bias, accepting machine recommendations uncritically. Others display algorithm aversion, abandoning a system after a visible error even when it outperforms human judgement on average. Research has found that people can prefer human judgement to algorithmic advice even when the algorithm performs better.

The answer is therefore neither unquestioning reliance on algorithms nor automatic human override. A local official may know that a particular village's soil is saline, or that a Primary Health Centre is understaffed, in ways invisible to a model trained on large datasets. The machine may be better at spotting patterns across thousands of cases; the person on the ground may be better placed to recognise the exception. Keeping humans in the loop is not merely a regulatory safeguard; it is a behavioural and institutional necessity.

Agriculture makes the stakes clear. The ICRISAT-Microsoft AI Sowing App used historical crop and climate data alongside weather information to recommend sowing windows, expanding from an initial 175 farmers in Andhra Pradesh to more than 3,000 across Andhra Pradesh and Karnataka. But technology did not eliminate the need for human judgement. Recommendations on pesticides, fertiliser or planting times can have serious consequences, and local extension workers provided context and maintained trust. The challenge is not simply to make a recommendation accurate, but also usable and defensible.

This has implications for how governments procure and evaluate technology. At present, success can too easily become synonymous with deployment: whether the software was installed, the server commissioned, the training conducted. But deployment is not the same as use, and use is not the same as impact. A more meaningful evaluation would ask whether officials continue to use a system after rollout, when they override its recommendations, and why. Recording the reasons for accepting or rejecting an AI suggestion could strengthen accountability while creating feedback loops that distinguish genuine model errors from cases where the system simply lacked context.

The IndiaAI Mission already recognises safe and trusted AI as a core pillar. But trust cannot be created through technical safeguards alone. It also depends on whether systems fit existing workflows, whether officials can challenge recommendations, and whether technology genuinely reduces workload.

India's AI ambition is therefore not only a question of whether the country can build better models. It is whether those models can become useful parts of institutions that were not designed around them. That means funding not just software development, but workflow integration, continuous training, local validation and iterative refinement based on field experience, treating implementation as a living process rather than the final box to be ticked in a procurement file.

India's governance challenges are too complex for algorithms alone to solve. They are also too important to be undermined by tools that ignore how public servants actually work. If the IndiaAI Mission is to succeed, it will need to invest as seriously in behavioural design as it does in code. The question is not simply whether India can build intelligent systems. It is whether it can build systems that people can, and will, use intelligently.

(The author is a Research Associate, Centre for Social and Behaviour Change (CSBC), Ashoka University, Delhi-NCR. Views expressed are personal and do not reflect the views of the University or THE WEEK.)

TAGS