Beyond the Hype: A Conversation on Building Trust-First AI in Healthcare

1. Your career spans building one of India's early clinical operating systems to leading an AI-focused healthcare technology company. What experiences along this journey most influenced your philosophy of building “trust-first” AI? 

Building and commercializing the Clinical OS about 15 years ago was an eye-opening experience. India’s scale, skewed clinician-to-patient ratio, income constraints, and patient behaviour make it a natural candidate for digital-first-and now AI-first-healthcare, yet it remains one of the lowest adopters of healthcare technology. There is a distinct lack of trust in software systems, particularly among clinicians. Healthcare has an extremely low margin for error, while generative AI is inherently stochastic and non-deterministic. Trust-first AI means harnessing generative AI without compromising reliability, with the right safeguards and human oversight. 

2. Healthcare organisations often invest heavily in digital transformation but continue to struggle with operational inefficiencies. In your experience, what are the biggest hidden operational blind spots that quietly impact hospital profitability and performance?

If it takes an extra two or three weeks to onboard a physician, that’s two or three weeks where that provider isn’t seeing patients or generating revenue. The same thing happens with claims denials. A claim may eventually get resubmitted and paid, so the problem looks resolved, but if nobody fixes the underlying reason it was denied, the organization keeps repeating the same expensive process. Prior authorization is another good example. Individually, these issues may seem small. But across thousands of transactions, providers, and patients, they add up to a significant financial and operational impact. 

3. Revenue cycle management, provider credentialing, and administrative workflows are often viewed as back-office functions. Why do you believe these areas present some of the greatest opportunities for AI-driven transformation?

Because they have many of the characteristics that make them ideal for AI: high volume, repetitive work, rules-based processes, and a lot of manual review. Credentialing is a good example. You may have people manually checking hundreds of documents for completeness, expiration dates, inconsistencies, or missing information. AI can do that first-level review very effectively and allow the team to focus on the exceptions that actually require judgment. You can introduce AI here without immediately asking it to make a clinical decision. That makes it easier for organizations to experiment, measure the results, build trust, and improve the technology. 

4. Many organisations begin their AI journey expecting immediate automation. From your experience working with healthcare providers, where do you think AI creates the fastest measurable business value, and where are expectations often unrealistic?

The fastest wins tend to come from processes that are repetitive, structured, and easy to measure. Claims review, eligibility verification, credentialing document checks, prior authorization support, those are areas where you can fairly quickly measure whether turnaround time has improved, whether errors have gone down, or whether staff are processing more work. Where expectations become unrealistic is when organizations jump from that to saying, “Can AI now replace an experienced person making complex decisions?” I see AI as very strong at doing the repetitive 70 or 80 percent and helping humans focus on the difficult 20 percent. 

5. One recurring theme in healthcare AI is the importance of keeping humans in the loop. Why has this model consistently proven more successful than pursuing complete automation, particularly in clinical and operational environments? 

Healthcare has too many exceptions for us to assume that every situation will fit a predictable pattern. You can build a system that performs extremely well on the majority of cases, but eventually you’re going to encounter something unusual-and that’s where human judgment becomes important. There’s also a trust component. People are much more willing to adopt AI when they see it as supporting their work rather than taking control away from them. So human-in-the-loop isn’t simply a safety mechanism. It actually helps adoption. The AI handles the repetitive work, the person handles the exceptions. 

6. Trust remains one of the biggest barriers to AI adoption in healthcare. Beyond accuracy, what factors determine whether clinicians, administrators, and operational teams are willing to embrace AI-powered decision support? 

Accuracy is obviously important, but I don’t think accuracy alone creates trust. People want to understand why the system is recommending something. If an AI tool tells a billing specialist, “This claim is likely to be denied,” the immediate question is going to be, “Why?” If it can point to the missing documentation or the coding issue, that becomes useful. Consistency also matters, and so does control. Users need to know that they can disagree with the system and override it when necessary. Ultimately, the best AI tools give the person enough information to make a better decision. 

7. AI systems are only as effective as the data they rely on. What common data quality or integration challenges do healthcare organisations underestimate before launching AI initiatives? 

The biggest surprise for many organizations is discovering how fragmented their data actually is. Information that looks standardized at a high level can be very inconsistent once you get into the details. The same provider may be represented differently across multiple systems. Fields may have different names, historical records may follow different formats, and sometimes two systems disagree on something as fundamental as whether a provider is active. Organizations naturally get excited about building the AI model, but cleaning the data, connecting systems, and defining a reliable source of truth is often where most of the effort goes. 

