§ 01 Abstract
A completed call is not a resolved concern.
Objective. Make signs of dissatisfaction visible while the clinic still has an opportunity to respond, and distinguish friction in the interaction from distress about the underlying situation.
Methods. We analysed sentiment and its apparent drivers in a 60-day production audit spanning medical and veterinary practices. We also implemented a staff review workflow that combined concerning mood readings with unfinished work and conversation history.
Findings. The model attributed 90.6% of frustrated calls to drivers associated with the assistant interaction. Anxiety and upset were predominantly attributed to other causes. This distinction made a negative sentiment label more useful: it distinguished concerns requiring different kinds of response.
Interpretation. Sentiment became a prompt to investigate and help. The work established visibility and a route to action; it did not measure reductions in complaints, negative reviews or patient loss.
§ 02 Background
The next complaint can start with unfinished business.
A call can be answered, routed and marked complete while the person on the other end remains confused or frustrated. The operational record says the interaction ended. It may say very little about whether the caller understood the next step or felt heard.
For a practice, that gap matters. If the concern resurfaces at the next appointment or in a public review, the team is responding after frustration has had time to accumulate. Earlier visibility creates an opportunity to clarify a misunderstanding, acknowledge a poor interaction or get an unresolved request to the right person.
We focused on a narrower, answerable question: could the clinic see the concern—and its likely cause—without listening to every call again?
§ 03 Methodology
We recorded the feeling, the cause and the unfinished work.
Our call-analysis pipeline extracted expressed mood, intensity, apparent drivers and perceived resolution from eligible completed calls. These readings were retained across interactions so staff could review a changing experience rather than a single isolated label.
The retrospective audit covered 21 July–19 September 2026: 8,920 scored calls at a veterinary practice and 613 at a medical practice. These were call records, not unique patients. The combined sample informed how we interpreted sentiment.
We separated drivers associated with the assistant interaction, such as confusion and waiting, from other drivers, including availability, cost, health concerns and prior experience. This was model-based attribution, not a determination of fault. A concern outside the assistant’s control could still require the clinic’s attention.
§ 04 Results
“Negative” concealed two very different problems.
Frustration and anxiety did not tell the same story. Among calls classified as frustrated, the model predominantly identified drivers associated with the assistant interaction. For anxious and upset calls, it predominantly identified causes outside that interaction.
| Expressed mood | Calls in category | Predominant attribution |
|---|---|---|
| Frustrated | 736 | 90.6% associated with the assistant interaction |
| Anxious | 949 | 96.8% associated with other drivers |
| Upset | 87 | 98.9% associated with other drivers |
We brought concerning mood readings into the portal’s Needs care view alongside unanswered messages and unresolved requests. Staff could see a short explanation, open the patient record and conversation history, reply, assign responsibility, or record the next step. A mood timeline preserved the context across calls.
The concrete change was visibility: an ended call with a concerning reading could still surface for attention. Staff no longer had to treat “call complete” as the last word on the experience.
§ 05 Discussion
The useful question is: what would help this person now?
Confusion about the next step calls for a clear explanation. Frustration about an unresolved request calls for someone to own it. Anxiety about a health concern may require reassurance or appropriate clinical review. Treating all three as the same negative score loses the information the team needs to respond well.
That is the practical value of proactive follow-up. A clinic can reach out with context, address what remains unresolved and correct misunderstandings before the next interaction begins on the same footing. Protecting trust starts with resolving the concern; a request for a better review does not do that work.
The score is a starting point for human review. Staff still need to read the interaction and decide whether contact is useful, who should handle it and what a meaningful resolution would look like. The patient’s account matters more than the label.
§ 06 Limitations
Visibility is the finding. Recovery is the next measure.
This was an internal operational audit and implementation study, not a controlled clinical trial. Sentiment and driver labels were model-derived and were not independently validated against a clinician-labelled reference set. The two practice settings and their unequal call volumes limit generalisation.
We did not establish how often staff completed proactive outreach or whether it prevented escalation, negative reviews or patient loss. A later calmer conversation alone would not prove recovery: it must be connected to a documented intervening action and a resolved concern. Sentiment scoring is not a clinical triage or diagnostic tool.
§ 07 Conclusion
Give the clinic a chance to respond before the concern returns.
We made concerning interactions visible beyond the call log and placed them beside the work still waiting for attention. The value was in connecting a feeling to its context and a possible next action. For a busy practice, that creates a more useful starting point than learning about an unresolved experience only when the person calls back angry.