08 - 10 - 2026

Inside Indian Hospitals, AI’s Biggest Problem May Not Be the Technology

Artificial intelligence has moved quickly from a technology demonstration to a purchasing question for Indian hospitals. But a closed-door discussion among hospital executives, clinicians, founders and investors suggests that the industry’s biggest obstacle may sit well before the AI model itself.

The problem is whether hospitals have the data, infrastructure, evidence and workflows needed to make these systems useful in everyday care.

That was the central finding of the first Signals Roundtable, organised by The HealthTech Collective and HCHE in Bengaluru on September 26. The discussion brought together 12 to 15 healthcare leaders to examine where AI is already being used, where it is failing to gain traction and what hospitals would actually pay for.

The clearest indication came from a simple question posed to the participants: where is AI making the most meaningful impact inside hospitals?

Diagnostics and imaging received 10 votes, the strongest response by a significant margin. Documentation and communication followed with six. Clinical decision support and operations and administration received four votes each, while four participants said AI was not yet making a meaningful difference for them. Revenue cycle and billing received three.

The pattern is revealing. AI is gaining ground fastest where tasks can be clearly defined, outputs can be measured and the consequences of an error can be contained.

The harder applications involve clinical judgement. The data bottleneck comes before the model

Hospitals already generate enormous quantities of information through laboratory tests, scans, vital signs and clinical notes. The difficulty, according to the roundtable participants, is turning that information into something useful at the moment a clinician needs it.

Results often arrive as individual reports rather than as a longitudinal picture of a patient’s condition. Abnormal values may remain unnoticed unless someone actively searches for them. In time-sensitive cases such as stroke, early imaging findings can influence treatment decisions within a window of only a few hours, making the speed at which information reaches the treating team critical.

The discussion also identified four documents that carry much of a hospital’s clinical and commercial information: consultation notes, rounds notes, nursing notes and discharge summaries.

Their contents, however, do not necessarily translate into action.

A note recommending a sleep study or follow-up consultation may never become an order, appointment or bill. That creates a gap not only in patient follow-through but also in the hospital’s ability to track where potential revenue is being lost. Participants identified order capture in electronic medical records as one of the significant unresolved problems.

The underlying infrastructure remains fragmented. Hospitals often operate separate systems for administration, laboratory, radiology and pharmacy, while few have dedicated resources focused on converting the information generated by those systems into action.

Even basic requirements such as reliable internet connectivity and functioning integrations can become barriers to deployment.

That creates a practical lesson for AI companies: a sophisticated model cannot compensate for a broken workflow.

Clinical AI faces a higher bar

The threshold for clinical AI is considerably higher than for administrative automation.

Doctors work within an evidence-based system, yet the roundtable found that relatively little clinical AI has undergone the kind of peer-reviewed, clinically validated testing clinicians expect. Much of the training data used by existing models comes from other countries, leaving Indian patient populations comparatively underrepresented.

That creates a familiar problem when technology moves from controlled testing into real hospitals.

A system may perform well on data close to its developers’ original environment but encounter different accents, languages, background noise and patient populations in practice. Performance can deteriorate as those conditions change. Participants also noted that models can produce different outputs for the same input and that benchmarks can become outdated within months.

One proposal emerging from the discussion was to treat clinical AI more like medicines after deployment: monitor it for adverse outcomes, establish reporting mechanisms and feed problems back into the system.

The distinction between objective and subjective tasks was equally important. Monitoring glucose or blood pressure and raising alerts was seen as relatively low-risk. Clinical judgement, where a decision can directly affect a patient’s life, requires substantially stronger evidence. Narrow applications such as preliminary flagging of chest X-rays or ECGs were considered more plausible early clinical use cases, but participants still wanted meaningful accuracy data before relying on them.

Why hospitals may prefer smaller models

The roundtable’s discussion of model architecture was notable for what it did not prioritise.

Rather than assuming that bigger models and greater computing power would automatically produce better hospital applications, several participants identified small, task-specific models trained on Indian data as among the more promising approaches.

Such systems could cost less, operate on existing hospital servers without expensive GPUs and keep sensitive information on hospital premises.

That approach also fits the operational constraints hospitals face.

AI budgets remain limited, with software competing for spending against medicines, equipment and staff. Hospital leaders therefore need technology to affect a metric they already monitor—whether revenue, cost, patient throughput or order conversion.

The expectation discussed at the roundtable was that hospitals would want to see returns within 10 to 12 months, partly because the technology itself is changing too quickly for organisations to plan much further ahead.

The question is no longer simply whether an AI product works.

It is whether a hospital can demonstrate, within a reasonable period, that the product changes an operational or clinical outcome that matters.

Non-clinical work may be the gateway

The strongest consensus at the meeting was that founders should begin with non-clinical problems.

Billing, claims and hospital operations offer a more straightforward route because they have measurable inputs and outcomes and do not require the same level of clinical validation.

Participants pointed to insurance deductions and claims processing as examples of under-addressed opportunities. If a hospital could identify likely insurance deductions while paperwork was still being prepared, it could potentially correct problems before a claim was submitted.

Other examples included collecting patient histories through voice or WhatsApp, directing patients to the appropriate sub-specialist and managing follow-up. The common recommendation was to understand the workflow first and then determine where AI fits.

That is a more restrained proposition than the usual pitch for AI in healthcare. It treats artificial intelligence as a tool within a process rather than the process itself.

It also gives startups something valuable before they attempt higher-risk clinical applications: access to hospital systems, relationships with IT and finance teams, and a record of delivering measurable results.

Another recurring issue was the distance between product teams and practising clinicians.

Hospital leaders said they regularly encounter tools that address problems they do not have or reproduce systems they already use. Products that force clinicians into unfamiliar workflows can quickly lose adoption; every additional button, step or delay creates friction.

Participants suggested that product teams should measure not only accuracy but also clicks, latency and continued usage.

The discussion went further than asking whether a clinician had been consulted during product development.

The more consequential question was whether a practising clinician has enough authority inside the company to stop a product or feature that does not make sense clinically.

Founders acknowledged the difficulty of finding clinical mentors with sufficient availability, while hospitals can take years to make purchasing decisions. That creates a structural mismatch with startup timelines and investor expectations.

The roundtable therefore pointed towards a model in which clinicians are involved early enough to shape what gets built, rather than being brought in only to validate a finished product.

The session ended with another revealing exercise: participants were asked how they would deploy $5 million.

The answers focused overwhelmingly on hospital operations—connecting departments, speeding insurance clearance, improving patient turnover, accelerating radiology reporting and developing asset-light models for reaching Tier-3 cities.

The choices reinforce the central finding of the discussion.

The immediate opportunity in healthcare AI may lie less in replacing clinical decision-making than in fixing the systems around it.

That does not make clinical AI unimportant. It suggests that its development depends on solving several less glamorous problems first: making hospital data accessible and usable, generating Indian clinical evidence, integrating software into existing workflows and establishing a credible economic case for deployment.

The unresolved questions are consequently as important as the areas where there is already agreement: what constitutes sufficient evidence for clinical AI in India, how hospitals can safely share data so models are trained on Indian patients, and how a hospital should calculate the return on an AI investment before purchasing it.

Hospitals are not waiting for AI to become more impressive. They are waiting for it to become useful, measurable, compatible with existing systems and trustworthy enough to fit the way care is actually delivered.

By Subramanya Joshi

If you spotted a mistake or have feedback on this story, write to us at editor@newsassociate.com. Email us your press releases on the same ID.