8. As healthcare organisations evaluate AI vendors, what questions should technology leaders ask to distinguish genuinely impactful AI solutions from products driven primarily by market hype?

I would ask the vendor to get very specific about the workflow they improve. Statements like “we reduce administrative burden” sound good, but they’re difficult to evaluate. I’d rather hear, “We reduced prior authorization turnaround from four days to one,” because now I have something measurable to validate. I would also ask what happens when the system is wrong. How are exceptions handled? Can a human override it? Is there an audit trail? I’d want to speak with customers who have been using the product for at least a year.

9. Many healthcare leaders worry that AI will replace jobs. You have argued that AI is more likely to expand the workforce than eliminate it. Could you explain how you see AI reshaping healthcare roles over the next five to ten years? 

I think we’re going to see jobs change much more than we see entire categories of jobs disappear. There will probably be less demand for work that is purely manual-data entry, document checking, repetitive verification. But that doesn’t mean the people doing those jobs suddenly become unnecessary. What changes is where they spend their time. Instead of reviewing every credentialing document, someone may focus on complicated exceptions, provider communication, or quality control. We’ll also see new roles emerge around monitoring AI systems, auditing their decisions, and managing workflows. 

10. Successful AI implementation requires more than technology. What organisational, cultural, or leadership changes are essential for healthcare institutions to achieve sustainable adoption and long-term value? 

One of the biggest mistakes is treating AI implementation purely as an IT project. You can implement technically excellent software and still have nobody use it. Leadership has to think about adoption from the beginning. Who owns the outcome after implementation? How are employees being trained? How is feedback collected? What happens when somebody identifies an error? You need a culture where employees are encouraged to challenge the AI. If people are afraid that reporting a problem will make them look resistant to innovation, they’ll stop reporting problems-and that’s when small issues become expensive ones. 

11. With increasing regulatory scrutiny around AI, privacy, and governance, how can healthcare organisations build AI systems that are both innovative and compliant while maintaining patient trust? 

Governance has to be part of the design process rather than something you add right before launch. From the beginning, organizations should know what information the system is accessing, why it needs that information, how outputs are being recorded, and who is accountable when the AI influences a decision. That also means involving compliance, privacy, security, operational, and clinical stakeholders earlier than organizations traditionally might. I don’t think governance and innovation have to work against each other. Good governance actually makes it easier to scale AI because everybody understands the boundaries. 

12. Looking across your work with hospitals and healthcare enterprises, can you share an example where AI significantly improved operational efficiency or financial performance, and what lessons other organisations can learn from that experience? 

One example was a credentialing workflow where a healthcare organization had a significant backlog. Providers were taking longer to onboard, which meant some physicians were ready to work but couldn’t begin seeing patients as quickly as the organization wanted. A lot of the credentialing team’s time wasn’t being spent on difficult credentialing decisions. It was being spent identifying missing documents, inconsistent information, and incomplete applications. We used AI to catch more of those issues before the application reached the reviewer. That meant the credentialing specialists could spend their time on genuine exceptions rather than chasing paperwork. 

13. As generative AI and autonomous agents continue to evolve, which emerging capabilities do you believe will have the greatest impact on hospital operations over the next few years, and which trends do you think are being overhyped? 

I’m particularly interested in agentic workflows for administrative processes. Today, a lot of automation handles one step: read a document, classify something, or generate a response. The next step is having systems that can coordinate several actions-identify a denied claim, understand why it was denied, retrieve the required information, prepare the follow-up, and route it to the right person. That could remove a tremendous amount of administrative friction. Where I think we need to be more cautious is the idea of fully autonomous clinical agents making consequential decisions with minimal oversight. 

14. If you could offer one piece of advice to healthcare executives planning their AI strategy today, what guiding principle would help them move beyond experimentation and build AI initiatives that deliver lasting clinical and operational value? 

Start with the problem, not with AI. I see organizations asking, “Where can we use generative AI?” or “What should our AI strategy be?” I’d turn that around and ask, “Where are we losing the most time? Where are people frustrated? Where are delays affecting revenue or patient experience?” Once you identify that, then ask whether AI is the right tool. And make the outcome measurable. If you can say, “Credentialing currently takes 45 days and we want to get it to 30,” you have a real initiative. You have a baseline, an owner, and something you can evaluate.