A story circulates that England’s NHS has rolled out automated triage in emergency departments, with independent trust-by-trust evaluations to prove it. Looking for the documents turns up three other things: an NHS England clinical guidance on initial assessment, a research prototype still in development, and a vignette study about self-triage, not about the nurse at the desk. The gap between those three and the story is the point.
Anyone walking into an English emergency department on a January afternoon meets, in order: a desk, a person asking a few questions, a quick set of observations, and then the wait. That first contact has a technical name and a written procedure. NHS England publishes Guidance for emergency departments: initial assessment, which carefully separates three processes that are routinely confused: streaming (sending the patient down the right pathway), triage (assigning clinical priority), and rapid assessment and treatment (starting assessment and care immediately). The guidance was developed with the Royal College of Emergency Medicine and in consultation with the Royal College of Nursing, the Care Quality Commission, regional urgent and emergency care advisors and the Society of Acute Medicine. It applies to England. And it describes a clinical process, carried out by people.
That is where the problem starts. For a couple of years a different story has been circulating around that desk: that the NHS, crushed by record waits, has extended artificial intelligence triage tools into emergency departments — machine learning models trained on clinical data (reported symptoms, vital signs, demographics, medical history) that output a priority code automatically. It is a plausible story. The technology exists, the pressure exists, the commercial interest exists. But the documents that should support it are not there. There is no NHS England announcement of a national rollout of automated triage in emergency departments, and no trust-by-trust dataset measuring what such tools did to waiting times or under-triage rates. Asked for “the numbers from three trusts”, the honest answer is that those numbers do not exist in public form.
What does exist, and at what stage
First, a guidance document that is not about automation at all. The NHS England initial assessment guidance replaces two earlier documents — Clinical streaming in the accident and emergency department (2017) and RCEM’s Initial assessment of emergency department patients (also 2017) — which recommended that initial assessment begin within fifteen minutes of arrival. It also connects to the NHS planning guidance for 2022/23, which required more structured streaming arrangements between urgent treatment centres and emergency departments. In substance it is an organisational intervention: who sees the patient first, to what end, and where they send them.
Second, there is a research project whose acronym tells you how far we are from the ward: DAISY, the Diagnostic AI System for Robot-Assisted Triage. The study published in Frontiers in Digital Health in 2023 describes it explicitly as a system in development to support triage in English NHS emergency departments, and gathers clinicians’ views on a hypothetical deployment. It does not measure clinical outcomes, does not compare triage decisions, does not follow patients. It is acceptability research — a legitimate and necessary stage, but the stage at which you ask clinicians what they would think, not the one where you look at what happened.
Third, the most frequently misquoted item. A study published in Cureus on 26 November 2025, authored by Hannah L. Brown and affiliated with East and North Hertfordshire NHS Trust, compared general-purpose AI platforms with the NHS 111 Online symptom checker on case vignettes, scored against standards built from NICE guidelines. The reported result: both AI platforms identified every emergency case, while NHS 111 under-triaged one; on non-urgent cases, NHS 111 classified them all correctly. Interesting. But it concerns self-triage — what a member of the public does at home in front of a screen before deciding whether to move — not the decision a triage nurse makes at the department desk. And it concerns simulated cases, not patients.
Three different things that the word “triage” holds together badly
Here the confusion becomes technically interesting, not merely journalistic. The same label covers systems with incompatible inputs, risks and lines of accountability:
- Self-triage: the user describes symptoms in their own words. The system has no measured vital signs, no verified history, no sight of the patient. The expensive error is the false negative — telling someone having a heart attack to stay home — which is why these tools are tuned to be cautious, that is, to send more people to hospital than strictly need to go.
- Decision support inside the department: the system sees measured observations, coded clinical data, sometimes the record. It works beside a professional who can overrule it. The expensive error shifts: it is no longer only under-triage, it is automation bias — the documented tendency to go along with the machine’s suggestion when you are tired, under pressure and short-staffed, which is precisely the state of an emergency department in January.
- Automated triage proper: the system assigns the priority code. That requires a chain of clinical and regulatory responsibility which, in the English public record, I could not find mapped out.
Treating these as more or less advanced versions of the same thing is what generates wrong headlines. A model that works beautifully as a cautious filter for someone on the sofa can be terrible as a priority-setter in a full waiting room, because in a waiting room caution is not free: every priority raised too high moves resources away from someone else standing there at the same moment.
The numbers that exist, and why they don’t add up
The pressure is documented. According to NHS Digital figures, around 347,707 patients waited twelve hours or more in A&E in 2022, against 1,306 in 2015. A RCEM review reported in the press puts 1.7 million people at over twelve hours waiting for A&E care in the year considered, with a government response citing £450 million invested in urgent and emergency care — a figure and year I could not confirm against a primary source.
The gap between 347,707 and 1.7 million is not somebody’s mistake: they are two different metrics. NHS England’s official statistic counts the twelve hours from the decision to admit; the measure used by RCEM counts from arrival at the department. The numbers are not comparable, and anyone placing them side by side without saying so is building a comparison that does not exist. Worth noting: this is exactly the ambiguity that poisons any evaluation of a triage tool. If you do not fix where the clock starts, any claimed “reduction in waiting times” is negotiable.
The most useful figure is the missing one
No source I found establishes a link between the adoption of digital tools and the course of winter waits. That does not mean none exists; it means it has not been measured publicly — which, in a health system that publishes monthly A&E statistics, is itself information.
And here is something a reader outside the UK is unlikely to know, and it explains why this verification is harder than it looks. In England, health technology procurement is decentralised: each NHS trust buys on its own account. There is no single national act to look for — there are hundreds of separate contract registers, published in different formats at different times. A claim of the form “the NHS has adopted X” is therefore almost always a generalisation from one or two pilot sites, and the structure of the system makes it extraordinarily hard to disprove. Absence of evidence does not become evidence of absence: it becomes noise. It is inside that noise that a technology still in development can look, for a winter or two, like a technology in use.