From GP Surgeries to Hospital Wards: The Unseen Test for DYAD and NHS AI
On a Monday morning in a GP surgery, the post has not stopped arriving. Discharge summaries and outpatient letters fill a shared inbox, and someone must read each one, decide what it asks of the practice and translate a paragraph of clinical prose into coded entries in the patient record. Nobody outside the building sees this work, and no waiting list statistic captures it. Yet it is where a great deal of the NHS's information quietly stalls.
That is the territory DYAD occupies. The company has raised $10 million in a Series A round led by Sangha Capital, and it intends to carry its clinical document-processing software beyond primary care into secondary care and mental health services, with further expansion planned in the Middle East and the United States. Its BetterLetter product is used in more than 200 GP practices and has processed over five million documents. Those are respectable numbers for a young company in a market that has chewed up many digital promises. They are also small against the volume of correspondence moving through the service every day.
The technology is worth understanding because it is less glamorous than the generative AI discussed in ministerial speeches. DYAD's approach, which it calls Graph Enabled Information Extraction, layers natural language processing, knowledge graphs and clinical coding taxonomies, and brings in a large language model only at the final stage to resolve ambiguity. Every suggestion is reviewed and approved by practice administrative staff before it reaches the permanent record. This is a sensible design for a setting where an error can follow a patient for years, and it reflects a lesson the service has absorbed slowly: the model is rarely the hard part.
The human review step is the crux of the argument. Efficiency gains from automation depend on checking being materially faster than doing, and that holds only if the software is accurate enough that reviewers do not slip into either rubber-stamping or redoing the work. Rubber-stamping is a patient safety problem. Redoing it is a productivity problem the NHS cannot afford, with practices stretched, waiting lists stubbornly high and a government that has staked its reform narrative on shifting the service from analogue to digital. Procurement teams and integrated care boards should ask suppliers for evidence on reviewer time and correction rates, which tells them more than accuracy percentages on a slide.
Secondary care and mental health raise the stakes considerably. Hospital letters vary enormously by specialty, author and trust, and mental health records contain narrative judgement, risk assessments and sensitive disclosures that resist tidy coding. A system trained on the settled conventions of GP correspondence will meet a harsher test there. Information governance obligations also grow heavier as data moves across organisational boundaries, and trust boards will want clarity on clinical safety cases, data protection impact assessments and where responsibility sits when an extracted entry proves wrong.
The international ambitions deserve a sceptical reading too. A company seeking growth in the Gulf and the United States is partly responding to commercial logic, and partly to the fact that scaling inside the NHS means negotiating hundreds of separate buyers with differing systems and appetites. Ministers say they want British health technology to thrive at home. The fragmented purchasing landscape, in which a product proven in 200 practices still has to win each new customer individually, makes that harder than the rhetoric suggests. A national route for assured, interoperable tools would do more for UK life sciences than another strategy document.
For NHS leaders the practical lesson is to treat correspondence handling as infrastructure rather than administration. Trapped information in letters delays referrals, medication changes and follow-up, and that delay is a quiet contributor to operational pressure. For policymakers, the lesson is that public trust in NHS AI will be built through dull reliability in routine tasks, and that a single high-profile failure in a sensitive setting such as mental health could set the whole agenda back.
DYAD may well succeed. Whether the NHS can absorb its success without lowering its safeguards is the more demanding question.
