How We Built an AI Voice Platform for a 40-Doctor UK Healthcare Network

What We Built

ValueStreamAI built Veda, an AI voice platform that manages inbound patient calls, appointment scheduling, and payment collection for a UK healthcare administration company overseeing 40 doctors across multiple clinic locations.

This is not a chatbot. Not a phone menu. Veda conducts real, intelligent conversations with patients: verifying their identity, finding the right appointment slot, collecting payment, and sending confirmation, all without any human involvement. It runs 24 hours a day across the entire doctor network.

The Technology

Veda is our own engineering wrapped around a managed conversation layer. The conversation itself runs on ElevenLabs Agents (formerly ElevenLabs Conversational AI). Everything the conversation needs to act on, the patient records, the diaries, the forms and the money, is a system we built and own.

We chose ElevenLabs deliberately: documented APIs for every part of an agent, leading speech recognition and voice models, and a real testing suite with tool mocking and repeated runs. Newer platforms often lead with benchmark scores nobody outside the vendor can verify, and open-source stacks vary widely with how and where they are hosted. For a medical line, we wanted the platform we could automate, test and audit through its API.

Core components:

• Conversation layer: ElevenLabs Agents, handling speech recognition, the language model, voice and turn-taking, connected to everything below through tools. The language model is a swappable component; we benchmarked four before settling
• Practice management system: our own PMS holding doctors, patients, calendars and bookings, with each doctor's records isolated from every other doctor's
• Doctor schedule overrides: custom per-doctor overrides for added clinics, dropped sessions and moved lists, because a real diary is never just opening hours
• Doctor forms: a custom forms system so each doctor gets the information they need before a consultation
• Voice & Call Routing: VoIPStudio, handling inbound call management and routing across all 40 doctor lines
• Payments: Worldpay, via two routes: secure hosted payment links generated mid-call and delivered by SMS, and in-call keypad capture with custom DTMF masking so card tones never reach the AI, the transcript, or the call recording
• Backend: Custom Python API with a secure PostgreSQL database: all patient data, calendar events, and payment records managed in one governed system
• Compliance: Patient records, calendars and payment records live in our own governed database rather than inside the voice platform, and card data never reaches the conversation layer at all, so the compliance boundary is set by the architecture rather than by policy

Business Value

Our client manages the administrative operations of 40 doctors across multiple UK locations. Before Veda, their admin team handled every inbound call manually: booking appointments, chasing payments, and managing rescheduling across dozens of separate calendars.

• Months 1–4: Automated 50% of inbound call volume, freeing the admin team for exception handling and patient care coordination
• Months 5–12: Reached 100% call capture rate and a 40% reduction in total administrative overhead
• Annual Impact: Projected £150,000+ recovered from previously missed or abandoned appointments

A full-time medical secretary in the UK costs at least about £2,200 a month at the April 2026 National Living Wage of £12.71 an hour, and considerably more with experience, employer costs and cover. And they are available 8 hours a day, 5 days a week. Veda covers every hour they cannot (evenings, weekends and bank holidays) across all 40 doctors simultaneously.

Human Hours vs Automation: Covering Every Hour of the Week

The fair comparison is not Veda against nothing. It is Veda against what it would take to answer the phones with people at the same hours.

The assumptions. A week has 168 hours. One full-time secretary covers 40 of them. Pay is the National Living Wage of £12.71 an hour (April 2026), deliberately the legal floor rather than a realistic wage, so every figure below is an underestimate. It excludes employer National Insurance, pension, unsociable-hours premiums, and holiday and sickness cover.

And that buys one person on one phone line at a time. A single secretary answers one call while the next caller waits; during peak hours the client was losing 30% of calls to abandonment. Veda answers every line at once, at any hour.

With automation, measured by the client. Veda handled 85% of routine bookings with no human involvement, and total administrative labour costs fell 40%. The admin team still exists: it handles the exceptions, the tickets Veda raises, and patient care coordination. Automation moved them off the phones, not out of the business.

