A single hospital letter can create several jobs for a GP practice. Someone has to match it to the patient, identify any follow-up, update the record and get it to the right team. At Northgate Medical Centre in Chester, one GP said in an NHS England case study that doctors there had spent up to 40 minutes a day reading correspondence. The practice then rewrote its protocol and trained its administrative staff. GPs now receive only 22% of the practice’s incoming correspondence and spend about 10 minutes a day on it, with most of the rest handled by administrative staff and medication changes sent to the practice’s medicines manager. Software is now beginning to take on that administrative work. What interests us is the level of automation that can be safely achieved, and its trickle-down effect on patient care. The AI that reads documents, handles speech and makes simple decisions is being built for almost every industry. A company processing letters for a practice like Northgate can use each improvement in it without funding the research itself. What is software already doing? A 2025 NHS South East report asked the AI suppliers working with the region’s 780 GP practices where their tools were in use, and around 60% of the 46 suppliers contacted replied. They reported automated letter filing and coding (labelling records with standard clinical terms) in 197 practices and consultation transcription and summarisation in 222, so each had reached roughly a quarter of the region’s practices by mid-2025. Source (NHS South East, October 2025). Supplier-reported in mid-2025, across the region’s 780 practices. A practice can appear in more than one category. Anima processes incoming letters for GP practices. Its 9 October release notes describe improvements to identifying implied actions and their timeframes, and fewer, more relevant task suggestions, which practice staff then confirm, edit or skip . If a letter asks a practice to arrange a blood test in six weeks, a summary saves some reading. A service that also prepares the task, with the timing and team attached, would save more of the work. A Bessemer survey of 226 healthcare executives across providers, payers and pharma companies, with research from Bain and published on 30 September, found self-reported returns concentrated in administrative work. Providers said revenue-cycle tools, which handle billing and payment, returned 4.0 times what they cost, against 2.9 times for clinical tools, although the survey reports no NHS figures. Source (Bessemer, with research from Bain, 30 September 2026). Self-reported by 226 executives; provider use cases shown. Why will software keep taking on more of the work? A company building one of these services can use a language model, the kind of AI behind ChatGPT, to read a letter, and its own rules to decide who should see it. It then records the result in the practice’s existing systems. A better model can mean fewer letters need a person to read them; a better connection to those systems removes another manual step. One newer option is TypeSafe’s Jev , released on 15 September. According to Vercel’s launch post, it takes a piece of text and a set of questions about it, each with possible answers, and returns a probability for each answer. A service can change the questions and answers whenever it needs to. TypeSafe’s documentation for its models says the questions in a single request can run together , with little extra response time. For a letter-processing service, those questions might cover whether the practice has been asked to act, whether a medication change is mentioned and which team should review it. After testing the scores, the service could accept straightforward answers and send uncertain ones to a person. The same approach is already used in hospital coding, where staff turn each patient’s notes into the standard codes used for payment and planning. Jigsaw says it codes around 80% of routine, low-complexity work and hands anything it is unsure of back to a hospital’s own coders. University Hospitals of North Midlands started using it for endoscopy day cases in June 2026. Its head of clinical coding says on Jigsaw’s site that the team “now focuses on the most complex cases”. Who is best placed to take on the work? We expect better models, and better connections to practice and hospital systems, to keep expanding what these services can offer. A service might begin by helping staff process letters, then take on more of the filing, task creation and chasing that surrounds them. Existing suppliers benefit from the same progress. In May 2025 Accurx announced it had added Tandem’s scribing technology to software already used in GP practices. A new company needs to make a particular job so much easier that customers change how they work. Making a job that much easier depends on knowing it in detail. A faster model has limited value if staff have to recheck everything, copy its output into another system or chase the same missing information themselves. Beyond Jigsaw’s own estimate for routine hospital coding, we have not seen published figures for the share of documents software handles without a person reading them, or for how often staff overrule it. Those two numbers, along with the staff time saved, are what we would want a company here to show, because they decide how much of the work can be safely automated. For Eka, the appeal is a business that can take on more of an essential service at a lower cost as the technology improves. It can specialise in healthcare while drawing on research done for a much wider market. The benefit we are looking for is less staff time spent on administration and more available for care. If you are building here, we would love to hear from you 👋 Subscribe now