It's 8am at a busy hospital. The phone lines open and the queues fill within minutes. Many of those calls are about the same things: What time do I arrive? Where do I park? Which entrance do I use? Can someone come with me? Each of these simple questions takes a slot in the queue, ahead of a patient who has a clinical concern.
That was where SpaMedica started. SpaMedica is one of the UK's largest providers of NHS eye care, with more than 60 hospitals. It put a voice AI agent behind its existing phone system to handle non-clinical FAQs. It went from proof of value to live production in under six months, and in its first couple of months the agent handled about 80,000 patient interactions.
In a recent HIMSS webinar, William Harris, Chief Transformation and Technology Officer at SpaMedica, talked with Druid AI's Leo Gaciu and Stuart Brougher about how they did it.
Starting at the Patient Front Door
SpaMedica's choice of first use case matches what Druid sees across healthcare deployments. Druid's AI Adoption Benchmark draws on 15 months of anonymized production data, not survey results. It shows that most healthcare AI activity is at the patient front door: identity verification, patient access, appointment management and FAQs. Voice is the top channel in healthcare, even though chat leads in higher education and HR/IT. Patients call. About 29% of healthcare interactions with AI agents also happen outside 8am - 5pm, when most contact centers are closed or short-staffed.
SpaMedica didn't replace its existing systems. It added Druid on top of them. "We see Druid a lot as if it was another person working within the organization," Harris said. "We provision it access to whatever it needs to have access to." The agent connects to SpaMedica's NICE contact center platform and to downstream systems, and uses those to answer questions within its guardrails.
Guardrails First, Technology Second
When a caller picks the non-clinical option in SpaMedica's IVR, the AI agent answers. It handles parking, directions, what to bring and the details in the pre-appointment information pack. If a caller's question moves into anything clinical, the agent hands them to a colleague through the contact center queue.
"Even if the AI has the competency, [we won't] allow it to get anywhere near diagnosis or treatment," Harris explained. That's partly because of MHRA regulation, "but also really out of care. When patients are asking questions of their own health, that's dealt with by one of our clinicians."
Today, about one in five conversations coming into the contact center are handled end to end by the AI. "That's not because of the limitation of the technology," Harris noted. "It's because of a limitation of what we will allow the technology to do." Patients rarely call only to ask where the car park is. The agent often answers the main reason for the call and then hands over the follow-up questions. SpaMedica therefore tracks how much of each call the AI handles, not just total containment.
The effect on staff is clear. Call volume hasn't fallen, but the work has shifted. "Our colleagues can focus more on clinically significant calls rather than operationally significant calls," Harris said.
From Pilot to Production in Under Six Months
When asked whether going live had taken two or three years, Harris answered: "No, much, much quicker."
• 4–6 weeks to build the pilot, put it into test and get data back
• A couple more months in production with hypercare, with the agent closely monitored
• Under six months from signing the proof of value to unsupervised live production
The build wasn't the hard part. Defining the guardrails was. Harris described it as a 3D model: what areas the AI may operate in, and how deep it may go in each. "The actual implementation could probably be measured in days to weeks," he said. Most of the program went into guardrails, safety and transparent, reportable AI. Automated reports now check that the agent doesn't breach a guardrail or fall for prompt injection.
Clinical buy-in was key. "Our clinical leadership was really behind it," Harris said. "Clinical staff were very happy that they didn't wind up ever having to deal with those calls again."
Advice for Other Healthcare Providers Getting Started with AI
William Harris gave three pieces of advice to other healthcare organizations:
• Anchor on care, not cost. "Organizations that center their transformation around pure cost or pure process efficiency are missing the trick," he said. Without buy-in from clinical colleagues, scaling becomes hard. He tells teams to "release" change through education and enablement, not "drive" it.
• Just get started. "You cannot predict what people are going to say to your AI. You can only predict how you can control your AI to answer whatever it is asked." Testing in a safe environment is better than spending a year planning every scenario.
• Think digital workforce, not point solutions. Telephony and EHR vendors now offer their own AI, but Harris warned against vendor lock-in. An orchestration layer works across systems and acts as "your digital HR layer," with all the governance in one place. New use cases become new agents in the same workforce, not rebuilds.
What Comes Next
SpaMedica will expand from answering questions to taking action: booking appointments and follow-ups, confirming attendance, checking for complications and rescheduling. Harris will prioritize based on patient wait time and "dead air" — the time patients spend on hold without knowing what's happening — and on where his colleagues are under the most pressure.
"Patients don't call in to chat to an AI," he said. "They call in to get their care moving."
