The AI Payback Problem Hidden Inside Pharma’s Most Measurable Workflows

In a business where even a small delay can matter, Novo Nordisk is treating AI less like a magic layer and more like a measurement problem. The company has turned to process intelligence and digital twins to map how drug-development work actually moves across systems, teams, and approvals. That matters because in regulated pharma, a faster model output does not automatically mean a faster path to market.

Fast Facts

• Novo Nordisk is using process mining to study drug-development workflows.
• The company says it has several hundred AI agents in active deployment.
• A process that looked like seven steps sometimes appeared as five or nine steps, depending on who performed it.
• The company’s AI ROI is still too early to confirm because drug-development cycles take years.
• Measurement gaps, hidden costs, and workflow drift can make AI gains hard to prove.

Why the numbers stay slippery

The technical issue is not whether AI can accelerate a task. It is whether an enterprise can see the whole workflow well enough to tell where time, cost, or risk is really changing. Process mining helps by reconstructing work from event logs inside enterprise systems, rather than relying on employee memory or survey answers. In a complex clinical setting, that can expose missing handoffs, duplicate steps, and process drift that quietly erase any local productivity gain.

That is also why baselines matter. If an organization does not record cycle time, exception rates, rework, and approval delays before rolling out AI, it loses the ability to prove what changed afterward. External technical guidance on AI evaluation and monitoring has made the same point in different language: deployed systems need ongoing measurement, because production behavior can differ from test behavior. From a defensive perspective, that is not just an analytics issue. It is a governance and security issue too.

The presence of several hundred AI agents adds another layer. Agentic workflows can increase the number of system calls, permissions, and automated actions that need to be tracked. If those actions are not instrumented well, costs can become difficult to attribute and anomalous behavior can be harder to spot. In a regulated environment, that also raises the stakes for auditability, access control, and change management.

The broader lesson is that AI ROI is often an observability problem before it is a model problem. A company may see task-level wins, but the business case only becomes real when the workflow is standardized, measured over time, and tied to downstream outcomes that matter. Until then, the smartest deployment can still be hard to defend on paper.

Conclusion

What Novo Nordisk is testing is not just AI, but whether an enterprise can make itself measurable enough to trust AI at scale. That is the quiet challenge behind many corporate AI programs: without visibility into the workflow, the value stays partial, delayed, or disputed. In cyber and in business operations alike, what you cannot instrument is hard to secure, hard to govern, and even harder to justify.

WIKICROOK

• Process mining: A method that uses system event logs to reconstruct how work actually flows across an organization.
• Digital twin: A data-driven model that mirrors real operations so teams can observe and test process changes.
• AI agent: An automated software component that can perform tasks, call tools, or make decisions within a workflow.
• Baseline: A pre-change measurement used as the reference point for judging whether a new system improved results.
• Telemetry: Operational data collected from systems to monitor behavior, usage, performance, and anomalies.