The one rule everything else follows

Rules you configure decide. AI only handles language.

The platform never lets a model decide policy, eligibility or clinical priority. Those decisions come from structured record data and rules a named person at the practice has set and can change at any time.

What the AI does

• · Writes the wording of an outreach message within an approved template and tone.
• · Reads a free-text reply and works out what the patient meant — confirm, cancel, reschedule, question, opt out, or unclear.
• · Flags anything it is not confident about so a person sees it.
• · Summarises a conversation thread for the staff member picking it up.

What the AI never does

• · Decide who is eligible to be contacted, or how many times.
• · Give clinical, diagnostic, medication or triage advice.
• · Override quiet hours, contact preferences or an opt-out.
• · Act on a reply it has read with low confidence — those go to a person instead.
• · Change a rule, a threshold or a template on its own.

A person stays in the loop

Every interpretation carries a confidence score. Below the configured floor, the work item is escalated to the “needs a person” desk with the full conversation attached. Sensitive categories — anything that reads as clinical urgency, distress or a complaint — are escalated regardless of confidence.

Transparency to patients

Where local rules require it, outreach messages state that they are automated and tell the patient how to reach a human or stop messages. Patients are never told they are talking to a clinician when they are not.

Data used with AI models

• · Only the minimum a conversation needs: first name, appointment or medication context, and the message thread.
• · No clinical notes are sent to a model.
• · We use providers under agreement that they do not train their models on your data.
• · Model requests are logged for audit, without unnecessary personal detail.

Accuracy, bias and monitoring

Language models can misread a message. We monitor interpretation accuracy, escalation rate and opt-out rate, review a sample of conversations, and tune templates and confidence thresholds. If accuracy drops, the safe fallback is always to escalate more to people, not to act automatically.

We check that outreach and interpretation behave consistently across patient groups, and treat any pattern of unequal treatment as a defect to fix.

Human override and correction

Staff can stop outreach, correct an interpretation, undo a written-back action within your record system's own rules, and suppress future contact for a patient permanently. Every override is recorded with the person, time and reason.

Alignment with AI regulation

We design against emerging AI governance expectations — the EU AI Act's transparency and human-oversight duties among them — by keeping decisions rule-based, keeping a person accountable for exceptions, logging automated actions, and documenting the intended use and known limits of every AI feature.