What problem does it solve?
A patient who goes home after surgery or a hospital stay is at a moment of real risk and little contact. Hospitals follow up in different ways. Some units have a clinician phone each patient, while others discharge uncomplicated patients with advice on warning symptoms and ask them to contact the hospital if these occur. The authors of an Oxford and Buckinghamshire evaluation note that each of these pathways either needs trained clinical staff time or depends on the patient making contact, which may delay the finding of a complication, and that finding enough staff to make follow up calls is challenging.
Most patients recover without a problem, which is what makes the check repetitive. In that evaluation, 63% of the patients who answered the AI call had no clinical concern. The few who do need attention are found by asking the same short list of questions each time. A voice agent can ask them on a set schedule, try again when nobody answers, and give the clinician a written report. In the United States, the Medicare Hospital Readmissions Reduction Program adds a payment reason to stay in touch after discharge.
- CMS says the payment adjustment under its Hospital Readmissions Reduction Program ranges from 0.97, a maximum payment reduction of 3 percent, to 1.0, which means no reduction.Hospital Readmissions Reduction Program (2026)
- A Clinical Ophthalmology evaluation of an automated cataract follow up call cites an estimated 20 million cataract procedures performed worldwide each year, with volumes projected to rise as populations age.Evaluation of Routine Clinical Deployment of an Autonomous Artificial Intelligence Assistant for Cataract Follow-Up in the National Health Service (2026)
How does it work?
- The care team lists the patient. Each pathway has inclusion criteria, such as an adult with an uncomplicated procedure who can talk on the phone in a language the agent supports. At Oxford University Hospitals and Buckinghamshire Healthcare, a clinician requests the call on the surgery listing or the operation note, and patients with other eye conditions stay on the usual pathway.
- The agent calls at a set time after the procedure. The timing follows the clinical protocol. Dora calls three weeks after cataract surgery. LOLA called valve replacement patients at Dr Balmis General University Hospital at weeks 1, 2 and 4 and months 3 and 12. If nobody answers, Dora tries another time on any listed number, and Milton Keynes University Hospital texts patients the time of the call.
- Design the opening so the patient knows who is calling. We recommend that the agent names the hospital, says it is an automated assistant and checks that it is speaking to the patient before it asks about their health. The sources on this page do not describe this step, and the guardrails below require it.
- It asks the protocol questions. The questions cover symptoms, medication and discharge instructions, and the patient can ask questions of their own. Universal Health Services reviews medication instructions and asks about new or worsening symptoms. Dora also records a preference for second eye surgery.
- Rules classify the outcome. In the Oxford and Buckinghamshire evaluation, red or painful eyes, new floaters or flashing lights, vision concerns and questions the agent cannot answer give a "potential concerns" outcome, and everything else is "no concerns identified". Clinical teams set these rules.
- A report reaches a clinician. The clinician reads the summary, discharges the patient or lists them for the next step, and calls back patients with potential concerns, within 48 hours at Oxford and Buckinghamshire. At Summerlin Hospital the system pushes a message to the nurse when a patient wants a call back.
- A person checks a sample. Trained reviewers check recorded calls against what the agent reported. At Oxford and Buckinghamshire at least 10% of calls were quality assured at the time of the evaluation.
- Audience
- Customer facing
- Autonomy
- Supervised agent
- Adoption
- Early adopters
- Channels
- Phone and voice, SMS
What is it worth?
Benchmarks are computed from the public deployments below: one data point per organization per KPI, with who made each claim.
| KPI | Median | Reported range | Data points | Claimed by |
|---|---|---|---|---|
| Interactions handled | Not pooled | 270 to 1636 | 3 | 3 organization |
| Accuracy | Too few to pool | 99.5% | 1 | 1 organization |
| Cost savings | Not pooled | about GBP 35 | 1 | 1 organization |
| Customer satisfaction | Too few to pool | 88.9% | 1 | 1 organization |
Value drivers: Employee productivity, Risk and loss reduction, Customer experience, Lower cost to serve.
