A practice manager opens an email on a Monday morning and finds a letter of twenty pages. It cites statutes, alleges breaches of duty and demands a response within days. Near the top, the sender has left behind a line addressed to the software, an instruction on tone that was never deleted. Somewhere in the document is a grievance about a referral that has not materialised. It takes an hour to find it.

Scenes of this kind are now routine. A survey of nearly 160 GP practice managers found that 93 per cent had received a complaint drafted with artificial intelligence, and 45 per cent said such letters were adding to administrative burden and damaging morale. The sample is modest and the findings come via a medical defence organisation with an interest in the subject, so they warrant care. Even so, the direction matches what practices describe, and the mechanism is easy to understand.

Complaining has always carried a cost in effort. Writing a coherent letter took time and confidence, and that friction filtered out some minor grievances while also deterring people with serious ones. Generative tools remove the friction for the sender. The cost of responding stays where it was. Every complaint a practice receives must be investigated and answered within set timescales, whether it runs to one paragraph or twenty pages. An asymmetry has opened between the price of making a complaint and the price of handling one, and general practice sits on the wrong side of it.

It would be a mistake to treat this as a story about patients misusing technology. The Medical Protection Society rightly acknowledges that AI can help people with limited English or low confidence in writing to put their case. That is a real gain in access to a process that has long favoured the articulate. The difficulty lies in the output: inflated legal language and length that buries the clinical concern. The ombudsman advises patients to describe what happened plainly and to avoid asking a tool for legal flourishes. It also warns against entering health details into commercial platforms whose data handling is opaque. Both points are sensible.

The wider context matters more. The Royal College of GPs observes that many of these complaints originate in delays the practice did not cause, such as hospital backlogs and gaps in community services. A patient waiting a year for an orthopaedic appointment has few formal channels for anger, and the surgery is the most reachable institution. The complaint lands on the GP because the GP is there. Seen this way, the letters amplify an existing signal, and the signal is about waiting lists.

For NHS England and integrated care boards, this is a capacity question that current planning does not capture. Complaint handling obligations sit within the GP contract, yet workload models rarely count them, and practices already contend with demand that outstrips staffing. Commissioners could issue proportionality guidance that lets practices triage complaints by substance. They could also establish clearer routes for hospital and community providers to receive complaints about their own delays, so that the surgery stops acting as the default recipient. Practices may use the same technology to summarise long correspondence and extract the factual allegations, provided data governance is sound. A sector answering machine-written letters with machine-written replies is an outcome nobody should want to normalise.

For policymakers, the episode previews a larger problem. Public services built on an assumption of friction now face a world where contact is cheap. Complaints and appeals share the same structure as many other citizen interactions with the state. Government has invested heavily in the promise of AI for efficiency inside public institutions, with less thought given to what happens when citizens adopt it first.

Complaints handling belongs in the NHS digital strategy, with resources attached. Until waiting lists fall, the letters will keep arriving, longer and more elaborately phrased, and general practice will keep paying for delays it did not cause.

‍