| Metric | Result |
|---|---|
| Doctors on one ecosystem | 40, across multiple UK locations |
| Routine bookings with no human involved | 85% |
| Booking accuracy | 99.2%, up from a 12% error rate under manual scheduling |
| Administrative labour cost | 40% lower |
| Components, one system of record | Voice agent, PMS, schedule overrides, patient forms, payments, team ticketing |
The Short Answer for Anyone Asking "Can an AI Agent Run Part of My Business?"
Yes, but almost never as one agent on its own. The question we hear most from business owners is some version of "can I get an AI agent for my business?" The honest answer from this build is that the agent was the easy part. What made it work was everything we built around it, so that the agent had something real to read from and write to, and so that the team had somewhere to pick up what the agent could not finish.
This case study is about that whole system. It covers Veda, which we built for a UK healthcare administration company running the operations of 40 doctors. Our original Veda case study focuses on the voice agent and payments. This one is for the reader who runs a business, not a call centre, and wants to know what AI agent development for a business actually involves when it has to work every day for a whole team.
Why "Build Us an AI Agent" Was the Wrong Brief
The starting problem looked like a phone problem: a 30% call abandonment rate at peak, after-hours calls going unanswered, and manual scheduling errors across dozens of separate diaries. The obvious brief is "put an AI agent on the phones".
A voice agent on its own would have answered the calls and then hit a wall on every one. It would have had nowhere reliable to check a doctor's real availability, no way to take money, no way to gather what each doctor needs before a consultation, and no way to hand work to the admin team except by sending an email someone might read. We have seen businesses build exactly that, and then a second agent next to it, and conclude AI does not deliver. Two agents is not an ecosystem. Each one is useful on its own and each one is an architectural dead end, because neither shares data, identity or follow-up with anything else.
So the brief became: build the system the business runs on, and put the agent at its front door.
The Ecosystem, Component by Component
Every component below was engineered for this business and connects to the same governed database. Nothing is copied between screens by a person.
1. A voice agent that answers every call
The agent answers all 40 doctor lines, at any hour, in natural conversation. It verifies whether the caller is registered with that specific doctor, collects details from new patients, checks real availability, quotes the correct consultation fee, books the appointment and offers payment, all within one call. The conversation layer runs on ElevenLabs Agents, which we chose for its documented APIs, leading speech models and a testing suite with tool mocking and repeated runs. The agent reads and writes everything else through tools.
2. A practice management system (PMS) built for 40 doctors
This is the system of record, and the part most "AI agent" projects skip. It holds doctors, patients, diaries, fees and bookings, with each doctor's records isolated from every other doctor's.
- The central admin team works from a dashboard that shows bookings, patients and payments across every doctor in the network.
- Each doctor sees their own diary, their own patients and their own schedule, and nothing belonging to anyone else.
Because the agent books from the PMS rather than from a loose calendar, the admin team and the doctors look at exactly the same truth the agent does.
3. Doctor schedule overrides
A real clinic diary is never just opening hours. A doctor adds a clinic, drops a session, or moves a list to another site. We built per-doctor schedule overrides into the PMS, so the slots the agent offers already reflect those changes. This mattered more than it sounds: a prompt rule that once told the agent to second-guess the availability it was given produced "that slot has just gone" on nearly every call. The fix was structural, and it is a lesson about trusting your own system of record. More on that in our 36 lessons from a year of live medical phone lines.
4. Per-doctor patient forms, sent by link
Each doctor configures their own forms, so a specialist collects what their specialty needs rather than a generic intake sheet. Patients receive their doctor's form by link to fill in before the consultation, and the answers land in the PMS. The agent does not try to gather clinical detail by voice, which is slower for the patient and more error-prone than a form they can complete at their own pace.
5. Payments, two ways
A booking that is not paid is a booking that can quietly fall through. Veda takes payment during the call in one of two ways:
- Pay-by-link: a secure payment link is generated mid-call and sent by SMS as a short, clean link.
- In-call keypad entry with DTMF masking: the caller types their card number on the keypad, and the tones are replaced with a flat tone before they reach the AI, the transcript or the recording, so those systems stay outside PCI DSS scope.
Unpaid bookings are released automatically after 24 hours, and a payment updates the doctor's diary immediately.
