New Digital Tools and AI

All software must comply with GDPR. A Data Protection Impact Assessment (DPIA) must be completed by your ICB Data Protection Officer (DPO) and signed off by the practice before implementation.

Clinical safety is governed by two standards - DCB0129 and DCB0160. DCB0129 applies to the software developer and sets out the clinical risks associated with a system and the safeguards in place. DCB0160 is completed by the practice deploying the product, and must be reviewed, localised, and signed off by the practice Clinical Safety Officer (CSO).

You can find out more about CSO training here.

The MHRA (Medicines and Healthcare Products Regulatory Agency) certifies medical devices. For most AI applications in primary care - where a member of staff reviews the output and the system is not acting autonomously - a Class 1 medical device classification is likely to be sufficient. Where higher-level or autonomous functions are involved, a higher classification may be required. This remains a developing area, and a cautious approach is recommended.

Generative AI tools such as Microsoft Copilot and ChatGPT can support a range of administrative tasks in primary care.

For guidance on using these tools appropriately, see: Generative AI in Primary Care - Considerations document.

AI scribes are currently the most widely used AI tool in primary care, partly due to their integration with existing clinical systems and the availability of free browser-based options. They aim to produce fuller, more accurate clinical notes while saving clinician time.

Key points to be aware of:

• AI scribes must be registered as a medical device.

• All entries must be carefully checked by the clinician for accuracy before being saved to the patient record.

• The scribe should be linked directly to EMIS to avoid the risks associated with copying and pasting into the wrong record.

• Most AI scribes are Class 1 medical devices and are not able to make diagnoses, so they will not code a problem title. They can capture a limited set of codes, typically a subset of patient observations.

• Patients must be informed about and consent to the use of AI scribes. Privacy notices should be displayed on the practice website and in waiting areas.

A comparison of the main AI scribes currently in use across NCL is available here.

Robotic Process Automation (RPA) can automate the processing of new GP registrations - a task that is often time-consuming for administrative staff. In most cases, patients will be registered instantly, reducing delays and freeing up staff time.

Automated recall systems can take much of the administrative burden out of managing patient recall. Most are designed to be set up once and then run in the background with minimal ongoing input.

When choosing a system, consider:

• Which contact methods it uses to reach patients (for example, SMS, letter, phone).

• Whether it can automatically gather additional information from patients, such as blood pressure readings or AST results.

• Whether it is compatible with your LTC LCS as well as QOF - not all systems support both.

AI document management tools aim to make handling incoming correspondence more efficient. Features typically include automatic coding of patient demographics and letter types, short summaries of letters for the clinical record, and highlighting of relevant codes and actions.

In their current form, these tools should be seen as supporting coders rather than replacing them.

AI voice agents can supplement your practice's telephone capacity by handling calls and completing online consultation requests on a patient's behalf. This can be particularly valuable for patients who face barriers to using digital services, or who do not speak English as a first language (many voice agents offer multilingual support).

Case studies

Robot Process Automation (RPA) “bots” can automatically file both normal and some abnormal pathology results directly into EMIS, running multiple times throughout the day. They can also notify patients of their results where this option is available.

Even with these tools in place, clinicians will still need to review abnormal and more complex results. Practices should consider whether the reduction in administrative workload justifies the cost.

AI triage tools have the potential to save significant time for practices operating clinician-led total triage models by helping to process incoming requests more quickly.

However, there are important limitations to be aware of:

• Current solutions do not have access to patient records, which makes them better suited to acute and new problems rather than follow-up or ongoing issues.

• Not all triage requests may be handled automatically, so practices need a clear plan for managing those that fall outside the system's scope.

• Smaller practices that may struggle to resource a traditional total triage model could find these tools particularly useful.

Any practice considering AI triage should carefully assess the clinical risks involved, complete the DCB0160 process thoroughly, and confirm that the solution holds appropriate medical device registration.