Coming soon — Shifaa AI is in private beta.Join the beta
Shifaa AI
Healthcare technology

AI for general practitioners: a practical guide

What the evidence says about AI in general practice: how many GPs use it, where it helps a GP's day, what it gets wrong, and how to adopt it safely.

Shifaa AI Team14 min read

AI is already in general practice, and the fastest-growing use is not diagnosis — it is paperwork. Across three annual surveys of UK GPs, reported use of generative AI in clinical practice rose from 20% in February 2024 to 25% in January 2025 and 52% in January 2026, and in every wave the leading uses were documentation and correspondence rather than diagnostic reasoning (Blease et al., BMJ Health & Care Informatics, 2024; Blease et al., Digital Health, 2025; Blease et al., BMJ Digital Health & AI, 2026). That is the honest shape of AI for general practitioners today: strong on the admin that surrounds the consultation, genuinely useful as a second opinion on a differential, and nowhere near good enough to be trusted unreviewed.

This guide is written for GPs and family physicians deciding whether, and how, to let AI into their week. It covers what the adoption data actually shows, why general practice feels the documentation load more sharply than most specialties, what the evidence says about time and burnout, where AI helps with diagnostic breadth, how it behaves outside English, and what safe adoption looks like in a practice. Every figure below is sourced, and where the evidence is weak, it says so.

How many GPs are actually using AI?

Just over half — 52%, or 522 of 1,004 UK GPs surveyed in January 2026 — reported using generative AI in clinical practice, roughly double the 25% recorded a year earlier and more than twice the 20% recorded in February 2024. Among the 2026 users, 70% said these tools had reduced their work burden.

The task mix is the more revealing number, because it shows what GPs actually trust AI with. In the 2026 wave, users reported applying it to:

  • Writing letters — 51%, the single most common use
  • Real-time documentation during consultations — 41%
  • Referrals — 31% and documentation after consultations — 30%
  • Differential diagnosis — 28% and treatment options — 27%
  • Patient summarisation or timelines — 12%, and medical certification just 3%

ChatGPT remained the most-used tool (67% of users), followed by a purpose-built clinical scribe (46%) and Microsoft Copilot (24%) — which tells you that a large share of this use is happening in general-purpose consumer tools rather than governed clinical software.

And the support around that use is thin. The 2025 wave found 95% of GPs had no professional training in using generative AI at work and 85% said their employer had not encouraged it in the previous 12 months; the 2026 authors note that earlier waves found minimal employer or commissioner encouragement and almost no formal training. Adoption is running well ahead of governance, which is precisely the condition in which avoidable errors happen.

All three waves drew convenience samples from the same doctors' professional network, which probably over-represents digitally engaged GPs, and no response rate could be calculated for the 2026 wave. Read the percentages as a direction of travel, not a precise national figure. The authors say as much themselves.

The one-line summary

For a GP in 2026, AI is a documentation and recall aid that can also widen a differential. It is decision support, not a decision-maker: the note, the diagnosis and the prescription remain yours to review, correct and sign.

Why does general practice feel the documentation burden so sharply?

Because the ratio of record-keeping to consulting time is worse in primary care than almost anywhere else, and the consultations themselves are short. A time-and-motion study of 142 family medicine physicians found they spent 5.9 of an 11.4-hour working day in the electronic health record — 4.5 hours during clinic hours and a further 1.4 hours after clinic — with clerical and administrative tasks accounting for 44.2% of that EHR time and inbox management another 23.7% (Arndt et al., Annals of Family Medicine, 2017).

The global picture compresses that further. A systematic review of 179 studies covering 28,570,712 consultations in 67 countries found mean primary care consultation length ranging from 48 seconds in Bangladesh to 22.5 minutes in Sweden, with 18 countries — representing roughly half the world's population — at five minutes or less (Irving et al., BMJ Open, 2017). The same review found shorter consultations significantly associated with the reduced-personal-accomplishment dimension of burnout and with physician dissatisfaction over time per patient, though it cautions that these associations rest on very few data points.

So the problem a GP is trying to solve with AI is specific: an undifferentiated caseload, a short slot per patient, and a documentation tail that follows you home. For the full picture of that burden and what else moves it, see our guide to physician burnout.

Where does AI genuinely help in a GP's day?

In five places, roughly in descending order of how widely GPs already use them: drafting letters and referrals, drafting the consultation note, summarising a long record before you open the door, widening the differential on a case that does not fit a pattern, and running the queue so the day stays on time. Everything else is still experimental.

