Adverse Event Capture on Medicines Information Lines

A caller rings a medicines information line to ask whether a tablet can be split. Halfway through, she mentions that her ankles have swollen since she started it. Under UK law that second sentence matters more than the first. The Human Medicines Regulations 2012, regulation 187, require a marketing authorisation holder to record every suspected adverse reaction to its product that is brought to its attention, whether a patient or a health care professional reports it. The caller did not ring to report anything. The duty does not care why she rang.

That is the operational problem this guide addresses for the Head of Customer Operations who runs a pharma company's medical information and patient support lines, and for the qualified person for pharmacovigilance who answers for them at inspection. Every inbound call is a potential adverse event intake, whether a human or a voice AI answers it. The line has to recognise a suspected reaction inside an unrelated conversation, capture enough to make a valid case, hand over to a trained person and never drift into medical advice.

Scope, stated plainly. This post covers the call journey: recognition, capture, routing, handover and the audit trail. It cedes the 15-day and 90-day submission clocks and the case pipeline to our map of where AI pays in UK pharma, and it cedes turning a transcript into a coded case draft to our clinical document extraction guide. The general pattern of spotting a regulated event mid-call in financial services is covered in our voice AI complaints handling guide; the pharmacovigilance mechanics here are different enough to need their own treatment.

This guide is shipped by the team behind Dilr Voice, a multi-agent voice AI platform that keeps a full audit trail on every call and warm-transfers to a human with full context. Or see DATS, our five-stage system for placing AI inside regulated operations.

What does regulation 187 require of a medicines information phone line?

Regulation 187 of the Human Medicines Regulations 2012 requires the marketing authorisation holder to record every suspected adverse reaction brought to its attention, from patients or health care professionals, in the UK or abroad. It must not refuse reports received by any appropriate means, and must keep them accessible at a single UK point. On our reading a phone line is an appropriate means, so every call can carry the duty.

The regulation has four paragraphs, and each one lands on the phone line in a different way. Paragraph (1) sets the recording duty. Paragraph (2) carves out reactions that occur inside a clinical trial, which sit under the Clinical Trials Regulations instead. Paragraph (4) requires the recorded reports to be accessible, electronically or physically, at a single point within the United Kingdom. Paragraph (3) is the one that most directly shapes call design:

"The holder must not refuse to consider reports of suspected adverse reactions to the product received electronically or by any other appropriate means from patients or from health care professionals." Human Medicines Regulations 2012, regulation 187(3)

Read against a contact centre, that sentence makes two common shortcuts hard to defend. A line cannot tell a caller to report side effects somewhere else and end the call, and an automated journey cannot route a reaction mention to a dead end because it fell outside the menu the caller chose. The duty binds the holder, not the agency or the software vendor running the line, which is why, in our view, the holder's pharmacovigilance system has to reach into every channel it owns. Our UK and EU voice AI compliance guide sets out the wider data protection and disclosure layer that sits on top of this.

What counts as a suspected adverse reaction on a call?

A suspected adverse reaction on a medicines information call is any spontaneous mention that something unwanted happened to a person using the product, unless the reporter states it was unrelated. The caller does not need to say side effect, know the cause or be the patient. Lack of effect, misuse, medication errors and exposure in pregnancy are also handled as special situations in the guidance.

The working definitions come from the European Medicines Agency's GVP Module VI, which still applies in the UK through regulation 205B of the 2012 Regulations, read alongside the MHRA's note on exceptions and modifications to EU GVP, last updated on 19 February 2025. That note says EU GVP continues to broadly apply to UK authorised products, and its Module VI section does not list a UK change to the validation, day zero or patient support programme passages this guide relies on.

Module VI is explicit that a spontaneously reported event meets the definition of an adverse reaction even when the relationship is unknown or unstated, unless the reporter says it was unrelated.

For call design, the practical consequence is that recognition has to be broad and the decision narrow. The agent, human or AI, should flag any plausible mention and pass it on. It should not decide that a mention is too vague, too indirect or obviously unrelated. Module VI says medical judgement should always be applied in deciding whether information is an adverse reaction or an event, and that where the primary source explicitly excludes a causal link and the holder agrees, the report is not a valid case. That agreement is a pharmacovigilance decision, made by the holder's trained staff, not a call-routing rule. In our view a voice agent should be tuned for recall over precision on this one intent, with the QA testing framework built around missed mentions rather than false alarms.

What are the four minimum criteria a call must capture?

The four minimum criteria for a valid individual case safety report, set out in GVP Module VI, are one identifiable reporter, one identifiable patient, at least one suspected medicinal product and at least one suspected adverse reaction. A medicines information call that captures all four from a first-hand reporter creates a valid case. A call that misses one creates an incomplete report the holder must still record and follow up.

