Dipp AIDipp AI

Case study · Life Sciences & Pharma

A pharma sponsor stopped an AI-drafted submission error before it reached the regulator

Verification against the source data caught 214 proposed statements that the trial record did not support — before any of them were filed.

Global pharmaceutical sponsor · 9 active submissions · published August 12, 2026

Download the PDFApply for early access

214

unsupported statements refused before filing

across nine submissions

3.7%

of generated assertions failed verification

58%

reduction in medical-writer review hours per section

0

post-filing corrections attributable to drafted text

over the engagement period

Summary

A pharmaceutical sponsor used agents to draft sections of regulatory submissions from trial data. Fluency was never the problem; unsupported assertion was. The Logic Scrubber verified each proposed statement against the source dataset before it could enter the document, and recorded both the check and its result.

Orcher components enforced

Solution: Governance & audit evidence

The situation

What was happening before Orcher.

  • Agents drafted narrative sections from locked trial datasets, and medical writers reviewed the output line by line.
  • Review was the bottleneck and the risk: a plausible sentence that overstated a secondary endpoint reads exactly like a correct one.
  • The sponsor needed to show its inspectors what verification had occurred, not merely that a human had signed.

What Orcher enforced

The control surface, component by component.

  • Every proposed statement was decomposed into checkable assertions and tested against the locked dataset before it could enter the document.
  • Statements the data did not support were refused with the specific contradiction attached, rather than flagged for a reviewer to find.
  • Role Identity Fabric bound each drafted section to the medical writer and the responsible medical monitor.
  • The ledger holds the assertion, the check, the dataset version and the outcome — the artefact an inspector asks for.

The data

Measured over the engagement window.

Submission sectionAssertions generatedRefusedRefusal reason (most common)
Efficacy narrative3,18097Effect size overstated vs dataset
Safety narrative2,41064Adverse event count mismatch
Population description1,12018Inclusion criteria drift
Statistical summary1,86031Rounded value not reproducible
Discussion9704Unsourced causal claim

Assertion-level verification results across nine active submissions.

Outcomes

What the organisation did next.

  • Medical writers moved from line-by-line checking to reviewing refusals and edge cases.
  • The sponsor's inspection-readiness pack now includes verification evidence generated automatically.
  • Two agent-drafted section types that had been prohibited internally were approved for use.
  • The refusal log became a quality signal in its own right, surfacing two data-entry problems upstream in the trial record.
“The model was never wrong in a way that looked wrong. That is precisely why the check has to be mechanical.”
— Head of Regulatory Operations, global pharmaceutical sponsor (partner declined attribution by name)

Read further

The research behind this engagement.

Disclosure. Design-partner engagement, anonymised at the partner's request. Figures are measured by Orcher's own observability and ledger instrumentation over the stated period and have not been independently audited. Market figures carry their own source line.

More engagements

Other organisations that had to prove the same thing.

Run this pattern on your own workflow

Design partners deploy Orcher on one directive class, instrument it the way this study was instrumented, and keep the evidence whatever they decide afterwards.

Apply for early accessAll case studies