The other side of the ledger is the build and the voice platform's per-minute usage, which scales with call volume. Even so, staffing out-of-hours phones with people starts at around £84,600 a year at minimum wage, before the extra people needed to cover peaks. Run your own figures through our hire versus automate calculator.

The Problem

Our client manages the administrative operations of 40 doctors across multiple UK locations. Each doctor runs their own patient list, appointment schedule, fee structure, and clinic calendar, and all of it was being handled by a central admin team fielding every call manually.

The results were predictable: a 30% call abandonment rate during peak hours, after-hours calls going completely unanswered, and manual scheduling errors creating double-bookings and frustrated patients. Missed appointments represented an estimated £150,000 in lost revenue per year. Adding more staff was expensive and still left the after-hours gap wide open. What the business needed was a system that could do the full job of a medical receptionist, across every doctor in the network, at any hour, without errors.

The Solution

ValueStreamAI built Veda specifically for multi-doctor healthcare administration: a conversation layer on ElevenLabs, and around it a practice management system, doctor schedule overrides, doctor forms and a payment system that we engineered ourselves.

When a patient calls any of the 40 doctor lines, Veda handles the entire interaction. It verifies whether the patient is registered with that specific doctor, collects their details if they're new, checks real-time appointment availability, quotes the correct consultation fee, generates a secure Worldpay payment link, books the appointment, and sends the patient an SMS with a clean payment link, all within a single phone call.

After the call ends, the system continues working in the background. Unpaid bookings are automatically released after 24 hours, freeing the slot for another patient. When payment is received via Worldpay, the doctor's calendar updates in real time. The admin team sees a clean, always-accurate view of every appointment across all 40 doctors, without manually managing any of it.

How It Handles Real Conversations

Medical admin calls are not simple. Patients call back to reschedule. They give incorrect details. They change their mind mid-call. Someone calls on behalf of a patient.

Veda is designed for this. It tracks everything confirmed during the conversation so a patient is never asked the same question twice. When a patient provides their email address, the system reads it back before recording it, because one wrong character means the patient misses their calendar invite, their payment link, and their appointment confirmation. The words the agent speaks ("at", "dot") are speech only: the stored address keeps the real symbols, a distinction we learned the hard way and describe below.

If a call is interrupted mid-booking, the system verifies whether the appointment was completed before deciding what to do next. It never creates a duplicate booking or charges a patient twice.

VoIPStudio manages the call routing across all 40 lines. Worldpay handles the payment layer with secure hosted payment pages. Payment links are delivered to patients as clean, short SMS links rather than the long strings that look suspicious on a phone screen.

Results

Within 90 days of full deployment:

• 40% reduction in administrative labor costs: Veda handled 85% of routine bookings without human involvement
• 100% inbound call capture, including evenings and weekends, with no missed enquiries after hours
• 22% increase in new patient registrations from after-hours calls that previously went unanswered
• 99.2% booking accuracy across all 40 doctors, up from a 12% error rate under manual scheduling
• Sub-2-second typical response time, and patients consistently report the experience as natural and professional

What the Client Says

"Veda didn't just replace our phone system. It transformed how we run the entire operation. We've eliminated the morning phone rush, our admin team focuses only on exceptions and patient care, and every doctor in the network has a live, accurate calendar without anyone having to manage it manually."

This is what it looks like when AI is built for a specific operational problem rather than adapted from a generic product. Veda was designed around how 40 real doctors actually work, and it runs that way every day.

Veda: Built as a Platform, Not a Plugin

Most AI receptionist builds stop at the conversation. A voice agent is connected to a calendar and a no-code tool such as Make or Zapier, and that is the product. Some of these are fine for simple use cases: FAQ answering, lead capture, routing to a human.

