Key takeaways
- Decide the access level first: no connection, read-only or write. Each step up needs more assurance.
- Use the least access that does the job, through an approved supplier interface, never a shared login or smartcard.
- Insist on an audit trail that shows every read and write, by which system, for which patient and why.
- Write access needs a DPIA, a clinical safety case, a processor contract and testing before go-live.
- Most practices get value from a standalone tool first, then add read-only access once the call patterns are clear.
Why integration is the question everyone asks
Once a practice has an AI tool answering calls, the obvious next step is to let it look at the appointment book, book a slot or add a task for the team. The pressure is understandable. Online requests to GP practices in England rose from 5,690,082 in August 2025 to 9,617,628 in March 2026, after practices had to keep online tools open throughout core hours, reaching 153 requests per 1,000 patients a month[1]. Every request that has to be re-keyed by hand costs reception time.
Source: Care Quality Commission, 2026[1]
Show the data
| Item | Value |
|---|---|
| August 2025 | 5,690,082 |
| March 2026 | 9,617,628 |
But connecting any tool to EMIS Web, SystmOne or a dental practice system changes the risk profile. A standalone tool can say the wrong thing; a connected tool can do the wrong thing to a patient's record. This guide sets out the levels of connection, the controls each one needs and the shortcuts to refuse.
Three levels of connection
| Level | What the AI can do | Patient data involved | Assurance before go-live |
|---|---|---|---|
| Standalone | Answer practice information, explain routes, pass requests to reception | Caller's name and number at most | DPIA, processor contract, privacy notice update |
| Read-only | Check appointment availability or confirm a patient's existing booking | Identity match and appointment details | All of the above plus identity checks, access scoping and audit logs |
| Write | Book, cancel or move appointments, create tasks or add entries | Record changes linked to a named patient | All of the above plus a clinical safety case, testing in a test environment and named sign-off |
Start with the least access that does the job
Health and care organisations in England work to eight Caldicott Principles[2], and two are directly relevant here: use the minimum necessary confidential information, and access it only on a strict need-to-know basis. For call handling that translates into a simple rule. If the task can be done without opening the record, do not open the record.
In practice many of the calls that feel like they need integration do not. Opening hours, how to order repeat medicines, how to register, where to park and whether the practice is taking new NHS dental patients are all answered from approved practice information. Requests about results, referrals and medication can be passed to reception with a short reason, which is safer than an AI tool searching a record on the caller's say-so.
Data concerning health is special category data under UK GDPR Article 9[3]. Every extra field the tool can see is a field you must justify in your DPIA and protect in your contract.
Questions to ask about the interface
- Which interface does the product use with our system, and is it an approved supplier route rather than screen automation?
- Exactly which permissions does the connection hold: read appointments, write appointments, read demographics, create tasks, anything else?
- How does the tool confirm it has the right patient before reading or writing anything, and what happens when two patients share a name and date of birth?
- Can we restrict the connection to certain appointment types, clinicians or sessions?
- What does the tool do if the clinical system is slow or unavailable?
- How quickly can we switch the connection off, and who at the practice can do it?
Identity checks that work on the phone
Any connection that reads or changes a record depends on matching the caller to the right patient. Receptionists do this instinctively, noticing when a voice does not fit the record or a parent is calling about an adult child. An automated tool needs explicit rules instead, and the rules should be stricter than the ones your team uses, because the tool cannot notice when something feels wrong.
- Use at least two identifiers, such as full name and date of birth, plus a check that does not rely on information a relative or partner would know, such as a one-time code sent to the mobile number on the record.
- Treat a caller ringing from a number that is not on the record as unverified, and limit them to general information.
- Never read out anything from the record to confirm identity; ask the caller to provide it.
- Route third-party callers, such as carers, parents of over-16s and care home staff, to reception unless consent is recorded.
- Fail safe: if any check fails or is ambiguous, the tool passes the call to a person and reads or writes nothing.
Test the rules with awkward fictional cases before go-live: twins, a patient who has changed their surname, two patients with the same name and date of birth, and a caller who gives the right details from a withheld number. Record the results in your clinical safety case.
The audit trail you should insist on
If something goes wrong you will need to reconstruct exactly what happened. Agree the audit trail before go-live and test that you can actually retrieve it.