Each criterion has detail that a scripted or AI-led journey can get wrong:

Two details matter for a voice line. First, Module VI treats a report as valid even when the reporter declines to give contact details, as long as the holder can confirm the case directly with that reporter, so a refusal is not a reason to drop the case. Second, second-hand or hearsay information is non-valid until it is verified with the patient, the patient's health care professional or someone with direct contact. A caller passing on a neighbour's story is still recorded as an incomplete report, but it stays non-valid until verified, so the follow-up path differs.

The capture step should feel like a short conversation, not a form. Our warm transfer guide covers how to carry what was captured into the human handover so the caller is not asked twice.

When does the reporting clock start on a phone call?

The reporting clock for a valid case starts on day zero, which GVP Module VI defines as the day information containing the four minimum criteria first reaches any personnel of the marketing authorisation holder, including medical representatives and contractors. Weekends and public holidays count. On our reading, when a contractor runs the medicines information line, day zero can be the day of the call itself.

The day zero rule is the reason intake design matters so much. The 15-day and 90-day submission windows, which the pharma map linked above explains in full, run from the moment the holder's organisation knows, not from the moment the pharmacovigilance team opens the record. A suspected reaction reported at 6pm on a Friday into an outsourced line, with all four minimum criteria captured, has on that reading already started its clock.

Two points need care. On our reading, an AI system is not itself personnel. The relevant question is whether the organisation operating the line, the holder itself or its contracted agency, has received the information. Module VI also expects explicit procedures and detailed agreements between a holder and any contracted organisation, so that the holder can meet its submission duties. That agreement is where the holder should set how fast the line passes a suspected reaction to the safety team, and how the two sides reconcile what was passed.

For a useful industry analogue, the ABPI and BHBIA guidance for market research (version 5.0, February 2021) expects agencies to report adverse events, product complaints and special reporting situations to the holder within one business day of becoming aware, and to keep records for reconciliation. That guidance governs market research agencies rather than medicines information lines, but it shows the kind of turnaround industry guidance already sets for one class of third party.

Is a patient support programme call treated differently from a medical information call?

Yes. GVP Module VI treats a suspected reaction reported through a medical enquiry or product information service as an unsolicited, spontaneous report. A reaction surfacing on a patient support programme or market research call counts as a solicited report, which needs a causality assessment. A reaction inside a clinical trial falls outside regulation 187(1) altogether and follows the trial regime.

That distinction decides how a single phone platform should label and route the same words. One company can run several kinds of line on one number range, which is where mislabelling starts.

For solicited reports, Module VI asks the holder to seek the primary source's view on whether the product caused the event, and to apply its own judgement where that view is missing. A voice agent can ask the caller's view in plain words and record the answer. It must not offer its own. The same platform that runs the adherence reminders many patient programmes depend on can carry the programme tag on every call record, which is what lets the safety team classify the report correctly later.

The same diagnostic logic underpins our AI operating model consulting, which decides who owns each channel, each tag and each handover before any agent goes live.

How should a voice AI line hand over a suspected reaction?

A voice AI line should hand over a suspected reaction by finishing the minimum capture, telling the caller plainly that a trained colleague will follow up, and passing the record and the caller's own words to the holder's safety intake team with full context. It should give no medical advice, and should direct anyone worried about their health to a doctor, pharmacist or NHS 111.

The no-advice boundary has a useful model in the regulator's own site. The MHRA's Yellow Card site tells worried visitors to speak to a doctor or pharmacist or call NHS 111, and states that it cannot provide medical advice. A medicines information line staffed by trained medical information professionals works within approved content, and an automated journey should stay inside approved content and escalate everything else. The Yellow Card site also notes that anyone can report an issue with a medicine, vaccine or medical device, including software, apps and artificial intelligence, so a holder may mention the scheme, but in our reading a referral to it does not replace the holder's own regulation 187 duty to record the report.

A sound handover design has five steps, and the order matters:

Three behaviours make the handover safe. The agent confirms what it heard back to the caller in plain words. It does not ask the caller to rate seriousness, because in the design we recommend seriousness is judged by the holder's safety team. It never closes a call with a suspected reaction unrecorded, even when the caller says it does not matter. Our escalation and human handover guide covers trigger design in more depth, and our enterprise voice AI guide covers the architecture behind a multi-agent line.

Dilr Voice fits this pattern as a voice layer rather than a pharmacovigilance system: it chains specialised agents into one call, cites answers from the customer's own data through RAG knowledge bases, warm-transfers to a human with full context, and keeps a full audit trail on every call. The safety database, the case assessment and the submission stay with the holder.