6. An internal ticketing system for the team
This is the part that makes it AI for the team, not just AI for the phones. Admin staff and doctors raise tickets for each other inside the same system, so internal requests between teams have an owner and a place to live instead of being scattered across emails, chat messages and sticky notes. It sits on the same records as everything else, so a ticket about a patient or a booking points at the real record rather than a description of it.
Patients reach a voice agent on ElevenLabs Agents. The agent reads and writes the practice management system, which holds doctors, patients, diaries, fees and bookings and applies per-doctor schedule overrides. Patients pay by link or by masked keypad entry, and complete their doctor's own form by link. The admin team works from a dashboard across all doctors, each doctor sees their own diary, and staff raise tickets for each other in the internal ticketing system.
What One Booking Looks Like, End to End
The point of the sequence is what is missing from it: nobody retypes anything. The agent books into the PMS, the payment updates the PMS, the form answers land in the PMS, and the doctor sees the result in their own diary. When something needs a person, the admin team works from the same records, and anything they need from a colleague goes into a ticket rather than an inbox.
The bot books into a calendar it cannot see properly.
A human calls back to take the deposit.
Collected on the phone, or not at all.
Sticky notes, emails and chat messages.
Ask someone what happened.
What It Changed for the Team
The measurable results are in the table at the top: 85% of routine bookings handled with no human, 99.2% booking accuracy across all 40 doctors, a 40% reduction in administrative labour cost, and every inbound call answered, including evenings and weekends. Rollout was staged: the first four months automated half of inbound call volume, and full call capture and the cost reduction arrived over months five to twelve.
The change the client describes is less about numbers. The morning phone rush is gone. The admin team handles exceptions and patient care rather than retyping bookings. Every doctor has a live, accurate diary without anyone managing it by hand. And adding a new doctor to the network now takes days, because the system they join already exists.
That last point is the economic argument for building an ecosystem instead of a bot. The first component is expensive. Each one after it is cheaper, because it plugs into the same data, the same identity model and the same follow-up path.
UK GDPR and Card Payments: How the System Handles Sensitive Data
A system that books patients, collects forms and takes card payments is handling two of the most regulated kinds of data there are, so compliance was designed into the architecture rather than added as a policy afterwards.
- Patient data stays in the UK. Data residency is kept within the UK, which is how the system meets UK GDPR by architecture rather than by policy alone.
- The AI never sees who the patient is when it does not need to. A custom data handling layer keeps patient identifiers out of AI processing, and patient records, calendars and payment records live in our own private, encrypted database with strict access controls, not inside the voice platform. Each doctor's records are isolated from every other doctor's.
- Card numbers never reach the AI. When a patient pays by keypad during the call, custom DTMF masking stops the card tones reaching the conversational AI, the transcript or the call recording, which keeps all three outside PCI DSS scope. Patients who prefer can pay through a secure hosted link sent by SMS instead.
If your business handles regulated data, this is the part to ask any provider about first. The full payment design is in our original Veda case study, and our guide to AI automation under UK GDPR and HIPAA covers what each regime allows an agent to do.
The Same Pattern in Other Businesses
None of these components are specific to medicine. The shape (an agent at the front door, a system of record behind it, structured intake, payment, and a place for the team to pass work to each other) recurs in most service businesses. Here is how it maps. These are illustrations of the pattern, not projects we are claiming to have built.
| Component | Healthcare (Veda) | E-commerce | Law firm | Other clinics (dental, physio, aesthetics) |
|---|---|---|---|---|
| Agent at the front door | Voice agent on 40 doctor lines | Order status, returns and order changes by phone or chat | New-enquiry intake calls | Bookings and rebookings |
| System of record | Our own PMS | The order management system | Case or matter management | Practice management software |
| Rules the agent must respect | Per-doctor schedule overrides | Returns windows, stock, delivery cut-offs | Conflict checks before anything is discussed | Practitioner diaries and treatment lengths |
| Structured intake | Per-doctor patient forms by link | Return reason and photos by link | Intake questionnaire by link | Medical history forms by link |
| Payment | Pay-by-link or DTMF-masked keypad | Refunds and exchanges | Money on account, where the firm's rules allow | Deposits to hold a slot |
| Team follow-up | Internal ticketing between staff | Warehouse and support hand-offs | Fee-earner and support-staff tasks | Front desk and practitioner tasks |
If you run one of these businesses, our guides cover the specifics: AI voice agents for e-commerce, secure document sharing for law firms, and the AI for medical practices hub.
