By Prasaanth Balraj, Associate Director, AI and Digital Transformation, APLYD
India does not lack AI pilots. Health departments have tested diagnostic tools, districts have tried chatbots for grievance redress, and state agencies have experimented with models that flag tax fraud or forecast crop yields. As of early 2026, the government had identified 762 potential AI applications across 62 ministries and departments, a scale of ambition few countries can match. What is missing is the layer that carries these experiments from a demonstration in a conference room into the daily routine of a block office or a municipal ward.
The implementation layer is the connective tissue between a working AI model and its everyday use inside a government office. An official needs to find the recommendation inside the software they already open each morning, not in a separate dashboard nobody checks. A technically sound model can still go unused for years without that connection. This layer includes the training that lets a department run and troubleshoot a system on its own rather than renew a vendor contract indefinitely. It also includes the governance rules around consent, access and correction that determine whether a citizen can question a decision made about them. This is the unglamorous half of AI deployment. It gets far less attention than the model itself, and it is the work many public-sector AI projects skip.
This pattern is not unique to India, but the stakes here are higher. Research has repeatedly shown that enterprise AI projects have a high failure rate, often because the challenge is not building the model but integrating it into how an organisation actually works. That pattern is familiar to anyone who has watched a government pilot move from a strong demo to a stall the moment it has to operate inside the messy, procedural, rule-bound world of real administration. An algorithm can be excellent and still fail inside government because government does not run on algorithms. It runs on file movements, sign-offs, service-level rules and staff already stretched across multiple duties.
Building Into What Already Works
A model that recommends the right action is not useful to a field officer who cannot act on that recommendation within the software, forms and approval chains they already use every day. This is why so many public-sector AI projects follow the same arc: a promising pilot, a favourable evaluation, a round of press coverage, and then silence. Once the pilot ends, the officials who ran it move to other postings, the vendor’s engagement winds down, and the tool has no natural home inside the department’s existing processes.
Bridging advanced AI models with ground-level administration means building for the systems that already exist, not around them. India’s digital public infrastructure, from identity and payments rails to health and education registries, gives the country an advantage that few others have: a base layer that AI can plug into rather than replace. The next step is to connect AI to these existing systems and workflows so that a recommendation appears where an official already works, rather than in a separate dashboard that must be remembered and checked. That distinction matters. The question should not be, “Where can we deploy AI?” It should be, “Where in an existing government workflow can AI remove a bottleneck, improve a decision or reduce repetitive work?”
Ownership, Not Dependency
There is a second, less visible problem. Many government AI deployments create dependency instead of capability. A vendor builds the model, tunes it and holds the institutional knowledge of how it works, while the department that owns the service is left to renew a contract each year rather than run and improve the system itself. This is outsourced intelligence with a public-sector logo attached to it, not sustainable delivery.
Real implementation has to include a transfer of capability. Government teams need to understand how to monitor model performance, work with new data, identify when outputs are unreliable and decide when a system should be overridden. The productivity gains from AI will only endure if the institution itself can sustain them. Without capability transfer, every change in administration or vendor relationship risks leaving behind a system that the department cannot maintain, audit or explain to the citizens it serves. The objective should therefore be clear from the beginning: when the technology provider eventually leaves, the capability should remain.
Governance Before Deployment
Citizen-first data governance has to sit alongside this capability transfer, not follow it. Public AI systems in India will increasingly touch health records, welfare eligibility and law-enforcement data. The institutions running them need clear rules on consent, storage, access and correction that citizens can actually understand and invoke.
Governance built after deployment tends to defend a decision already made. Governance built into the implementation layer from the outset protects both the citizen and the institution. This is particularly important as AI moves from advisory applications to systems that influence eligibility, prioritisation and service delivery. The implementation layer therefore cannot be treated as a purely technical integration function. It has to carry accountability into the way the system is actually used.
From Pilots to Public Infrastructure
India’s AI ambition is not short of talent or funding. What it needs is a discipline dedicated to making models function inside real institutions: adapting them to existing systems, building the capability within government to own and improve what it deploys, and putting citizen-first governance around their use. Three shifts can help build that implementation layer.
First, make capability transfer a procurement deliverable. Contracts should identify the government team that will operate the system, the skills it needs to acquire and the knowledge that must be transferred before the engagement ends.
Second, build governance in at launch. Consent, access, correction and accountability mechanisms should be established before a system goes live, rather than added after complaints arrive.
Third, measure adoption, not announcements. Ministries should ask whether an AI system has become part of a daily workflow, whether officials actually use its outputs and whether the department can sustain it after the pilot or vendor engagement ends.
None of this requires another layer of technology. It requires giving implementation the same priority currently given to model development. India has already demonstrated that it can build AI models and run ambitious pilots. The next test is whether those systems can survive contact with the institutions they are meant to serve. That is where the implementation layer matters.