Indicative value
A hospital group that discharges 40,000 patients a year into a routine follow up call
EUR 40,833 to EUR 259,200
Clinician time released from routine follow up calls per year
How this is calculated
Formula: patients * eligible * handled * minutes / 60 * hourlyCost. The low scenario uses every low input, the high scenario every high input.
| Input | Low | High | Basis |
|---|---|---|---|
| Patients discharged each year who are due a follow up call patients, patients per year | 40,000 | 40,000 | The reference organization. |
| Share of those patients the agent can call eligible, fraction of patients | 0.5 | 0.9 | Editorial assumption. At Oxford University Hospitals and Buckinghamshire Healthcare 97% of patients listed for a Dora call were eligible, but that is one narrow pathway; replace with the share for your own pathways. |
| Share of called patients that need no clinician time beyond reading the report handled, fraction of called patients | 0.35 | 0.48 | Based on the Oxford and Buckinghamshire evaluation, where 767 of the 1580 eligible patients Dora contacted (48.5%) had no clinical concerns. The 63% figure in that paper is relative to the 1213 patients who completed the call, and patients who did not answer still cost clinician time because their notes were assessed. The low value allows for pathways with less reach. Replace with your own. |
| Nurse time a manual follow up call takes minutes, minutes per patient | 10 | 15 | Editorial assumption, replace with your own time study. |
| Fully loaded cost of nurse time hourlyCost, EUR per hour | 35 | 60 | Editorial assumption, replace with your own cost. |
What it leaves out: Counts nurse time saved on patients with no concern only. It leaves out the cost of the AI, the telephony and the quality review, the extra callbacks that false flags create, and any effect on readmissions or complications, which none of the deployments on this page report.
Who already uses it?
5 public deployments, strongest evidence first. Grades: A regulator or audit, B the organization itself, C vendor case study, D anonymous or estimate.
Universal Health Services
United States · Healthcare · 2025
Universal Health Services (UHS) and Hippocratic AI deployed generative AI agents that phone patients after they leave hospital, starting at Summerlin Hospital Medical Center in Las Vegas and Texoma Medical Center in Denison, Texas. The agent reviews discharge and medication instructions, asks about new or worsening symptoms and answers questions, and the system pushes a message to the nurse when a patient wants a call back. UHS says thousands of patients were contacted from the start of 2025 and reports an average patient rating of 9.0 out of 10, with plans to extend the program to more of its acute care hospitals.
No outcome disclosed.
Dr Balmis General University Hospital
Spain · Healthcare · 2023
The cardiology department of Dr Balmis General University Hospital in Alicante enrolled 274 of 302 screened patients discharged after transcatheter aortic valve implantation in 2023 in the TeleTAVI study (28 were excluded for language, access route or consent), in which Tucuvi's voice assistant LOLA phoned patients at weeks 1, 2 and 4 and months 3 and 12 using clinically designed question protocols. Calls that raised alerts went to the clinical team, and 73% of the 274 patients went home within 48 hours. The authors report no significant differences in complications or readmissions between very early, early and standard discharge. It was a single center, single arm study.
- Interactions handled: 1039, calls to the 274 patients enrolled in the TeleTAVI study in 2023, with follow up scheduled up to 12 months
"A total of 1039 calls were made using the virtual assistant, with a total of 385 h of automated conversation."
Claimed by: organization - Customer satisfaction: 88.9%, patients satisfied or very satisfied, CSAT 4.68 of 5
"The vast majority of patients (88.9%) reported to be satisfied or very satisfied (CSAT 4.68/5); 86.3% of them would recommend the virtual assistant, ‘Lola’, according to the ‘NPS’ scale."
Claimed by: organization
Oxford University Hospitals NHS Foundation Trust
United Kingdom · Healthcare · 2022
Oxford University Hospitals and Buckinghamshire Healthcare NHS Trust use Ufonia's Dora, an automated telephone assistant, as the standard follow up for patients who had routine cataract surgery. Dora calls three weeks after surgery, asks symptom questions, takes second eye surgery preferences and hands a report to a clinician, who discharges the patient or lists them for second eye surgery, or calls back within 48 hours when the report shows potential concerns. A peer reviewed service evaluation of the first seven months at each site covered 1636 patients. An earlier trial led by Newcastle University with Oxford and Imperial College Healthcare estimated a saving of around 35 pounds per patient.
- Interactions handled: 1636, first seven months of use at each of two NHS trusts
"Across both clinical sites, 1636 Dora R2 consultations were included."
Claimed by: organization - Accuracy: 99.5%, negative predictive value of a "no concerns" outcome, 763 of 767 patients
"The negative predictive value for a “no concerns” outcome was 99.5% (763/767), indicating that when Dora R2 identified no concerns, patients were highly unlikely to require clinical intervention."
Claimed by: organization - Cost savings: about GBP 35, per patient, compared with standard care, estimated in the 2024 trial
"Using this telemedicine follow-up assessment was estimated to have saved around £35 per patient compared to standard care."