Is This the Right Shape for Your Business?
Probably not, if you have one workflow and good software already. If your systems all expose APIs and the problem is one repetitive job, buy a scoped build or an off-the-shelf tool. Our guide to AI agents for business automation covers what has to exist before an agent is worth building, and a $29 a month AI receptionist is the right answer for plenty of businesses that only need calls answered.
Probably yes, if the agent keeps hitting the edge of your systems. The signs are familiar: the AI books but a human has to take the payment, the information a specialist needs is collected badly by phone, staff pass work to each other through inboxes, and nobody can see across locations or teams. That is a missing system of record, and no agent fixes it on its own.
The Competitor Pulse Check
| Factor | Agentic ecosystem (this build) | A standalone AI agent | Off-the-shelf SaaS stack |
|---|---|---|---|
| System of record | One governed database for every component | Whatever the agent can reach | Several, each with its own copy of the customer |
| Rules of the business | Built into the system the agent reads | Written into the prompt | Limited to each product's settings |
| Payment in the conversation | Yes, by link or masked keypad | Rarely | Rarely at the small-business tier |
| Team follow-up | Internal ticketing on the same records | Email or chat | A separate helpdesk tool |
| Cost of the next use case | Falls, because the foundation exists | Starts again from scratch | Another subscription and integration |
| Who owns it | The business | Depends on the vendor | The vendors |
What an AI Agent for Your Business Costs
We build ecosystems the same way we build single agents: one workflow first, measured, then outward. Engagements start with a $5,000 fixed-scope pilot on one workflow. A typical single-agent system runs $15,000 to $25,000, and multi-agent systems with their own data layer, monitoring and compliance work run $40,000 to $100,000+, which is the band an ecosystem of this shape sits in, depending on how much of the system of record already exists. You get a fixed price before committing to a full build: our automation quote tool gives a ballpark for your own workflow in about a minute, and our pricing page sets out the ongoing retainers that keep a system like this maintained. Our AI agent development service sets out how a pilot is scoped, and the hire vs automate calculator compares a build with another hire.
A year of running Veda in production also taught us what to watch: test runs that fire real messages, agents that claim actions they never took, and why the system of record must be the one source of truth. We have published all of it in why AI voice agents fail in production.
Frequently Asked Questions
Can I get an AI agent for my business without rebuilding everything?
Usually yes. If your core systems publish APIs, an agent can be connected to what you already run, and a first pilot can be live in weeks. Veda included a new practice management system because the business was running dozens of separate calendars rather than one system of record, which is the exception rather than the rule.
What does AI agent development for a business actually include?
More than the agent. In this build it included the voice agent, a practice management system, per-doctor schedule overrides, per-doctor patient forms sent by link, two payment routes, and an internal ticketing system for staff. The agent is the visible part; the system of record and the team's follow-up path are what make it reliable.
How does AI help my team rather than replace it?
In this case the admin team stopped retyping bookings and chasing payments and moved to exceptions and patient care. Staff raise tickets for each other in the same system the agent writes to, and each doctor sees an accurate diary without managing it. The work that needs judgement stays with people.
How much does it cost to build an AI agent ecosystem for a business?
Our pilots start at $5,000 for one workflow. A single-agent system typically runs $15,000 to $25,000, and multi-agent systems with their own data layer, monitoring and compliance work run $40,000 to $100,000+, depending on how many systems already exist. You get a fixed price before a full build.
Can this kind of system work for e-commerce, law firms or other clinics?
The pattern can: an agent at the front door, a system of record, structured intake by link, payment, and team follow-up. The specifics change, such as returns rules for e-commerce or conflict checks for a law firm, and each needs its own compliance review. We would scope one workflow first rather than the whole pattern at once.
Which AI platform did you use to build Veda?
The voice conversation runs on ElevenLabs Agents. The practice management system, schedule overrides, patient forms, payment flows and ticketing system are our own engineering on a governed database that the business owns.
Want This Shape for Your Business?
If your team is drowning in the same calls, the same retyping and the same internal chasing, start with one workflow. Talk to us about an AI agent for your business →, or read how we approach it on our AI agent development service.