An AI-native GP practice in 2030 could use connected software to support access, preparation, documentation and follow-up while clinicians retain responsibility for clinical judgement. The useful vision is not a doctor-free surgery. It is a practice in which information and unfinished work are easier to manage without making care less personal or accountable.
The journey below is fictional. It is a scenario informed by public technology descriptions checked on 15 October 2026, not a prediction that every practice will adopt it or a claim that any named vendor already supplies the complete workflow.
Begin with the patient's route into care
A fictional patient contacts the practice about a change in their ability to manage ordinary activities. They can use a digital form, telephone or an appropriate assisted route. The practice does not assume that everyone can describe symptoms confidently in writing.
Software could help organise the request, identify missing administrative information and prepare it for the appropriate team. The important safeguard is that an incomplete digital interaction must not become an invisible exclusion from care.
An AI-native practice should therefore be defined by the quality of its accessible pathways, not by how successfully it directs everyone into one interface. Patients who need another route should remain part of the service design rather than exceptions handled reluctantly.
Preparation should expose uncertainty, not settle the story
Before the appointment, an authorised system could assemble relevant records and show what has changed. It should distinguish the patient's current account from historical information and separate explicit negative findings from questions that have not been asked.
In the fictional case, a previous investigation appears reassuring, but it concerns a different episode. A useful summary makes that chronology visible. A poor summary places the result beside the new concern without qualification, encouraging the clinician to assume the issue has already been assessed.
This is why record access is only the beginning. The GP Connect service description, checked on 15 October 2026, illustrates that information-sharing functions and availability vary. It does not imply that an arbitrary AI tool has complete access to every patient's history.
During the consultation, the patient remains the primary encounter
A documentation assistant may reduce typing, but the clinician still needs to establish the patient's priorities, clarify contradictions and decide what assessment is necessary. The patient's explanation should not be subordinated to a polished pre-visit narrative.
The AI could make relevant information easier to inspect or prepare a draft record. Whether that improves the consultation depends on the interaction: does the clinician listen more effectively, or spend the encounter reconciling several competing panels?
The GMC's guidance on innovative technologies, checked on 15 October 2026, keeps professional judgement and discussion of uncertainties relevant. A futuristic interface does not remove the need for a patient to understand the proposed plan and available alternatives.
Separate what exists from what is imagined
| Capability | Position as checked on 15 October 2026 | What the 2030 scenario adds |
|---|---|---|
| Ambient documentation | Commercial products already describe consultation-note support | Better integration and locally demonstrated whole-workflow benefit |
| Record-connected AI review | Published organisational integrations exist, including read-only examples | More consistent handling of relevant context across the local pathway |
| Coding and draft workflow support | Vendor announcements describe specified functions in particular markets | Dependable local adoption with explicit approval and exception handling |
| Cross-service autonomous coordination | Not established as a universal GP capability | A speculative, bounded system with named ownership and verified completion |
| Independent learning linked to clinical questions | Reference and education tools are already available | A proposed, consented learning workflow without unnecessary patient-data transfer |
Examples include OpenAI's documented read-only Epic connection and Oracle's US August 2026 coding and chart-review release. These do not establish equivalent availability in EMIS, SystmOne or Medicus.
Prescribing and referrals need more than a completed draft
In the fictional practice, the clinician agrees a plan that creates further work. Software may prepare a letter, assemble the relevant information or create a review task. The appropriate level of automation depends on the action and the information available.
A draft is not an issued prescription. A sent referral is not an accepted referral. A reminder is not a completed review. The practice needs visible task states and a responsible team for anything unresolved.
The scenario should not rely on an agent independently deciding that the work is complete because the conversation ended. Confirmation must come from the relevant operational system, and changed circumstances may require the clinician to reconsider the original plan.
Follow-up is where the design is tested
Suppose the fictional patient replies that they cannot complete the agreed next step. A simplistic system records non-completion and sends another reminder. A better pathway preserves the reason and routes it to someone who can decide whether the plan needs changing.
This is a useful test of an AI-native practice because it distinguishes coordination from repetition. More reminders are not necessarily better follow-up. The service needs to understand whether the obstacle is administrative, practical or clinically significant.
The patient should also know whom to contact. A familiar automated interface is not a substitute for identifiable responsibility when something does not fit the expected pathway.
The core system remains a major operational choice
Doctolib's May 2026 Medicus announcement makes the future of core GP software a live competitive question. Its October AI ambitions do not make record migration, reliability and national-service connections less important.
A practice choosing between established and newer platforms needs to examine the complete operating environment. AI functionality should not distract from continuity during migration, access to historic information, staff training and the ability to recover from outages.
Nor should an organisation assume that all supplementary tools must come from the core-system supplier. A modular approach may be useful where interfaces, contracts and assurance permit it. The choice should be practical rather than ideological.
Capacity is a service decision, not an automatic dividend
The fictional practice might use saved administrative time for longer complex consultations, reduced overtime or additional access. Those are different outcomes. A system can also create more work by identifying unmet needs or generating additional review tasks.
Evaluation should therefore include patient understanding, unresolved work, staff experience and total effort across the team. Counting AI-generated notes or messages would not establish that the practice has become more effective.
The scenario becomes credible only when the practice can explain what improved, for whom and at what cost. A quieter desktop is useful, but it is not the same as a safer or more sustainable service.
Keep professional learning outside unnecessary patient-data transfer
A clinician can turn a general uncertainty from the fictional encounter into a learning question without uploading an identifiable record. As described for September 2026, Ask-iatroX, Tutor and CPD tools support reference and learning, not autonomous practice management.
The 2030 opportunity is a better connection between clinical questions and deliberate development. That connection should preserve independent understanding rather than merely file an AI-written reflection. The practice of the future still needs people who can recognise when its software has misunderstood the present.
Frequently asked questions
Is this a description of an existing fully AI-run GP practice?
No: it is a fictional 2030 scenario. The table separates published current functions from developing or speculative arrangements.
Will AI automatically increase GP appointment capacity?
No: the result depends on total workload, implementation and how any released time is used. New access can also reveal unmet need and generate appropriate additional work.
Would an AI-native practice still need independent clinical learning?
Yes: clinicians need the understanding required to question recommendations and respond to exceptions. Reference and learning tools do not replace supervision or clinical experience.
Explore clinical learning alongside a changing GP workflow →