They fall apart when the job requires managing real state across a complex operation. A booking flow across 40 doctors is not a single step. It involves verifying patient identity, checking doctor-specific availability, confirming the correct fee, generating a payment link, completing the booking, sending confirmation, monitoring payment status, and releasing unpaid slots automatically. Every step depends on the previous one. Some steps need to happen while the patient is still on the call.

No-code platforms cannot model that reliably. They are built for linear flows and happy paths. Real medical administration does not follow linear flows.

The other problem with those setups: when something goes wrong, you cannot fix it. You are at the mercy of the platform's debugging tools. You do not own the code.

What We Built Instead

We did not try to out-build a voice platform at speech recognition and voice. ElevenLabs runs the conversation. What we engineered is everything the conversation acts on: an AI system architecture designed around the actual relationships between doctors, patients, calendars, and payments. Our practice management system holds the records. Custom doctor schedule overrides keep every diary honest. A custom doctor forms system collects what each doctor needs. VoIPStudio manages the call layer, and Worldpay handles payment through both pay-by-link and DTMF-masked keypad entry. Every component connects through a single purpose-built system that ValueStreamAI owns and can modify at any level.

When something needs to change (a new doctor joins the network, a fee schedule updates, a new clinic location opens), it takes effect immediately across the whole platform, because the business logic lives in our systems rather than in the voice platform's configuration. No waiting for a feature release.

Multi-Doctor Architecture

The platform serves 40 doctors from a single deployment with complete isolation between them. Each doctor operates as an independent practice within the system:

• Their patient list is accessible only to their practice, with no visibility across other doctors on the network
• Their appointment availability reflects their specific clinic schedules and locations
• Their fee structure and consultation types are configured independently
• Their calendar and payment records are fully separated from every other doctor

The AI understands which doctor it is representing on every call. A patient registered with one doctor cannot appear in another doctor's records. Bookings go to the correct calendar, with the correct fee, at the correct location, automatically, across all 40 practices.

The Payment Flow

Veda supports two payment routes, because patients differ in what they are willing to do on a call. Both keep raw card data out of the AI layer entirely.

Route 1: Secure payment link by SMS. When a patient agrees to a consultation fee during a call, Veda generates a secure Worldpay payment link in real time and sends it to the patient's phone via SMS as a clean, short link. The patient pays on their own device at their own pace. Card details never enter the call at all.

Route 2: In-call payment with custom DTMF masking. For patients who would rather pay there and then, and for older callers who do not want to leave the call to tap a link, we built keypad payment capture directly into the call flow. The patient types their card digits on their phone keypad. Those keypress tones are intercepted and replaced with a flat monotone before they reach any downstream system, so the AI never hears the digits, the transcript never contains them, and the call recording never captures them. The captured values are passed straight to the payment service provider rather than into our own environment.

The reason this matters is scope. Under PCI DSS, systems that transmit, process, or store cardholder data fall inside the compliance boundary. Masking the tones at the point of capture keeps the conversational AI, the transcript store, and the recording archive outside that boundary, which is a materially different architecture from simply promising not to log the numbers.

Unpaid bookings are monitored automatically on both routes. If payment is not received within 24 hours, the appointment is released and the slot returns to availability. When payment arrives, the doctor's calendar updates immediately, with no manual intervention, no chasing, and no orphaned appointments sitting on a calendar that will never be filled.

The complete payment lifecycle, from capture through confirmation and calendar update, runs without the admin team needing to touch it.

Why This Scales

We have written about why no-code fails at this kind of complexity. A quick-build version of Veda would handle straightforward calls. The moment a patient needed to correct a detail, reschedule across locations, or had a booking interrupted mid-call, it would break.

This platform handles those situations reliably because it was built as software, designed from the start around the operational reality of a 40-doctor network, not adapted from something built for a simpler problem.

For decision-makers evaluating AI for healthcare administration: the question is not whether AI can handle appointment booking. It is whether the system you are looking at was built for the complexity of your actual operation, or built for a demo.

