The AI Gold Rush Hits a Hard Wall: Proving What Actually Changed

When a process spans research, manufacturing, documentation, and regulatory handoffs, a productivity win in one team can vanish in the next. That is the tension at the heart of Novo Nordisk’s AI work: the company can see improvements in specific steps, yet still struggle to show a clean return on investment across the full drug-development chain.

The company’s use of process mining and digital-twin style visibility is revealing why. A workflow that appears to have seven steps can look like five steps to one employee and nine to another. In practice, that means the organization is not only measuring AI performance - it is measuring whether the process itself is stable enough to measure at all.

Fast Facts

• AI benefits are easier to show at the task level than in long-cycle financial results.
• Novo Nordisk is using process data to compare documented work with how work is actually done.
• A one-week delay on a blockbuster drug can carry very large economic consequences.
• Baselines matter: without them, productivity claims are hard to verify.
• Hidden AI costs include integration work, usage growth, and ongoing maintenance.

Why the math stays messy

This is a measurement problem, not a headline problem. In regulated workflows, the value of AI may be spread across many small changes: fewer handoffs, less rework, better document readiness, or faster approvals inside a single team. But financial ROI only appears later, if it appears at all. NIST’s guidance on AI evaluation reflects that reality: trustworthy AI depends on disciplined measurement, not intuition.

Process mining helps because it pulls event data from enterprise systems instead of asking people to recall what happened. That matters when the real bottlenecks are invisible to management or scattered across departments. A digital twin of operations can expose drift between the official process and the one employees actually follow, which is often where AI projects succeed or stall.

The cost side is just as slippery. Subscription fees are easy to count, but enterprise AI also brings integration work, data movement, model calls, retraining, and the upkeep of new workflows. In more automated environments, usage can grow faster than teams expect, which is why granular telemetry and usage tracking are becoming part of the finance conversation, not just the engineering one.

That is the broader lesson for any organization chasing AI returns: task-level wins are real, but they do not automatically become company-wide value. If the baseline is weak, the process is fragmented, or the time horizon is long, the ROI story will stay cloudy even when the tool is genuinely useful.

Conclusion

AI does not fail simply because it is slow or expensive. Often, it fails to prove itself because the organization cannot see the full system clearly enough to attribute the gain. In industries built on long timelines and interdependent work, the real advantage may belong to the companies that instrument their processes first and ask for ROI only after they can actually measure change.

WIKICROOK

• Process mining: Analysis of event data from enterprise systems to reconstruct how work really flows.
• Digital twin: A data-driven model of operations used to compare real performance with expected behavior.
• Baseline: The starting measurement used to judge whether a later change improved performance.
• Telemetry: Automatic collection of usage and performance data for monitoring and analysis.
• ROI: Return on investment, a measure of whether gains outweigh the total cost of a project.