Gen AI is fundamental shift

This isn't just "easier AI".
It fundamentally changes what AI can do and who can do AI.

Validating ICFs is a $500K-$2M annual cost

Let's take a pharma firm that runs:

• 50-100 clinical trials at a time
• 20-50 sites per trial
• each submitting informed consent forms (ICFs) in local languages
• with senior clinical ops verifying each against the 68-page ICH E6 guideline
• 2-4 hours per review
• $500K-2M annual cost in personnel time alone

Same with compliance docs - every country, every site, every amendment

What if LLMs could do this?

Let's upload the ICH E6 guideline into ChatGPT
and ask it to validate the following five informed consent forms:
ICF 1 | ICF 2 | ICF 3 | ICF 4 | ICF 5

The prompt is simple:

Validate the following ICF against the guidelines at https://database.ich.org/sites/default/files/E6_R2_Addendum.pdf and list all non-compliances.

After 2 min 41 seconds of thinking, it found 15 errors.

Below are the places your ICF does not meet ICH E6(R2) Addendum expectations for informed consent content (primarily 4.8.10, plus a couple of related clauses).

• 4.8.10(a) – “Trial involves research” not stated clearly. Your ICF describes a study, but never explicitly tells the subject this is research.
• 4.8.10(e) – Subject responsibilities not described. Beyond contraception and “training,” it doesn’t clearly list what the participant must do (e.g., dosing compliance, visit attendance expectations, AE reporting expectations, glucose diary requirements, restrictions).
• 4.8.10(f) – Experimental aspects not explained. It doesn’t clearly state what is investigational/experimental.
• ... and a dozen more.

It's like a genie. Free wishes (but careful)

But what if it makes a mistake?

Well, what do we do when our teams make mistakes?

We have protocols for that.

• Checklists
• SOPs
• Peer reviews
• Audits

These are designed to catch human errors. They catch AI errors too!

Use AI to generate checklists. Checklists are robust

Use AI to run peer reviews. Validation is harmless

It's stunningly cheap, and growing cheaper by the day

Extract Schedule of Assessments from a protocol PDF

Say you have a medical trial protocol for testing a new diabetes medication called "Glucofix" vs placebo.

Extract the "Schedule of Assessments" tableinto a tabular format suitable for direct upload using the REDCap Instrument-Event Mapping CSV format.

A single missed assessment costs $50K-200K in rework/lost data.
LLMs for validation reduces errors dramatically.
Error rate: ~5% manual → ~0.5% with LLM + human QC. Cost: $50

Find all errors and correct them. List your corrections.

Find anomalies in lab data results

You have 50 pages of lab listings data for a diabetes clinical trial.
200 patients × 8 visits × 8 lab tests = ~12K data points

• Edit checks, e.g. "ALT > 500 = flag", works only for extreme values.
• Listings review, on pages of numbers, humans miss subtle patterns
• Statistical outlier detection: "2 SD", misses systematic bias

Analyze this clinical trial lab data. Find data quality issues. Look for: temporal patterns, site-specific anomalies, impossible correlations, gradual drift, etc.

10,000 patients in 5 minutes. Daily surveillance. Flag within 24 hours.

Filter patient list for eligible candidates

A day's delay in clinical trial enrollment from patient data costs $40K-$8M.
Not enough matching patients, high screen failure rate (40% average), ...

"Let me check if Mrs. Johnson's HbA1c is 7.5-10.5%... and her creatinine... and her medication history..." takes 30-45 minutes per patient chart. With errors.

Find patterns predict enrollment failures [TIME CHECK]

Based on enrollment history:

Analyze which inclusion/exclusion criteria predict enrollment failure. Quantify impact of each criterion

Instead of a therapeutic lead saying "I think we should use 7.5-9.5% HbA1c range" and clinical ops saying "That seems reasonable", LLM says "Historical data shows HbA1c ranges <2.0% have 23% higher screen failure rates"

Data-driven decision lead to better site selection.

AI is already better than humans in many areas