What should the audit trail show an inspector?

The audit trail for a medicines information line should show, for every call, that a suspected reaction was recognised, what the caller said in their own words, which minimum criteria were captured, when the record reached the safety team and how it was reconciled. GVP Module VI asks for the verbatim text or an accurate translation, and for storage that traces every entry and change.

The verbatim point is easy to miss. Module VI says data from the primary source should be treated in an unbiased and unfiltered way, with inferences avoided, and that reports should include the verbatim text the source used. A summary written by an agent, human or AI, is not a substitute for what the caller said. Where calls arrive in other languages, the guidance accepts an accurate translation of the caller's words.

Reconciliation is the second pillar. The line's own log and the safety database should agree on every suspected reaction passed across. When they do not, the gap is a weakness in the pharmacovigilance system itself, and the MHRA's good pharmacovigilance practice inspections examine that system as a whole, not only individual cases. The holder's QPPV and pharmacovigilance system master file guidance from the MHRA is where the line's role in that system should be described.

Retention needs the same discipline as capture. Call recordings and transcripts that carry health data sit under UK GDPR as well as pharmacovigilance rules, and our voice AI data retention guide sets out how to keep what the safety system needs without keeping everything forever. For the evidence side of automated decisions, see our voice AI auditability guide.

What is the best way to run adverse event intake on a voice AI line in 2026?

The best way to run adverse event intake on a voice AI line in 2026 is a hybrid: the AI answers, recognises any plausible reaction mention, captures the four minimum criteria and the caller's words, then hands over to trained medical information or pharmacovigilance staff. In our view, a fully human line remains the stronger choice for low call volumes or complex specialist products.

The criteria for judging any platform, Dilr Voice included, are concrete. Does it recognise a reaction mention inside an unrelated intent? Does it capture the four criteria without turning the call into an interrogation? Can it tag the line type on every record? Does it warm-transfer with the captured context? Does the audit trail hold the caller's own words and every change?

On those criteria, enterprise contact centre suites such as Genesys and Five9 suit a holder whose medical information team already runs on them and wants AI added inside an existing human operation. Conversational specialists such as PolyAI focus on enterprise customer service voice assistants. Developer platforms such as Vapi and Retell AI give an engineering team the most control, at the cost of building the pharmacovigilance logic themselves. A holder with a small specialist portfolio and a few calls a day may find a well-trained human team, with no AI at all, the better answer. Our comparison of AI voice agents for 2026 sets the wider market out.

Whichever route a holder takes, the pharmacovigilance design is the holder's to own. The AI placement diagnostic is how we test whether a medicines information line is the right first placement, and the AI execution office is how a placement is governed once live.

Can a voice AI agent decide that a call is not an adverse event?

A voice AI agent should not decide that a call is not an adverse event. GVP Module VI asks for medical judgement on whether information is an adverse reaction, and regulation 187 forbids refusing reports received by any appropriate means. The agent's job on a medicines information line is to flag any plausible mention, capture what it can and pass it to trained staff.

A false alarm costs a pharmacovigilance professional a few minutes. A missed report is a gap in the holder's legal record. That asymmetry is why we recommend tuning the recognition intent towards recall, and testing it with deliberately indirect phrasing such as a caller who blames the weather for a rash.

Does a vendor running the line carry the reporting duty?

No. The regulation 187 recording duty binds the marketing authorisation holder, not the contact centre agency or the voice AI vendor running its medicines information line. GVP Module VI expects the holder to have explicit procedures and detailed agreements with any contractor, covering how fast suspected reactions pass across and how both sides reconcile their records.

That allocation shapes procurement. A holder should ask any provider, Dilr Voice included, for the call record format, the handover mechanism and the audit trail, and write the transfer time and reconciliation cadence into the agreement. The obligation stays with the holder whichever provider answers the phone.

How the pieces fit

Adverse event capture is one placement among several in a pharma operation. The Cognibl, from DILR.AI operations desk pattern and the Dilr Mira extraction models are candidates for later stages of the same safety pipeline, and our DATS methodology decides which placement comes first. For the regulator's own view on AI in clinical voice settings, our post on the MHRA AI Airlock and ambient voice is the companion read, and every regulator-led guide we publish sits in the compliance category.

Want to see this in production? Try Dilr Voice live, read the best AI voice agent guide, see the Dilr Voice platform, or read about our approach to placing AI inside regulated systems.

Written by the Dilr.ai engineering team, practitioners who ship enterprise AI in production. Follow us on LinkedIn for shipping notes, or subscribe via the RSS feed.