A GP consultation used to involve two people. Increasingly, it involves three. An AI tool may be listening, transcribing and drafting a summary of what was said. Are the words spoken in a consultation always the same as what the patient meant?
The promise is understandable, and the need is real: less time spent typing and more attention on the patient. But a consultation is more than a sequence of words. Tone, context, hesitation and what a patient chooses not to say can all matter.
As these tools move into practice, we need to look beyond the time saved and ask what happens to the clinical record once AI becomes part of how it is created.
Ambient voice technology (AVT), sometimes called an AI scribe, is no longer confined to small pilots.
In July 2026, NHS England’s Midlands region became the first to procure ambient voice technology at scale for all 1,239 GP practices in the region, as well as more than 70,000 clinicians across 15 acute and community trusts.
Use among GPs is already high. A 2026 survey of 598 practising UK GPs found that 40% were currently using AI scribes, with a further 23% having used them in the past.
These tools are therefore already affecting how consultations are documented, and their use is likely to grow.
There are clear benefits.
An NHS England-sponsored evaluation led by GOSH DRIVE included more than 17,000 patient encounters in outpatient clinics across nine London NHS sites. It found a 23.5% increase in direct patient interaction time and an 8.2% reduction in overall appointment length when AI scribes were used. Separately, a conference abstract from a single London emergency department reported a 13.4% increase in patients seen per shift.
But they are largely measures of time and activity. They tell us less about what happens to the quality and accuracy of the clinical record, the patient’s story or the clinician’s own relationship with that record over time.
A University of Edinburgh review of 27 published articles looked beyond productivity at how ambient scribes interact with clinical work and patient-clinician relationships. It found that much of the evidence focuses on efficiency rather than wider consequences, and argues that scribes should be understood as complex interventions that can shape workflows, interactions and clinical documentation.
That matters because an audio-based system does not experience a consultation in the same way a clinician does.
Facial expressions, gestures, emotional state and context may carry meaning that is not captured in speech alone. A summary may prioritise clinical facts while giving less weight to the patient’s own account, the clinician’s interpretation or the uncertainty around what was said.
There are practical risks too. Drug names and diagnoses can be missed or misspelt, and in clinical data and medicines safety a small slip matters.
At DXS, we know that once incorrect information begins moving through records, letters or referrals, fixing it becomes much harder than preventing the error in the first place.
What counts is the product’s intended purpose. Tools limited to administrative functions such as transcription, summarisation, drafting letters or matching explicitly stated terms to clinical codes may fall outside medical-device regulation. Where a product is intended to support diagnosis or treatment, or acts autonomously on clinical information, medical-device requirements may apply.
That distinction does not remove the need for clinical responsibility. NHS England guidance continues to emphasise appropriate governance, clinical safety, transparency and clinician review when these products are adopted. Safety therefore depends on local clinical-safety assurance, transparency with patients, appropriate permission where required, and clinicians reviewing outputs before they enter the record.
On 10 September, the National Commission into the Regulation of AI in Healthcare published recommendations for a more proportionate, lifecycle-based and system-wide approach to AI regulation and assurance.
On 6 October, the government accepted all 44 recommendations and set out its first delivery priorities. This is a commitment, not yet a new set of rules, and an implementation plan and roadmap are due by spring 2027.
A consultation note is not just paperwork. It can influence coding, prescribing, referrals, follow-up and the legal clinical record.
If an AI-generated summary changes the emphasis of what was said, even subtly, that change can travel with the patient.
If AI-drafted text feeds into a referral, an error can move from one service to another faster and further.
For those of us building connections that carry clinical information between systems and services, that means being clear about where AI-generated content comes from and maintaining a traceable clinical-safety process from the start. Structured clinical terminology such as SNOMED CT can support consistency and standardisation across systems, but it does not remove the need for human review and clinical-safety controls.
AI scribes are already part of clinical practice. The question is whether the words said and the meaning are always the same. The answer is: not always, and that gap is where the risk sits. The technology still has a place, but accuracy and clarity depend on how people and tools work together – and on asking patients, not only clinicians, how the experience feels.
Clinicians need enough time and responsibility to review what AI produces properly, not as a tick-box action. Healthcare organisations also need to keep monitoring how these systems behave in everyday use.
Saving time matters, but the record still has to reflect the patient and clinical details accurately.
If an AI-drafted note or code feeds a referral and something goes wrong, everyone involved – clinician, practice, patient and technology supplier – needs to understand where responsibility sits.
Interested in how digital healthcare systems can support safer, more reliable clinical information across the patient journey? Get in touch – hello@dxs-systems.co.uk
