5 weeks to 5 minutes

Outpatient medicine has a problem: the consultation happens now, but the letter arrives later. You see the patient, make the decisions, dictate after the consultation, and only after transcription and review is the letter finally sent to the GP and the patient. Sometimes weeks have passed. The gap is so familiar that we hardly notice it.

Ambient voice technology (AVT) attacks that gap. In my view, that is its real benefit.

We have been told about productivity and AI liberating us from administration, but the real shift is simpler: a clinic letter that once took five weeks to reach the reader can now be reviewed, printed, and given to the patient in five minutes.

For the past 6 months I have used AVT in my cardiology clinic. The workflow is simple. The patient comes in, verbal consent is obtained, I switch the microphone on, consult as usual, except that I say the medications and examination findings out loud, then review the generated letter, print it, and the patient leaves with it. Consultation, documentation, and communication now sit in the same time frame. That is a meaningful change, and probably the most important one.

The evidence base, at least so far, broadly matches this experience, though with more modest gains than the hype suggests. A 2025 randomised trial of ambient AI scribes in outpatient practice found reductions in time spent on documentation for one platform and possible improvements in burnout and task load. But it also reported occasional inaccuracies and emphasised the need for clinician vigilance. A 2026 multisite JAMA study showed only modest reductions in total EHR time and documentation time, alongside a small increase in visit volume. These are not transformative gains.

So, this is not about saving time. Clinic does not suddenly run faster. You still read the letter, check the medications, and decide whether the generated prose reflects what happened. You still carry the medicolegal responsibility. The work is not abolished. It is relocated.

What used to be delayed dictation and deferred correction becomes immediate review. Time is not removed from the system; it is used differently.

That matters because deferred work is toxic. It occupies mental bandwidth, produces evening admin, invites recall errors, and leaves a persistent sense of incompleteness. The old model asks the clinician to do two jobs separated by time: first the consultation, then the reconstruction of the consultation. Ambient systems collapse those two jobs into one episode. That is why the technology feels better. Not because it removes work, but because it removes delay.

There is another reason it feels better: the patient’s voice is captured more faithfully. Anyone who has run a busy clinic knows the weakness of traditional letters. By the time we dictate, we are no longer simply recording the story - often we are rewriting it. The temporal sequence has been cleaned up and compressed. Once we think we know the diagnosis, confirmation bias tends to emphasise the features that fit and quietly discard those that do not.
​
Ambient systems are often better at preserving the patient’s account of breathlessness, chest pain, and palpitations, and more accurate in capturing onset, duration, and triggers. That is not trivial in cardiology, where chronology often drives diagnosis.

But fluency is not fidelity.

These systems can generate confident falsehoods. In my experience, the clearest example is medication allergies. If allergies are not explicitly discussed, the draft may still insert “no known allergies” or an equivalent negative statement. That is not a stylistic error. It is a safety problem.

The danger is not that the text looks wrong, but that it looks so polished as to be credible. A generated letter can be wrong in a far more seductive way than a bad dictation. It is structured, coherent, and confident. It invites approval.
This changes the clinician’s task. You are no longer asking, “How do I say this?” You are asking, “Did this system just say something I never established?” Editing is not the same as writing. It requires a different, and in some ways sharper, form of attention.

What happens when the system fails, if the output is incomplete or if you forget to switch the microphone on. Safe use doesn’t just depend on good drafts. If the AI is unavailable, uncertain, or clearly wrong, the clinician needs an obvious manual fallback rather than a gap or a polished but misleading output. In other words, these systems do not just need accuracy; they need a safe backstop.

One potential downside, rarely discussed amid the enthusiasm for these systems, is what happens to trainees. Writing a clinic letter is not clerical work; it is a cognitive act. It forces you to organise a history, prioritise signal over noise, commit to a diagnosis, and articulate a plan coherently for another clinician. That process - the translation from messy clinical reality into structured narrative - is where much of the learning sits.

If ambient systems assume that role, trainees risk becoming editors rather than authors. Editing is reactive, not generative. Over time, that may erode the ability to independently synthesise complex information, particularly in specialties like cardiology where nuance in symptom chronology, investigation interpretation, and risk framing matters. The analogy is obvious: just as reliance on imaging can atrophy examination skills, reliance on generated documentation may atrophy clinical reasoning as expressed in writing.

Patients, for their part, seem comfortable with the technology. When the purpose is explained clearly, most accept it readily. In practice, verbal consent is straightforward and often reassuring. Some patients like the idea that their account is being captured accurately. Most value leaving clinic with a printed letter in hand.

That last point should not be underestimated. A letter delivered at the point of care is not just faster; it is more transparent. The patient can read it, correct it, and leave with a record of the plan rather than waiting for a delayed and often forgotten communication.

The same immediacy helps with referrals. Ambient systems automatically draft referral letters, reducing the friction of onward referral while the clinical reasoning is still live. The responsibility remains, but the barrier is lower.

The question is whether this early experience generalises in a robust way. Here the evidence is still thin. Most studies are multispecialty outpatient evaluations or primary care implementations. There is little cardiology-specific data. That matters. Cardiology is not generic outpatient medicine. It is a field of valve gradients, ECG intervals, medication doses, and risk-laden omissions. It is one thing for an ambient system to produce a competent note for a straightforward follow-up. It is another for it to reliably handle the patient with mixed angina and dyspnoea, prior PCI, LV impairment, complex medication decisions, and multiple referral pathways.

We should stop asking whether ambient voice technology is “good” and start asking more precise questions. Where does it work well? Where does it fail? Which errors matter clinically? How often does it introduce false negatives or misplaced certainty? These are the questions that will determine whether this becomes routine practice or remains a promising but imperfect tool.

Regulators have already recognised that this is not a trivial addition. NHS England has issued guidance on the use of AI-enabled ambient scribing systems, including requirements around consent, transparency, and information governance. That is appropriate. These are clinical documentation tools, not convenience software.
So where do I stand after using it for half a year?

Positive, but not evangelical.

It works. It handles a range of patient voices better than expected. It captures symptoms well. It produces usable letters. It removes the need for traditional dictation and creates a clearer real-time record of the consultation. Most importantly, it collapses the interval between consultation and communication from weeks to minutes.

But it is not magic. It does not abolish work. It does not remove the need for review. And it does not earn trust simply by sounding competent.

It may not shorten clinic, but it shortens the distance between decision and documentation.

That is enough to matter.