Every action logged
Each read and write is recorded with the date, time, patient, the action taken and the call that triggered it.
Clear attribution
Actions by the AI tool appear under a distinct system user, never under a member of staff's name.
Call linked to record
The call transcript or summary can be found from the record change, and the record change from the call.
Retention agreed
Logs are kept for a period you set and can be exported in a readable format.
Regular review
A named person checks a sample of AI actions every week during the first months, then monthly.
Shortcuts to refuse
- Shared logins or smartcards used by a supplier or a bot to reach the clinical system.
- Screen automation that types into the system as if it were a person, with no supplier interface and no distinct audit identity.
- Free-text entries written into the clinical record by the AI tool without human review.
- Automatic filing of documents or coding of entries based on a phone conversation.
- Broad permissions granted 'for later' that the current task does not need.
Security is not theoretical. In the 2025 Cyber Security Breaches Survey, 43% of UK businesses reported a breach or attack in the previous 12 months, and phishing was the most common type, affecting 85% of those that were breached[4]. A connection that relies on a shared credential multiplies the damage one compromised password can do.
Building the assurance pack
Write access in particular is a clinical change, not just an IT one. The MHRA says many software and AI products used in health and care are regulated as medical devices[5], so start by asking the supplier for its regulatory position. Then assemble the pack below and keep it with your DPIA.
- DPIA covering data flows, lawful basis, Article 9 condition, risks and controls.
- A processor contract that covers instructions, confidentiality, security, sub-processors and deletion[6].
- The supplier's clinical safety case and hazard log, plus your own deployment safety case signed by a clinical safety officer.
- Evidence of testing in a test environment with fictional patients, including mismatched identities and system outages.
- A go-live checklist with a named person who can pause the connection immediately.
Ava follows this order itself: appointment requests in EMIS Web or SystmOne are labelled Planned, not live, because they need clinical safety assurance, a DPIA and a data processing agreement before they can be offered.
Example: moving from standalone to read-only
Example (fictional): a GP practice of 14,000 patients runs a standalone AI phone tool for three months before considering integration.
In this example the call records show that the most common request that the tool cannot finish is 'Can you confirm the time of my appointment?' The practice agrees a read-only connection limited to future appointments, with identity confirmed by date of birth and a code sent by text to the number on record. Write access is deferred. After six weeks of weekly audit checks with no wrong-patient events, the clinical safety officer signs off continued use, and the practice starts a separate assessment for cancellations only.
Where national policy is heading
The government's 10 Year Health Plan commits to rolling out AI scribes, digital triage and the single patient record within two years[7]. That direction will bring more connected tools into practices, and more suppliers offering integration as a headline feature. The principles above will not date: least access, approved interfaces, a clear audit trail and clinical sign-off before any tool writes to a record.
Frequently asked questions
Can an AI phone tool book into EMIS Web or SystmOne?
Some products offer it through supplier interfaces, but it should only be switched on after a DPIA, a clinical safety case, a processor contract and testing with fictional patients.
Is read-only access safer than write access?
It carries less risk because the tool cannot change the record, but it still exposes patient data, so identity checks, scoped permissions and audit logs are still needed.
Why not let the AI use a spare login?
Shared credentials break the audit trail, breach smartcard and access rules and make any security incident much harder to contain.
Do we need integration to get value from AI call handling?
Often not at first. Many calls are answered from approved practice information or passed to reception, so a standalone tool can show where integration would genuinely help.
References
Care Quality Commission (2026). State of Care 2025/26: capacity in primary care (opens in a new tab) ↑
GOV.UK (National Data Guardian) (2026). The Caldicott Principles (opens in a new tab) ↑
legislation.gov.uk (The National Archives) (2026). UK GDPR Article 9: Processing of special categories of personal data (opens in a new tab) ↑
GOV.UK (Department for Science, Innovation and Technology) (2025). Cyber security breaches survey 2025 (opens in a new tab) ↑
GOV.UK (Medicines and Healthcare products Regulatory Agency) (2025). Software and artificial intelligence (AI) as a medical device (opens in a new tab) ↑
legislation.gov.uk (The National Archives) (2016). UK GDPR Article 28: Processor (opens in a new tab) ↑
GOV.UK (Department of Health and Social Care) (2025). Fit for the future: 10 Year Health Plan for England (opens in a new tab) ↑