New Delhi, Nov. 10 -- As foundation models become increasingly capable and interchangeable, pharmaceutical companies must shift their focus to proprietary knowledge, model-agnostic architecture and workflow assurance. The challenge is to make AI outputs reliable, traceable and auditable in an industry where errors can have regulatory and patient-safety consequences, explains Tarun Mathur, Chief Technology Officer at Indegene, a technology-led life sciences services company that works with pharmaceutical and biotechnology firms across clinical development, medical, regulatory and commercial operations. Edited excerpts.

You have advocated an AI-led platform approach over point solutions in pharma. How has that thinking evolved as AI models advance, and where should enterprises focus their investments now? The platform thesis has won, and the argument has moved on. Point solutions created fragmented data, controls and validation processes, with value leaking out in the gaps between them. What has changed is what sits at the centre of the platform. Models are converging, and orchestration is becoming a commodity. What does not commoditise is domain knowledge: why a claim was approved, how a market-specific exception is handled or how an experienced reviewer interprets a safety narrative. Much of this knowledge resides in expert judgement rather than documentation. The winners will own this knowledge and be able to demonstrate the reliability of every AI output.

With cheaper, open-weight models becoming increasingly capable, does model-agnostic architecture become essential for regulated pharma workflows? Absolutely. Locking a regulated workflow to one model is a bet against the pace of the market. We evaluate models workflow by workflow, first on their ability to meet accuracy and traceability requirements, then on economics and finally on data privacy and intellectual property constraints. A safety-signal workflow and a first-draft summary should not use the same model simply for convenience. Our principle is simple: rent the intelligence, own the context. Models are replaceable components; domain knowledge and controls are the real assets.

What does trust engineering look like in practice for medical writing, MLR review and regulatory operations? Trust in regulated AI must be engineered and provable. First, AI must work from approved, governed sources rather than relying on what a model remembers. Then, every output must point to its underlying evidence so reviewers can verify it. Finally, the knowledge itself must be continuously evaluated, not just the answers generated from it. Accountability remains with people. AI can draft, retrieve and check, but medical writers own the scientific narrative, MLR reviewers sign off on claims, and designated individuals approve regulatory submissions. AI should make experts more productive, not replace their judgement.

Does running models within an organisation's own environment necessarily make sovereign AI safer for regulated use? Control is not the same as safety. Running models within an organisation's environment also means taking responsibility for validation, security patching, drift monitoring and red teaming. This can significantly increase the governance burden. There is a case for sovereign models in highly confidential applications and organisations with deep AI engineering capabilities. For others, sovereignty over data and domain knowledge matters more than ownership of model weights. Our approach keeps client knowledge in a private instance under their control, while allowing the underlying model to be replaced.

How should pharma companies monitor model drift and maintain an auditable evidence trail as AI systems evolve? Companies must assume that models will change, whether through their own decisions or provider updates. First, they need fixed workflow benchmarks covering accuracy, cost, latency and reliability. Then, every output must be traceable to its model version, prompt, source data and approval. Finally, model changes must be treated like modifications to validated systems, with testing before deployment. Pharma already understands validated systems. The challenge is to apply that discipline to AI rather than treating it as a permanent experiment.

Could workflow assurance become as important as model governance in enterprise AI? It will matter even more. Regulators and pharma leaders ultimately care about whether a workflow produces a correct, defensible outcome, not which model was used. Before AI influences regulated work, organisations need a defined intended use, validation against agreed test sets, role-based access, source traceability and complete audit logs. They must also establish human approval checkpoints, exception handling, rollback mechanisms and a named owner accountable for the outcome. Building these controls in from the beginning is far more effective than adding them after deployment.

What should CIOs and CTOs measure to establish whether AI is delivering business value in regulated environments? Documents generated, hours saved and tokens consumed measure activity, not value. The real question is whether better work gets approved faster. Companies should first measure quality through first-pass approval rates and review cycles per asset. They should then track end-to-end cycle times and, finally, business outcomes such as faster time to market and lower costs per approved asset. These baselines must be established before AI deployment. Usage, spending and latency remain important operational metrics, but they should not become the board's primary scorecard. No productivity gain justifies a high-severity safety incident.

What will distinguish an AI-native pharma technology organisation over the next three years, and how will Indegene's Cortex evolve? An AI-native organisation will redesign work rather than simply accelerate existing processes. AI will produce the first draft, experts will govern decisions, and every correction will improve the system. At Indegene, Cortex, our Domain Intelligence platform, is central to this shift. It captures expert knowledge, including judgement that may never have been formally documented, and makes it available to AI agents and applications. Every answer is linked to its source, and the knowledge is tested and improved through expert corrections. Our ambition is to make Cortex the governed knowledge layer powering our platforms and, eventually, clients' AI systems. Models will keep changing; validated expertise is what compounds.

Published by HT Digital Content Services with permission from MINT CIO CIRCLE.