Claimed by: organization
Peking Union Medical College Hospital
China · Healthcare · 2019
The orthopedic department of Peking Union Medical College Hospital (PUMCH) adopted an AI assisted follow up system in April 2019 that phones patients after surgery in daily batches, recognizes their speech and dialects, asks scripted questions about recovery, nursing, health education and the ward environment, and passes feedback to surgeons and nurses. A published exploratory study compared 270 patients followed up by the system with 2656 followed up by manual calls. Reach was similar (92.2% connected for the AI system against 93.3% for manual calls), and the AI calls collected feedback from 10.3% of patients against 2.5%.
- Interactions handled: 270, patients followed up by the AI assisted system, April to May 2019
"A total of 270 patients were followed up through this system."
Claimed by: organization
Milton Keynes University Hospital
United Kingdom · Healthcare · 2024
The ophthalmology department at Milton Keynes University Hospital started a 12 month trial of Dora, Ufonia's conversational AI assistant, in July 2024. Dora calls patients after cataract surgery to ask about symptoms and whether they want their second eye operated on, answers common questions and collects feedback on the service. Patients with a mobile phone get a text with the date and time of the call, and Dora tries other times and days when the first does not suit. No results are reported yet.
No outcome disclosed.
How do you implement it?
A model agnostic playbook: what to prepare, the order to build in, and what goes wrong.
Data you need
- Daily list of discharged or operated patients with procedure, date, contact numbers, language and consent
- Clinician approved question protocols and escalation rules per pathway
- Inclusion and exclusion criteria per pathway
- Outcome data for the safety audit, such as unplanned visits, treatment changes and readmissions within a set window
Systems to integrate
- Electronic health record or patient administration system (read and write)
- Telephony and SMS provider
- Clinician worklist or task queue for call reports and callbacks
- Nurse line or contact center for handover
- Call recording and quality review tool
Complexity: High
The build needs a feed of discharged and operated patients from the electronic health record or patient administration system, call protocols signed off by clinicians, a written rule for what counts as a concern, a worklist where a clinician reads each report within a fixed time, and a quality review of recorded calls. In the UK and the EU, software that classifies symptoms can count as a medical device, as it does for Dora and LOLA, which adds a conformity route and post market surveillance.
- 1
Start with one routine pathway
Pick a pathway where most patients recover without events and a clinician already follows a script, as Oxford and Buckinghamshire did with uncomplicated cataract surgery. Write the inclusion criteria first and keep patients with complicating conditions on the existing route.
- 2
Set the flag rule to miss almost nothing
Agree with clinicians that doubt counts as a concern. The Oxford and Buckinghamshire evaluation reports a positive predictive value of 16.8% for "potential concerns" and a negative predictive value of 99.5% for "no concerns", described by its authors as an intentionally cautious design. Expect many flags that turn out fine and size the callback team for them.
- 3
Decide what happens to every outcome
Write a rule for no concerns, potential concerns, no answer and an incomplete call, with a time limit for each callback. At Oxford and Buckinghamshire patients with potential concerns are called back within 48 hours, and patients who do not answer are handled like a missed clinic booking.
- 4
Audit safety against what happens next
Compare patients with a "no concerns" outcome against unplanned visits and treatment changes in a set window after the call. The Oxford and Buckinghamshire evaluation used three months and traced four unexpected management changes within two weeks of such calls.
- 5
Check who the agent cannot reach
Report reach and outcomes by age, language and ethnicity. The Oxford and Buckinghamshire evaluation found no significant difference in call outcomes by age, gender, ethnicity or eye status, but only for patients who met the inclusion criteria, and 3% of listed patients were ineligible, mostly for complex surgery, language, cognitive or hearing reasons. Keep a human route for those patients.
- 6
Add pathways one at a time
Move to the next procedure only after a clinician has reviewed a sample of calls and the audit shows no missed concerns. Each pathway needs its own protocol and its own flag rule.
Guardrails
- Clinicians own the question set and the flag rules, and the agent never reassures beyond its script
- A question the agent cannot answer counts as a concern, not as no concern
- The agent discloses that it is automated and verifies the patient before asking about health
- Patients can ask for a person at any point on the call
- Patients with an urgent clinical concern or a communication need are kept on the human route
KPIs to instrument
- Answer rate and completed call rate per pathway
- Share of calls flagged, and the share of flags a clinician confirms
- Unplanned visits and treatment changes after a "no concerns" outcome
- Time from flag to clinician callback
- Patient rating and the share who ask for a person
- Clinician minutes per patient followed up
- Reach and outcomes by age and language
Human in the loop
A named clinician reads every report that carries a flag and decides the next step, and a sample of "no concerns" reports is read as well. A quality reviewer may raise a flag but should not clear one without a clinician seeing the call, and unanswered calls are assessed against the patient record instead of being discharged by default.
