Most of the conversations about AI in healthcare still treats it as a tool. Something a clinician picks up, uses for a task, and puts down again. A faster dictaphone, a smarter search bar. A way to get the note written before the next patient walks in.
That framing made sense when AI in medicine meant a single purpose algorithm doing a single purpose job. The systems we are now putting in front of clinicians can hold context, reason across a patient's history, and produce structured documents. It is no longer whether AI can be useful in a clinical workflow, It is what kind of relationship the clinician is actually having with it.
Earlier this year a study in npj Digital Medicine ran a randomised trial of two collaborative workflows, AI as a first opinion and AI as a second opinion, and found both improved clinician diagnostic accuracy compared with conventional resources. As large language models start performing at expert level, the interesting question is no longer whether they can contribute. It is how they integrate into the way clinicians actually work.
Where does Behavioural Health fall in this era
Psychiatrists carry one of the heaviest documentation burdens in medicine. The published estimate is around three hours of documentation per working day, which most psychiatrists would say is conservative. Mental health assessments are long, narrative, and dense with context. A good initial assessment can take an hour and produce a note that takes another hour to write properly. Multiply that across a clinic week and the maths becomes obvious: the documentation is, in effect, a second clinic that nobody is paying for.
This is the gap that AI co-working actually fills. Not by replacing the clinical thinking but by holding the parts of the work that do not need clinical thinking, so the parts that do can have more room.
What the evidence is starting to show
In October 2025, a JAMA Network Open study of 263 clinicians across six US health systems found that after thirty days of using an ambient AI scribe, burnout dropped from 51.9% to 38.8%, Clinicians reported being able to give patients more focused attention. The senior author at Yale described it as a tool that allows technology to fade into the background and allows care to come to the foreground.
In the UK, evidence comes from a Central and North West London NHS Foundation Trust pilot in CAMHS, evaluating ambient voice technology in ADHD medication reviews, general medical reviews, and developmental history assessments. The pilot reported significantly reduced documentation burden and is now in phase two, expanding across five boroughs. It is one of the only published mental health AI documentation pilots in the NHS, and the early signal is consistent with the international evidence.
The point worth holding onto is that none of these studies describe AI replacing the clinician. They describe AI sitting alongside the clinician, taking a specific kind of work off their hands, and freeing them to do the work only they can do. That is the teammate model we are moving towards.
What clinic owners should actually ask
If you are evaluating AI tools for a psychiatry practice, the question is not whether the tool can produce a note. They almost all can, to varying degrees. The questions worth asking are:
-
Does the tool understand psychiatric narrative, or has it been trained on general medical conversations and adapted afterwards? The difference shows up in the quality of the assessment more than anywhere else.
-
Is the clinician genuinely in charge, with the ability to verify and correct what the AI has produced, or is the system opaque?
-
What is the tool doing with the time it gives back, in terms of how it shapes what happens before and after the consultation? And is the company building it accountable for clinical safety in a way you would be comfortable defending to a regulator?
Where this is going
The teammate framing is going to feel obvious in five years, in the same way that the shift from paper notes to electronic records feels obvious now. We are at the point in the curve where the technology has caught up with the ambition, the early evidence is starting to cohere, and the practical question for clinicians and clinic owners is no longer whether to engage with this but how to engage with this.