Shadow AI on Your Panel: A Director's Policy

A panel expert pastes a claimant’s GP records into ChatGPT to speed up a chronology. The output looks cleaner and the expert files on time. However, six months on, an opposing barrister asks how the medical history was prepared, which tool touched it, and where the patient’s data went. The expert answers for their own practice. You answer for your agency.

This is shadow AI. Experts on your panel using AI tools you have not sanctioned, cannot see, and do not control. The work runs faster but the exposure lands on you.

You need a position on AI before an instructing firm asks for one. This guide sets out what the evidence shows, where the rules are heading, and the policy you put in place this quarter. For the wider agency picture, read our guide for medico-legal agencies and our full guide to what a medico-legal report is.

The numbers a director needs

Expert AI use is climbing. The Bond Solon Expert Witness Survey 2025 found 20% of experts had used AI in their role, double the 9.31% the year before. The same survey found 89% want specific guidance on AI use, and 86% would refuse an instruction to work from an AI-drafted report.

Observe those three figures together. Most reach for guidance no one has given them and this gap is where agencies sit.

What the courts and regulators have said

The direction of travel is clear, and recent.

In June 2025, Dame Victoria Sharp, President of the King’s Bench Division, pointed to cases where fake authorities reached the court through careless AI use. Judicial guidance on AI followed in October 2025. On 30 January 2026, the Academy of Experts published guidance for expert witnesses on AI use.

The rule change to watch sits with the Civil Justice Council. Its consultation proposed amending the Practice Direction 35 declaration so an expert states what AI use was made, beyond transcription and admin, and names the tools used. The consultation closed on 14 April 2026, with a final report to come.

The settled position across these sources reads the same. AI helps organise records and pull material from long documents, the opinion stays with the expert and any AI use must be documented, verifiable, and disclosed where it shaped evidence.

Where the exposure lands on the agency

Directors carry three risks their experts often do not weigh.

Confidentiality - A claimant’s records in a personal AI account sit outside your data protection framework. If those records were disclosed or retained, the instructing firm looks to the agency who placed the expert.

Defensibility - An AI-assisted fact with no source trail will not stand up under a Part 35 question. Missing it will have the report challenged, the case stalled and your standing with the instructing firm drops.

Reputation - One expert’s shortcut becomes the whole agency’s problem. Instructing solicitors choose agencies on trust. A single AI misstep on a live case spends the trust fast.

The policy to put in place

A workable AI policy fits on a page and answers what an instructing firm would ask. Build yours around six points.

• Permitted tools. Name the tools experts may use. Ban patient data in consumer AI accounts outright.
• Data handling. Records stay inside tools with the right data protection and clinical safety standing. No pasting into personal accounts.
• Human ownership. The expert forms every opinion. AI prepares facts, never conclusions.
• Source trail. Every AI-assisted fact links back to the record it came from, checkable at source.
• Disclosure. Record what AI touched each report, ready for the PD35 declaration once it changes.
• Review and sign-off. The expert reviews and signs every output. No unread AI text reaches a report.

Give experts the policy in writing at onboarding, and again when you send an instruction. Most experts want the guidance and you must supply it.

Turn the policy into an answer you give

A policy on paper is a start. An instructing firm wants to see it work on a live file.

Health Narrator gives your panel one sanctioned route. It takes the full bundle, typed, handwritten, scanned, or photographed, and produces a structured, source-referenced chronology shaped to the expert’s specialty. Every finding links to the source line, so any fact is checkable in one click, with a full audit trail. It holds NHS DSPT registration and DCB0129 clinical safety certification. It uses no patient data for model training. It prepares the facts and leaves the opinion to the expert.

The answer covers the confidentiality question, the defensibility question, and the disclosure one with a singular tool, the same for every expert on your panel. When an instructing firm asks how your panel handles AI, you have a straight answer.

See what Health Narrator does for your panel. Book a demo.