Beyond the Voice Agent

This case study focuses on the calls and the payments. Veda also includes per-doctor patient forms sent by link, an admin dashboard across every doctor with each doctor seeing only their own diary, and an internal ticketing system where staff raise tasks for each other. The whole system, and how the same pattern maps to e-commerce, law firms and other clinics, is written up in our AI agent ecosystem case study.

What a Year in Production Taught Us

The results above are the outcome. Getting there meant a year of finding out what breaks on a live medical line, and most of it was not in any vendor's documentation. We have published the full set as 36 lessons from a year of live medical phone lines. Five of them shaped Veda more than any others:

• Our own tests once stopped patient messages for a day. A benchmarking session ran with tool mocking off, booked hundreds of real appointments, and exhausted the email quota that live patient traffic shares. Every write tool is now mocked in testing, and any test data a human could see carries the word TEST.
• "I've logged that" is speech, not a tool call. About 1 in 21 message-taking calls confirmed logging to the patient without calling the tool, and the platform's success flag still marked them successful. We monitor tool-call records, never the flag.
• The agent must trust the availability endpoint. A well-meant rule to double-check slots against the raw calendar made the agent tell patients a slot "had just gone" on nearly every call. The practice management system is the single source of truth.
• Payments key off the result, not the HTTP status. On our payment provider, the same success-range status code covers both authorised and refused payments, and a new payment identifier on an already-paid session would mean a double charge. Both are guarded in the database.
• Fixes do not propagate across agents. Every doctor line is diffed against a reference agent through the API, at go-live and after every change, because one agent once missed months of fixes, including the ability to hang up.

Frequently Asked Questions

Which voice platform does Veda run on?

The conversation layer runs on ElevenLabs Agents, chosen for its documented APIs, leading speech models and a testing suite with tool mocking and repeated runs. Everything the conversation acts on, the practice management system, doctor schedule overrides, doctor forms and the payment system, is engineered and owned by ValueStreamAI.

How do you handle patient data and compliance?

Patient records, calendars and payment records are stored in our own private, encrypted database with strict access controls, rather than inside the voice platform. Each doctor's records are isolated from every other doctor's. Card details never reach the conversation layer, the transcript or the recording, because payment happens either on a secure hosted page or through DTMF-masked keypad entry passed straight to the payment provider.

Does Veda work with different accents and medical terminology?

Yes, though not by default. We feed the speech recogniser the network's real consultant surnames, hospital names and place names, which is the cheapest accuracy fix available. Background noise needed explicit handling: a transcript containing only punctuation is noise, not speech, and the agent stays silent rather than answering it. When a caller corrects themselves mid-number or mid-name, the correction replaces what came before.

What happens if a call is dropped or interrupted?

The system handles interruptions without creating errors. If a call drops mid-booking, the system verifies whether the action completed before deciding what to do next. It never creates duplicate bookings or generates two payment requests. If connectivity fails at the clinic level, calls continue to be handled via the cloud infrastructure.

Can Veda handle rescheduling and cancellations, not just new bookings?

Yes. Veda manages the full appointment lifecycle (new bookings, rescheduling and cancellations), applying the correct notice period and cancellation policy for each doctor automatically. Patients can call to make any of these changes and the system handles them end to end, including updating the calendar and notifying the relevant parties.

How long does deployment take?

Adding a new doctor to an existing Veda network typically takes days. Full platform deployment for a new client organisation, including calendar integration, call routing via VoIPStudio, and Worldpay payment configuration, takes 4–8 weeks depending on the size and complexity of the doctor network.

Ready to Build This for Your Practice Network?

If you manage multiple doctors and admin overhead is the bottleneck, this is the kind of system that changes the economics of the operation. Talk to ValueStreamAI about what Veda would look like for your network.

To check those economics against your own network, the hire vs automate calculator compares another reception hire with a one-time build, and the savings calculator totals what your current admin tooling and manual call handling cost each year. Both are on the free tools page.

Talk to Us About Your Practice →