  • The note. An ambient scribe listens to the consultation and drafts a structured note — usually SOAP — which you then correct and sign. This is the highest-value, lowest-risk use, and the one GPs reach for first. See what an AI medical scribe is and how to write a SOAP note for the format it should produce.
  • Letters and referrals. Turning an agreed plan into a letter is reformatting, not reasoning, which is why this is now the most common use of all — 51% of GP users for letters, 31% for referrals. The clinical content still has to be checked against what you actually decided.
  • Record summarisation. Building a patient summary or timeline from prior documentation is useful for a 10-minute slot with a 15-year record, and 12% of GP users do it — the lowest-uptake major task, which likely reflects how little most GPs trust a summary they cannot see the workings of. Verify anything you intend to act on.
  • Differential breadth. Decision support can surface the dangerous mimic you had not voiced, with a reference you can open. It should widen your list, not shorten it.
  • Flow. A queue or token system that keeps the clinic moving is not AI at all, but it is often the change a busy OPD feels first.

Shifaa AI, which is in private beta, is built around exactly that split: the AI medical scribe drafts the note from the conversation (OpenAI's Whisper for multilingual transcription, Anthropic's Claude to draft — both disclosed sub-processors), differential-diagnosis support surfaces ranked possibilities with the guidance they come from, and the patient queue keeps the list in order. The clinician reviews, edits and signs everything. There is a general practice overview if you want to see the workflow end to end, and a free printable SOAP note template if you would rather start on paper.

What does the evidence say about time saved and burnout?

The best current evidence is encouraging but soft. In a quality-improvement study of 263 clinicians — 131 of them in primary care — across six US health systems, 30 days of ambient AI scribe use was associated with burnout falling from 51.9% to 38.8% (adjusted odds ratio 0.26, 95% CI 0.13–0.54) and self-reported after-hours note writing falling by 0.90 hours per week, about 10.8 minutes per working day (Olson et al., JAMA Network Open, 2025).

Read the caveats before you quote the numbers. There was no control group, so background trends could not be separated out. Outcomes were self-reported on a non-anonymous survey, with no EHR log data to corroborate them. Participants were early adopters who chose to take part. The scribe vendor facilitated the data collection and two authors held roles with it. The authors themselves describe the time savings as modest and note that burnout improved more than documentation time did — which suggests part of the benefit is cognitive relief rather than minutes on a clock.

The practical reading for a GP: expect roughly ten minutes a day back and a meaningful drop in the feeling of being behind, not an hour of reclaimed clinic time. Anyone promising the latter is ahead of the published evidence.

Can AI help with diagnosis in general practice?

It can help you think of more possibilities; it cannot make the diagnosis. Some 28% of the GPs using generative AI in 2026 already used it to suggest differentials, and the gap it speaks to is real: an estimated 5% of US adults — about 12 million a year — experience a diagnostic error in outpatient care, around half of them potentially harmful, and the authors believe that is an underestimate (Singh et al., BMJ Quality & Safety, 2014).

General practice carries a particular version of that risk. You see undifferentiated presentations at low prevalence, which means the serious diagnosis arrives disguised as something ordinary. A tool that reliably says "consider this too, and here is the guidance" is genuinely useful at 5pm on a full list. A tool that ranks a single answer confidently is a liability, because the failure mode is not the tool being wrong — it is you agreeing with it. That is automation bias, and it is the main reason decision support has to show its reasoning. Our explainers on clinical decision support systems and whether AI can diagnose patients go deeper.

An illustrative example of the useful behaviour: a 54-year-old attends with three weeks of fatigue and mild breathlessness on stairs. The working thought is deconditioning and poor sleep. Decision support that lists anaemia, new heart failure, atrial fibrillation, hypothyroidism and occult malignancy alongside it — each with the finding that would raise it — has done its job, because it prompts the FBC, the ECG and the question about weight you might otherwise have deferred to the next visit. It has not diagnosed anything. You still decide what to test and when.

Does AI work when the consultation isn't in English?

Less reliably, and the gap is documented. NHS England's guidance on ambient scribing warns explicitly that these products "may have varying success with some accents or regional dialects, and where the participants speak English as a second language, or are affected by speech disorders or impairments" — and asks adopting organisations to confirm the product works for their own population and to monitor it over time.

There is early evidence that bilingual documentation can work well. A prospective evaluation of a bilingual Arabic–English ambient scribe in Riyadh rated 55 real-world consultations (40 Arabic, 15 English) in family medicine clinics using a modified 9-item Physician Documentation Quality Instrument, scoring 42.4 of 45 for Arabic notes, with accuracy 4.53 of 5 for Arabic against 4.77 for English — a difference that did not reach significance (Khan et al., JMIR Medical Informatics, 2026). It is a small, single-arm study with no comparator, no objective time measurement, no formal error audit, dialect-level performance unanalysed, and several authors financially tied to the product — so treat it as a promising signal, not proof.

For a GP working across languages, the operational conclusion is simple: test the tool on your own consultations, in your own languages, before you rely on it, and keep reading every note. Our pillar on language barriers in healthcare covers the wider problem these tools are reaching into.

What does safe adoption look like in a general practice?

It looks like a short, boring governance list completed before the first patient, not after the first incident. NHS England's guidance on AI-enabled ambient scribing products (Version 3, first published 27 April 2025, last updated 29 July 2026; for England only) sets out the steps a setting should take, and the shape of it transfers reasonably well to any jurisdiction.

