Healthcare · · 8 min read
How to evaluate AI patient support agents
Acumen Editorial · Updated November 22, 2026
The patient support line is the next place AI agents are going to work. In one week at HLTH 2026, three launches made that plain.
On November 16, Sempre Health launched Atlas, an AI voice service for life science hubs, payers and PBMs. It handles enrollment, benefits and prior authorization calls, refills and adherence outreach. On November 17, The Rx Assistant unveiled Builder, a no-code platform for pharma brands to deploy AI experiences for patients and HCPs across voice, SMS, email and chat. On November 18, Cognizant introduced Patient Hub Operations, which pairs AI agents with specialists who review, approve or escalate the actions the agents prepare.
Patients are already there. Healthcare Dive, citing KFF, reports that roughly one-third of U.S. adults have used an AI chatbot for health information. Apple and OpenEvidence just opened a study of conversational health AI for up to 200,000 participants, focused on trust, comprehension and response quality.
For patient services teams, the question is no longer whether to use agents. It is how to prove they are safe before they talk to your patients.
What do AI patient support agents actually do?
They take the conversations that sit between a prescription and the first dose. Based on the launches above, that includes:
- Access and affordability. Benefits checks, prior authorization and appeals calls, copay and financial assistance questions.
- Logistics. Enrollment, pharmacy issues, shipping and refills.
- Adherence. Outreach and reminders after therapy starts.
- Brand questions. Product and condition questions, answered from approved content.
- Handoff. Transfer to a human with context when the question needs one.
The vendors describe guardrails. Sempre says every call is recorded, transcribed and reviewed, and that it captures and reports adverse events and product quality complaints. The Rx Assistant says it automates adverse event logging and regulatory review workflows. Cognizant says specialists keep final authority over regulated decisions. These are vendor descriptions. Your evaluation is what turns them into evidence.
Why is patient support harder to test than a website chatbot?
Because the stakes are regulatory and clinical, not just commercial.
Every conversation is a safety data source. Under 21 CFR 314.80, applicants must promptly review adverse drug experience information from any source. Serious and unexpected events go to FDA as 15-day Alert reports, no later than 15 calendar days from initial receipt. In the EU, GVP Module VI treats reports from organized data collection, including patient support programs, as solicited reports. The MHRA Inspectorate expects contracts with third parties to cover collecting and forwarding safety information, plus training, reconciliation and audits. An agent that misses an adverse event is a compliance gap, not a quality issue.
Patients do not speak in clean categories. A patient will not say "I am reporting an adverse event." They will say they have felt dizzy since the new box arrived, while asking about a refill. The agent has to answer the refill question and catch the safety signal.
The right answer is often a handoff. Dosing changes, worsening symptoms, pregnancy and off-label questions belong with a pharmacist, a nurse or the patient's own doctor. Knowing when to stop is part of the job.
Plain language decides whether the answer works. A correct answer about a prior authorization denial that a caregiver cannot follow still fails.
What should you measure?
Score every test conversation on the same rubric. These are the dimensions we use as a starting point.
| Dimension | What good looks like | Who should score it | Typical failure |
|---|---|---|---|
| Content accuracy | Answer matches the prescribing information and approved program content | Pharmacist or clinician in the therapy area | Correct-sounding detail that is not in approved content |
| Adverse event capture | Every possible AE or product complaint is flagged, with the minimum case details collected | Pharmacovigilance specialist | AE mentioned in passing, buried under a logistics question |
| Escalation and handoff | Clinical, off-label and urgent questions go to the right human with context | Clinician | Agent answers a dosing question it should have routed |
| Urgency recognition | Signs of an emergency trigger immediate direction to care | Clinician | Agent keeps troubleshooting while symptoms worsen |
| Plain-language fit | Patient or caregiver can explain the answer back | Patients and caregivers with the condition | Accurate answer the patient cannot act on |
| Tone and trust | Answer sounds like it understands life with the condition | Patients and caregivers | Scripted empathy that reads as dismissive |
| Access accuracy | Coverage, assistance and next steps are correct for the program rules | Hub or access specialist | Wrong eligibility guidance or missed next step |
| Record quality | Call summary and handoff notes are complete and correct | Safety or hub specialist | Summary drops the detail the human needed |
Report results by severity, not just as an average. One missed serious adverse event matters more than a hundred slightly awkward greetings.
