Why does this matter now?
Artificial intelligence can help India’s security institutions notice important signals sooner, reduce repetitive work and make better use of scarce specialist time. Whether it improves a decision depends on what the system is allowed to do, how well it works under local conditions and who is accountable when it gets something wrong. A tool that prioritises a maintenance inspection faces a different standard from a system that informs an action affecting a person’s safety or rights. Keeping those distinctions clear matters more than attaching the label ‘AI’ to a system.
Where is AI already relevant to India’s security?
At its July 2022 AI in Defence symposium, India’s Ministry of Defence announced 75 AI products and technologies spanning areas such as cybersecurity, logistics, surveillance, operational analytics and manufacturing. It highlighted three examples from public sector defence manufacturers: voice transcription and analysis, driver fatigue monitoring, and assistance with spotting welding defects in X-ray images. These examples illustrate the variety of tasks AI can support. The launch announcement is a record of what was presented at the event. It does not establish how widely any individual system was subsequently deployed or how well it performed in service.
Start with a use case where a qualified person can check the output. A maintenance team could use image analysis to rank equipment images for inspection. The operator would still see the original image, record why they accepted or rejected a suggestion, and retain responsibility for clearance. If the system performs poorly on images from a particular camera or production line, the team can narrow its use or fall back to the normal inspection process. That design makes it possible to improve throughput and measure errors without handing the final decision to software.
What makes a security use case harder than a demo?
A model can score well on a tidy dataset and fail when the operating environment changes. Image quality varies with weather, lighting and equipment. Transcription quality can vary across Indian languages, accents, background noise and code-switching. A cybersecurity model trained on one network may flag legitimate work on another. In each case, averages can hide the error that matters most. Evaluators should measure missed events as well as false alerts and report results separately for the settings where the tool will be used. They should also measure how much work a reviewer needs to confirm or correct the output.
Security testing has to ask what happens when inputs are misleading or unavailable. An incident response team using AI to summarise alerts should be able to trace the summary back to the original records. A log entry or retrieved document may contain text intended to mislead an AI system. It belongs in the evidence the system examines, never in its operating instructions. If the model, data source or connection fails, staff should know how to continue with the original workflow. Keep any ability to take consequential actions outside a tool built only to summarise information.
Why does India’s AI infrastructure matter?
In March 2024, the Union Cabinet approved the IndiaAI Mission. The announced programme included public-private AI compute capacity, support for Indian AI innovation, access to datasets, and work on safe and trusted AI. That investment can help Indian researchers and suppliers test and maintain systems for local needs. It does not make every model suitable for sensitive workloads. Procurement still needs to account for who can access training and operational data, where each type of data is handled, what happens if a supplier changes terms, and whether an authorised team can reproduce the results.
Domestic capacity also depends on the less visible parts of a deployment. A locally developed application may still rely on a model API, an external logging service or specialised hardware. Teams should document each dependency and decide what data it can see. They should test operation when a dependency becomes unavailable, and agree on export and replacement arrangements before deployment. For some work, a small model running in a controlled environment will be easier to test and operate than a general-purpose system. The right choice follows from the task and the risks of failure.
What does meaningful human oversight look like?
A named reviewer should have enough time, evidence and authority to disagree with the model. A green indicator or a confidence score alone is insufficient. Interfaces should show what the system saw, what it omitted and when the input falls outside tested conditions. For a critical decision, an organisation can require corroboration and a clear escalation route. It should also define who can suspend a system and keep a record of which model version, inputs and human decisions produced each outcome. An audit trail helps only if someone can reconstruct an incident from it.
The NIST AI Risk Management Framework offers a practical vocabulary: govern, map, measure and manage. A team can use those functions to assign responsibility, describe how the tool will be used, test risks and decide when to intervene. The framework is a reference for organising that work. Indian institutions and suppliers still need to apply their own legal, contractual and operational obligations to each deployment.
What should an Indian supplier prove before a pilot expands?
First, define one decision the product will support and one action it has no authority to take. Use representative data that the buyer can lawfully and appropriately use. Agree in advance on error thresholds, review time, logging, access, an incident procedure and the conditions for switching the tool off. Run in parallel with the existing process long enough to measure whether the tool actually changes a useful outcome. Ask the operators what evidence they need to act and how often they would reject a suggestion. Recheck those assumptions when the model or input source changes.
A credible first deployment might shorten the time it takes to identify a suspect component, direct an analyst to relevant records, or identify a duplicate alert. Each result can be measured against the current process. Reserve broader authority for cases with stronger evidence, a clear public purpose and meaningful safeguards. India’s advantage will come from systems people can test, maintain and challenge under real conditions. The aim is more dependable decisions, with responsibility still where it belongs.
Research sources
Ministry of Defence: AI in Defence symposium, July 2022 — https://www.pib.gov.in/Pressreleaseshare.aspx?PRID=1840740&lang=2®=48
MeitY: Cabinet approval for the IndiaAI Mission, March 2024 — https://www.pib.gov.in/PressReleasePage.aspx?PRID=2012357&lang=2®=48
NIST: AI Risk Management Framework core — https://airc.nist.gov/airmf-resources/airmf/5-sec-core/
AI and automation become valuable when they are useful, controlled, tested, and connected to the real customer journey.
Ready to put this into action?


