A modern injury or clinical negligence bundle can run to several thousand pages: GP records, hospital notes, imaging reports, prescriptions, correspondence, much of it scanned, some of it handwritten, often with the same letter appearing three times. Reading it is the slow, unglamorous part of an expert's work, and it is exactly the part where AI now makes a real difference. The question worth asking is not whether to use it, but where it helps and where it has to stop.
Why is medical record review the bottleneck?
The opinion is the skilled work, but the reading is what consumes the hours. Before an expert can form a view, someone has to go through every page, find the relevant entries, put them in order and notice the gaps. On a large case that is days of work, and it is work that does not scale by trying harder. It is the main reason reports take as long as they do and cost what they cost.
What is AI genuinely good at here?
AI is well suited to the reading load specifically. It can ingest a large PDF, pull out typed facts such as diagnoses, medications, procedures and admission dates, and build a first-pass chronology in minutes rather than days. It can answer a direct question across the whole record, for example whether a drug was ever prescribed or when a symptom was first recorded, and point to where the answer sits. None of that is interpretation. It is fast, thorough reading, which frees the expert to spend their time on the part only they can do.
Where should AI stop and the expert take over?
The moment a question becomes one of clinical judgement, the machine should hand back. Whether a standard of care was breached, whether an injury was caused by the event in question, how to weigh two conflicting accounts in the notes: these are matters of expertise, and they are what the expert is instructed and paid to decide. AI that drafts an opinion is solving the wrong problem. AI that hands the expert an organised, sourced record so they can form the opinion faster is solving the right one.
Why does every AI answer need a source page?
This is the single requirement that separates a usable tool from a liability. In medico-legal work an assertion is only as good as the document behind it, and an answer that cannot be traced back to a page is an answer no one can stand behind under cross-examination. Every extracted fact and every response should carry the document and page it came from, so the expert can open it and confirm it in seconds.
An AI answer without a citation is a prompt to go and check, not a finding. With the source page attached, it becomes something the expert can verify and rely on.
What makes medical AI dangerous, and how is it managed?
The failure modes are specific. Scanned records with no text layer, handwriting, duplicated documents, missing pages, and the well-known tendency of a model to answer confidently when it should say it does not know. The way to manage them is in the engineering rather than the optimism: read each page in tiers, starting with the text layer, falling back to the page image, then to a vision model where needed, and verify a fact against its source before presenting it. A system built to cite is also a system built to admit when the record does not support an answer.
Where does the patient data go?
Medical records are special category data, and the governance question comes before the clever features. An AI tool fit for this work keeps patient identifiers encrypted, runs on infrastructure the organisation controls rather than posting bundles to an external service, restricts access to invited and authenticated users, and records every step for audit. If you cannot say where a patient's records went when a question was asked of them, the tool is not suitable, however good its answers look.
Who is responsible for the opinion?
The expert, without qualification. AI changes how quickly the records are read and organised; it does not move the duty to the court or the authorship of the conclusions. A report is still the expert's reasoning on facts they have satisfied themselves are accurate, and the citations are what let them do that satisfying quickly. Used this way, the technology supports the Part 35 duty rather than competing with it.
Why it matters
That is the principle DOQSynth is built on. It reads the whole record, extracts the facts into typed tables, builds a chronology and answers questions in plain language, and it attaches the source page to what it returns so the expert can check it. The model runs on infrastructure you control and every step is logged. The reading gets faster, and the judgement, which is the point of the instruction, stays exactly where it belongs.
