Point of No Return. AI Infrastructure as Production Systems.

The pattern we're seeing is teams running ~6 AI pilots that work in testing, then six months later they're still pilots.

The way to quickly diagnose?

• Ask your team which AI pilots are actually on the critical path for a decision you need to make in the next 6 months.

If the answer is "none" or "still evaluating," you've reached a paralysis. Models may work but they're not integrated into workflows, no one owns their performance because 'who should?', and finance sees mounting costs with no promising ROI.

What Successful Teams Did

The companies we saw grow by leaps and bounds made at least these three changes:

1. Picked two workflows and went deep

They chose cases where:

• Data was clean

• Regulatory risk was clear

• There was a specific decision the model would inform

Usually these were low-risk internal tools first, such as target screening or site selection, before anything patient-facing.

2. Named system owners instead of project managers

This is someone accountable for ongoing model performance who can make the call to take it offline if it's not working. A decision-maker who doesn't hesitate; someone who interfaces with the team actually using it.

If you don't have someone with both technical depth and workflow understanding, you might not be ready to deploy.

3. They had tiered validation

The breakthrough was that you don't validate a discovery triage tool (low-risk, internal use) the same way you validate a manufacturing QC model, which is high-risk and touches regulatory filings.

They look something like:

• Low-risk (internal): Basic benchmarks, version control

• Medium-risk (decision support): External validation, drift monitoring

• High-risk (patient/regulatory): Full qualification, SOPs, audit trails

Most teams apply the same validation burden to everything, which kills momentum.

Where the ROI is Likely to Show Up

It won't likely be one big explosion, it might be separate smaller gains that compound when the AI infrastructure is reusable:

• Discovery: 2-3x faster screening cycles
  (But only when validation can keep up—that often becomes the new bottleneck)

• Clinical ops: 10-20% fewer screen failures, faster site activation
  (But only if ops changes their workflow to use it, not runs it in parallel with existing processes)

• Manufacturing: 10-15% fewer deviations, faster review cycles
  (But only if QA trusts it enough to change their review cadence)

Gains only showing up when you change the workflow to incorporate the AI. Running AI alongside existing processes just adds work, and so teams abandon it within months. AI has to do the work, and we have to trust it.

Example: 18 Months of Pilot to Production

A Phase 2 oncology company had ran pilots across target screening, patient stratification, site selection, deviation prediction, and more. All areas were promising but not in production.

When they picked two workflows (target screening and site selection), named system owners, defined what success meant (not "model accuracy" but "does this change how we decide?"), and documented decision logic from day one, things changed.

In ~three months, both were live and more importantly, when they started their third workflow, they reused some 60% of the data contracts and governance framework they already had. That third tool took only six weeks instead of sixish months.

Their Head of Data Science said they, "stopped debating whether AI works and started focusing on making it work."

When is it premature to implement AI infrastructure?

The honest answer is that most pre-IND companies should not build AI infrastructure yet. There just isn't enough programs or data to justify reusable systems.

But you're probably ready if:

• IND-stage or later (multiple programs to create reusable infrastructure)

• Data already structured (or resources to structure it)

• Someone internal can own model performance technically

Probably not ready if:

• Pre-IND with single asset

• Data is mostly unstructured (PDFs, inconsistent spreadsheets)

• No one who understands both the science and the models

• Core R&D workflows are still in flux (don't automate unstable processes)

Does this match your experience with AI pilots? What are the constraints?

Warm regards

—Roop