  • Name someone accountable. NHS England asks adopters to assign a Clinical Safety Officer and identify the technical and clinical risks, including missing or incorrect information and delayed outputs.
  • Do the paperwork. Complete the DCB0160 clinical risk documentation (safety case, hazard log, monitoring framework) and a Data Protection Impact Assessment, and ask the supplier for their DCB0129 clinical safety case report.
  • Check the regulatory position. Confirm the product has been correctly considered under medical device regulation — a tool that summarises and suggests is a different proposition from one that only transcribes.
  • Integrate it properly. The guidance notes that in most GP settings an ambient scribing product needs to integrate with the principal electronic patient record, in UK primary care via IM1, to avoid copy-paste errors and notes landing in the wrong record.
  • Keep the human in the loop, in writing. The guidance is explicit that settings should "ensure users review and approve any product outputs prior to further actions", and that training should reinforce practitioners' ongoing responsibility to review and revise what the tool produces.
  • Monitor after go-live. Audit documentation accuracy, review incidents, and collect your own monitoring data independently of the manufacturer's. Keep pilots time-limited — the guidance caps proof-of-concept at four months and says a pilot does not exempt a supplier from compliance.

Professional bodies elsewhere land in the same place on the clinical duty. The RACGP's guidance for Australian GPs states that GPs are ultimately responsible for ensuring the patient health record is accurate and up to date, requires patient consent before an AI scribe is used in a consultation, and warns that as use becomes routine clinicians may pay less attention to critical detail or stop checking the output (newsGP, August 2024). The WHO has published governance guidance for large multi-modal models in health, which is worth reading if you are setting practice policy (WHO, March 2025).

Where should AI stay out of a GP's work?

Out of anything that depends on what was not said out loud, and out of any step that ends in a decision without you. The RACGP guidance is blunt about the blind spot: a scribe cannot capture a hospital discharge summary, an imaging report, a non-verbal cue, or data sitting in a device — only what was spoken in the room. NHS England names a second, subtler failure: misreading clinical context, such as "mistaking a patient's hypothetical statement for a confirmed diagnosis".

  • Never let it sign. The note is a draft until a clinician reads it against what actually happened. Build the review time into the slot, not into the evening.
  • Never paste patient data into a general consumer chatbot. Use a tool whose data handling, retention and sub-processors you can name, under your practice's own agreement.
  • Never treat silence as a negative. An absent finding in an AI-drafted note means it was not said, not that it was excluded.
  • Never skip consent where it is required. Recording a consultation is a change the patient is entitled to know about and decline.
  • Never outsource the safety-netting sentence. What you told the patient to watch for, and when to come back, is the part of the note most likely to be litigated and the part a tool is least able to invent.

Used inside those limits, AI in general practice is a modest, real improvement: a draft note waiting when the patient leaves, a differential a little wider than the one you would have reached alone, and a slightly shorter evening. That is a worthwhile change — and it is a very different claim from the ones usually made for it. For the specialty view on the same split, see clinical AI for cardiologists, and for the broader landscape, our overview of AI in healthcare.

Frequently asked questions

How many GPs use AI in clinical practice?

In a January 2026 survey of 1,004 UK GPs, 52% reported using generative AI in clinical practice — up from 25% in January 2025 and 20% in February 2024. All three were convenience samples from the same professional network, so they indicate a trend rather than an exact national figure.

Does an AI scribe actually save a GP time?

The published evidence suggests a modest saving. A 2025 quality-improvement study of 263 clinicians found self-reported after-hours note writing fell by 0.90 hours per week — about 10.8 minutes per working day — with no control group and no EHR log data to corroborate it.

Can AI diagnose patients in general practice?

No. AI can widen a differential and surface a possibility you had not considered, with the guidance behind it, but the diagnosis, investigations and treatment decisions remain the clinician's. Tools that present a single confident answer invite automation bias.

Do I need patient consent to use an AI scribe?

Usually yes, and you should check your own regulator and indemnity provider. The RACGP requires patient consent before an AI scribe is used in a consultation, and some medical defence organisations ask for written consent before a conversation is recorded.

References

Medical disclaimer. This article is for general information for healthcare professionals. It is not medical advice, and Shifaa AI provides clinical decision support only — it does not provide a diagnosis, and the treating clinician is responsible for all decisions and patient care.

Keep reading

AI in healthcare

AI hallucinations in healthcare: risks and safeguards

Read article

Clinical reference

Shortness of breath differential diagnosis (dyspnoea)

Read article
Coming soon · private beta

Be first to try Shifaa AI.

Shifaa AI is in private beta. Join the beta to get early access, or share your feedback and interest — we’d love to build this with the doctors who’ll use it.

Or write to [email protected]

  • Free at launch
  • iOS & Android
  • Doctor stays in control