How do you build a test set that reflects real patients?
Start from real conversations, not internal guesses.
- Collect real patient and caregiver input. Consented, de-identified questions in the patients' own words, across the stages of the journey: diagnosis, access, first dose, maintenance and switching.
- Seed the hard cases. Clinicians write conversations with embedded adverse events, product complaints, pregnancy mentions, off-label questions and emergencies. Vary how directly they are stated.
- Cover who actually calls. Caregivers, low health literacy, other languages, rare disease patients and callers who are angry or confused. Voice agents need spoken test cases, not just typed ones.
- Write reference answers and rubrics. For each item, a clinician states what a safe answer must include and must avoid, and whether a handoff is required.
- Score with the right people. Clinicians and safety specialists score accuracy, safety and escalation. Patients and caregivers score clarity and trust.
- Retest on every change. New content, a new channel or a new model version means a new run on the same set.
Simulated patients help with volume, but they should not be your only test. As we covered in how to validate AI-simulated patients, simulated patients tend to answer like an average patient, which is exactly where the hard cases are not. For the safety side, see how to evaluate AI for pharmacovigilance.
Checklist: what to ask a patient support AI vendor
- What content can the agent answer from, and who approves changes to it?
- How does the agent detect adverse events and product complaints, and what is its measured miss rate on seeded cases?
- What minimum case details does it collect, and how fast does it forward them to our safety team?
- Which topics trigger a human handoff, and what context travels with the transfer?
- How are emergencies handled on voice, SMS and chat?
- Who reviews recorded conversations, how many and how often?
- Was the agent tested with real patients and caregivers in our therapy area, or only with staff and simulated patients?
- Do you reconcile safety data with our pharmacovigilance system, and how often?
- What gets retested when content, channel or model changes?
- Can you show results by severity on a test set we provide?
How Acumen helps
We give pharma and digital health teams the human side of the evaluation. That means consented, de-identified patient and caregiver input, matched by condition and journey stage, plus clinicians and safety specialists who write hard cases, reference answers and rubrics and score the agent's conversations. You get item-level scores, failure examples by severity and documentation you can bring to safety, medical and compliance review.
See our patient experience data for AI service for what a test set and scoring program include.
FAQ
What is an AI patient support agent?
An AI patient support agent is software that handles patient and caregiver conversations for a pharma brand, hub, payer or pharmacy. It answers access, affordability, enrollment, refill and adherence questions by voice, text or chat, and hands complex questions to human staff. Launches in November 2026 from Sempre Health, The Rx Assistant and Cognizant show the category moving into production.
How do you test an AI patient support agent before launch?
Build a test set from consented, de-identified patient and caregiver questions, then add clinician-written hard cases with embedded adverse events, off-label questions and emergencies. Score each conversation on a fixed rubric for accuracy, safety, escalation and plain language. Clinicians and safety specialists score safety; patients and caregivers score clarity. Retest after every change.
Do AI patient support agents have to report adverse events?
The company does. Under 21 CFR 314.80, applicants must review adverse drug experience information from any source, and serious, unexpected events go to FDA within 15 calendar days of initial receipt. In the EU, GVP Module VI treats reports from patient support programs as solicited. An agent is part of that intake, so its capture rate needs testing.
Can simulated patients replace real patients in testing?
No. Simulated patients add volume and help rehearse common flows, but they tend to answer like an average patient. The cases that break patient support agents are the unusual ones: a side effect mentioned in passing, a confused caregiver, a rare disease question. Use real patient and caregiver input for the core test set and simulated patients to extend it.
Who should score a patient support agent's answers?
Match the scorer to the question. Pharmacists or clinicians in the therapy area score content accuracy and escalation. Pharmacovigilance specialists score adverse event capture and case details. Hub specialists score access guidance. Patients and caregivers with the condition score whether the answer was clear, usable and trustworthy. One reviewer type cannot cover all of these.
How often should a patient support agent be re-evaluated?
Re-evaluate on every material change: new approved content, a new indication, a new channel such as voice, or a new model version. Between changes, review a sample of live conversations on a set schedule and feed new failure types back into the test set. The MHRA expects periodic safety data reconciliation for patient support programs, which fits the same rhythm.