A 95% Accurate AI Model Can Still Be Fragile in Regulatory Review — Here's Why

95% Accuracy Isn't Enough to Defend a Pharma AI Model

A pharma AI model can post 95% accuracy and still be fragile under regulatory review.

The cause is rarely the metric. It's usually a pre-specified document that nobody on the AI team thought to write.

ICH M15: What Changed on January 29, 2026

ICH harmonizes pharma guidelines across the FDA, EMA, PMDA, and the other member authorities.

M15, their first global guideline on Model-Informed Drug Development, reached Step 4 on January 29, 2026. Agencies are now entering the implementation phase.

AI and machine learning are explicitly listed as modeling methods, on the same footing as:

• popPK
• PBPK
• QSP
• disease progression models

The Real Question Isn't Whether AI Is Accepted

The real question isn't "is AI accepted?"

The real question is "is the evidence the model produces pre-specified, justified, and traceable?"

That's where the Model Analysis Plan comes in.

What a MAP Must Contain, at Operational Minimum

The MAP frames each modeled analysis before it runs. At operational minimum:

• Question of interest — what decision is the model meant to inform?
• Context of use — where, when, and by whom the model will be used
• Data and its qualification — sources, quality, representativeness
• Chosen method and justification — why this architecture, not another
• Technical criteria — performance metrics, acceptability thresholds
• Planned evaluation — how you'll judge the model is fit for purpose
• Link to model risk — effort proportional to the stakes

Pre-dated. Versioned. Traceable.

The Final Metric Doesn't Carry the Evidence Alone

Direct consequence: an AI model submitted with impressive accuracy but no coherent MAP stays fragile under regulatory review.

The final metric doesn't carry the evidence on its own. The pre-specified plan is what makes the evidence defensible.

The question to ask today in pharma or biotech: are you documenting your AI models with a proper MAP, or with an end-of-project README?

---

If you're running an AI project in pharma or biotech and want to frame it so it holds up under regulatory review, let's take 30 minutes to talk.