Common failure modes
- A flag cleared by a reviewer
- In the Oxford and Buckinghamshire evaluation the call of a patient with a retained lens fragment was flagged by Dora as potential concerns, and the human quality assurance step relabeled it as "no concerns". The patient later needed an unexpected change in management. Reviewers who can lower a flag need a clinician to see the call first.
- Callbacks swamp the clinic
- A cautious rule produces many flags. With a positive predictive value of 16.8%, most callbacks end with reassurance. Track flag volume against callback capacity and tune the questions with clinicians.
- Reach gaps nobody sees
- Patients without a phone, with hearing or cognitive difficulty, or who speak another language drop out of the automated route, and an average hides it. Report reach by group and keep a manual route.
- A satisfaction score with no method
- A rating such as 9.0 out of 10 says little without the number of responses and the question asked. Record the method before the score is quoted to a board or a regulator.
- Drift into triage and treatment advice
- Once the agent judges symptoms, it can become a regulated medical device. Keep scripts inside the protocol and check the device rules in your market before widening the scope.
What are the risks and rules?
EU AI Act
Depends on design
A voice agent that talks to patients must tell them they are interacting with an AI system (Article 50). If the software is itself a medical device, or a safety component of one, that needs a third party conformity assessment under the EU Medical Device Regulation, it is high risk under Article 6(1) and Annex I. A script that only collects answers for a clinician to read is closer to limited risk, while a system that classifies symptoms and decides who is reviewed can move into device territory. Dora is described as a UKCA medical device and LOLA as a class I medical device under the earlier EU directive.
Rules that apply
Guidance
- Declaratory ruling on AI generated voices under the TCPA (FCC 24-17) (Federal Communications Commission, North America). AI generated voices count as an artificial or prerecorded voice under the TCPA, so US follow up calls made with AI need the consent the TCPA requires for such calls.
- Medical devices, software applications (apps) (Medicines and Healthcare products Regulatory Agency, Europe). Explains when stand alone software counts as a medical device in Great Britain, which decides whether a symptom classifying follow up agent needs UKCA marking.
- Article 50, transparency obligations for providers and deployers of certain AI systems (European Union, Europe). People must be informed that they are interacting with an AI system unless this is obvious from the context.
Controls to put in place
- Clinical sign off of every question protocol and flag rule, with a version history
- Written decision on whether the software is a medical device in each market it is used
- Call recordings and reports kept under the health record retention policy, with access logs
- Safety audit of "no concerns" outcomes against unplanned care, reviewed by a clinical lead
- Consent and opt out for calls and texts, logged per patient
Frequently asked questions
- Do patients accept an AI follow up call?
The deployments report acceptance, though by different measures. At Oxford University Hospitals and Buckinghamshire Healthcare 86% of eligible patients answered the Dora call, and the patient acceptability Net Promoter Score was 47. At Dr Balmis General University Hospital 88.9% of patients said they were satisfied or very satisfied, and UHS reports an average rating of 9.0 out of 10. Newcastle University's earlier trial found most participants accepting, with some concern about the lack of a human element when complications arise.
- Does an AI follow up call reduce readmissions?
None of the deployments on this page reports a readmission result from the calls themselves. The Alicante study found no significant difference in complications or readmissions between patients discharged very early, early or on the standard timeline, but it had no control group without the assistant. Treat readmission claims from vendors without a named hospital and a method with caution.
- Is a post discharge follow up agent a medical device?
It can be. Dora is described as a UKCA medical device and LOLA as a class I medical device under the earlier EU directive. A script that collects answers for a clinician to read is a different product from one that decides who is flagged, so ask your regulator which applies before you launch.
- Which patients should stay on the human route?
Patients with an urgent concern, complex or combined procedures, and patients who cannot hold a phone conversation in a supported language or have a hearing or cognitive impairment. At Oxford and Buckinghamshire 3% of patients listed for a Dora call were ineligible, mostly for these reasons.
How to cite this page
Blits.ai AI Use Case Library, "AI agent for outbound post discharge and recovery follow up calls", last verified 14 October 2026, https://www.blits.ai/ai-use-cases/outbound-post-discharge-follow-up-agent. Licensed under CC BY 4.0. Method: how we verify use cases.
Changelog
- 14 October 2026: First published