The next phase of enterprise AI will not be won by simply adopting the most powerful model. As agents begin to take on real work, the differentiator is shifting to the system around the model: the knowledge it can access, the expertise embedded into its workflows, the controls governing its decisions, and the feedback that allows it to improve.
At the Indegene Digital Summit 2026, Tarun Mathur, CTO (Indegene) explored this shift and what it means for life sciences organizations moving from AI-assisted information retrieval toward agentic systems capable of performing increasingly complex work. The discussion pointed to a fundamental change in how leaders should think about AI investment: the model is becoming one component of a much larger enterprise capability.
The Model Is No Longer The Whole System
AI models are improving rapidly. The session highlighted benchmarks in which state-of-the-art models performing specific work-related tasks moved from a 48% human-evaluated score to 85% within a year. But the more consequential insight was what sits behind that improvement.
As AI moves from answering questions to performing work, model capability alone becomes an incomplete measure of enterprise readiness. Performance increasingly depends on the surrounding architecture, or “harness”: the prompts, tools, semantics, ontologies, methods, checks, expert gates, and knowledge that shape how an agent approaches a task.
For pharma leaders, this changes the investment question from “Which model should we use?” to “What system do we need to build around the model to perform this work reliably?”
More Information Does Not Automatically Create More Intelligence
Building on this idea, the discussion challenged a common assumption: giving an AI system access to more information does not necessarily make it better.
In one healthcare benchmark discussed during the session, providing an agent with web access actually reduced performance. Yet when the same underlying model was surrounded by domain-specific context, including the semantics and ontology of the work it was performing, performance nearly doubled for certain domain-intensive tasks. The effect was not universal, reinforcing another important point: sophisticated agent architectures create the greatest value where the work itself requires deep domain understanding.
For pharma, where critical workflows are shaped by specialized terminology, SOPs, protocols, scientific evidence, regulatory requirements, and defined processes, domain context cannot simply be retrieved. It needs to be engineered into how the agent works.
Enterprise Knowledge Must Become Agent-Ready
That requirement extends directly to the enterprise knowledge layer. Much of pharma's institutional knowledge still resides in unstructured formats: protocols, literature, SOPs, documents, tables, code, and repositories designed primarily for human consumption.
The session highlighted research in which an agent answering questions directly from a research paper achieved a 56.7% quality score. When the same information was transformed into structured datasets and tools that allowed the agent to compute against tables and execute relevant code, performance increased to 82.7%. The underlying information had not changed. Its usability to the agent had.
This creates an important priority for life sciences organizations. Preparing knowledge for AI is no longer simply a content-management exercise. Metadata and structured attributes, ontologies, and machine-usable tools increasingly determine how effectively agents can reason over enterprise knowledge and translate it into work.
AI Changes Where Human Expertise Enters The Workflow
Extending the discussion from knowledge to process, perhaps the more important productivity opportunity is not removing experts from workflows but changing their starting point.
The session illustrated this through the creation of an operational checklist from a clinical trial protocol. Thirteen clinicians took approximately 150 minutes to complete the task manually. When AI produced the first draft and clinicians reviewed and refined it, total time fell by roughly 40%. Yet the deeper shift was structural: clinicians no longer began with a blank spreadsheet.
They entered the process as reviewers and editors rather than initial synthesizers. That distinction matters because it reframes AI transformation from task automation to workflow redesign. Instead of asking which human activities can be eliminated, leaders can ask where AI can move expert intervention downstream, allowing scarce expertise to focus on judgment, validation, and exceptions.
Feedback Loops Become Part of The Operating Model
Once agents begin performing meaningful work, another question becomes unavoidable: what does good look like?
There is no universal answer. A high-quality output may require accurate citations in one workflow, comparative evidence in another, or adherence to a specific format or methodology elsewhere. Those criteria must therefore be defined with the people who understand the work.
The emerging opportunity is to turn those definitions into continuous feedback loops. Rather than relying only on techniques that modify the underlying AI model, newer approaches can improve the prompts, context, tools, and other components surrounding it. The session cited one example in which a feedback-driven approach improved a lightweight model's score from 29% to 78%, while another stronger model moved from 84% to 94%.
For pharma leaders, this makes evaluation and feedback architecture a core operating capability, not simply a technical quality-control step. Expert judgment must be translated into measurable rubrics that allow agentic systems to learn how the organization defines quality.
Enterprise AI IP Is Moving Into The Harness
This shift also changes where organizations should think about proprietary advantages.
Frontier models are increasingly accessible to every enterprise. Organizations can adopt the same underlying models and, over time, replace one with another. What is far harder to replicate is the accumulated organizational intelligence surrounding those models: proprietary prompts, workflows, tools, ontologies, evaluation criteria, knowledge structures, and feedback.
That makes the agent's harness strategically important. If the underlying model can be changed while the organization's optimized context and workflow remain intact, the enterprise becomes less dependent on any single model provider while retaining the intelligence it has built around its own way of working.
The implication is significant: competitive advantage in enterprise AI may increasingly come from how effectively an organization codifies and continuously improves its expertise, rather than which foundation model it selects.
From Deploying AI to Building Learning Systems
Taken together, these shifts suggest a different playbook for agentic AI. Organizations first need to embed expertise through semantic structures, ontologies, structured knowledge, methods, controls, and expert gates. They then need to define what good looks like, capture feedback from actual performance, and use that feedback to continuously refine the system surrounding the model.
This is a departure from treating AI deployment as a one-time technology implementation. The more powerful model is a learning system, one in which enterprise knowledge shapes the agent; human expertise defines quality, and real-world feedback continually improves how work gets done.
For pharma, that may ultimately be the defining shift of the agentic era: AI transformation will depend less on access to intelligence and more on the organization's ability to surround that intelligence with its own expertise, controls, and capacity to learn.