How We Built a Voice AI System for a UK Restaurant Group in 4 Weeks

How we built an AI voice system for a UK restaurant group in 4 weeks

Earlier this year, a UK restaurant group came to us with a straightforward problem: their phone lines were constantly busy, staff were spending a significant portion of every shift answering the same booking and menu questions, and they were losing reservation calls during peak periods because no one was available to pick up.

They had looked at off-the-shelf chatbot products. None of them could handle natural voice conversations in real time. They had also looked at call centre outsourcing, which would have cost more than the problem was worth.

We proposed a custom solution: a voice AI system built on Amazon Nova Sonic that could answer calls, understand natural language, handle bookings, and escalate to a human when needed. Four weeks later, it was live.

The technical challenge

Voice AI is harder than text AI. The challenge is not just understanding what someone says — it is understanding it fast enough that the conversation feels natural. Latency is everything. A 2-second pause in a phone conversation feels like an eternity.

Amazon Nova Sonic, part of the Amazon Bedrock foundation model family, is built specifically for real-time voice interactions. It processes speech-to-speech natively — meaning it does not need to convert voice to text, process text, and convert back to speech. This collapses the latency significantly.

Our target was sub-200ms end-to-end response time. We achieved it.

What we built

The system handles three primary call types:

• Reservations: The AI can take a full reservation — date, time, party size, contact details, and any special requirements — and write it directly to the booking system.
• Menu enquiries: Questions about allergens, daily specials, dietary requirements, and opening hours — answered instantly from a structured knowledge base.
• General enquiries: Directions, parking, private dining availability, gift vouchers.

For anything outside these categories — or when a caller explicitly asks to speak to someone — the system transfers to a human staff member with a brief handover summary.

The architecture

The stack is entirely serverless on AWS:

• Amazon Nova Sonic (via Amazon Bedrock) — core voice AI model
• AWS Lambda — call handling logic and API orchestration
• Amazon DynamoDB — session state and conversation context
• Amazon API Gateway — telephony integration layer
• AWS Step Functions — booking confirmation and handover workflows

There are no servers to manage, no infrastructure to patch, and the system scales automatically to handle simultaneous calls without pre-provisioning capacity. In testing, we confirmed 10 parallel calls with consistent sub-200ms latency.

The build timeline

| Week | Milestone | |------|-----------| | Week 1 | Requirements scoping, knowledge base preparation, architecture design | | Week 2 | Core Nova Sonic integration, Lambda handlers, Amazon DynamoDB session management | | Week 3 | Booking system integration, telephony testing, latency optimisation | | Week 4 | User acceptance testing with restaurant staff, go-live, monitoring setup |

What this means for other UK businesses

The Bella Cucina build is a restaurant use case, but the architecture is reusable across any business that receives high volumes of similar inbound calls: GP surgeries, estate agents, letting agents, legal firms, accountants, and logistics companies.

If your team spends a meaningful portion of its time answering the same questions on repeat, a voice AI system on AWS will almost certainly pay for itself within months.

The starting cost for our AI solutions is from £8,000. The Bella Cucina system was in the lower range of that bracket — and was live in four weeks.

One important point

We always start with a 2-week PoC. If we cannot demonstrate a working system in a controlled environment that proves the use case for your business, we do not proceed to a full build. You own the PoC if we do not proceed. There is no obligation.