The starting point
The question was whether conversational AI could make ordering feel natural, from browsing through to building an order. We were evaluating vendors to develop a proof of concept.
At that stage, resource and technical constraints meant the vendor prototypes would use mocked data. I realised that I could build a benchmark experience myself and use it to understand the problem more directly.
The experience behind the demo
My first working version came together in roughly an hour using AI tools. It could respond to a user, browse mocked menu data, and build an order, without payment integration.
Trying the experience revealed the harder questions. A request to change a drink to iced could restart the order. The bot asked too many follow-up questions, and suggestions could feel arbitrary. A technically working response was only the beginning of a useful interaction.
My contribution
I built the benchmark chatbot and refined the conversational experience. I then worked with a vendor to develop the initiative into a working POC and progressed it to senior-leadership review.
The benchmark gave me a more concrete basis for evaluating proposals and output. It also exposed gaps in the requirements that were difficult to see while the idea existed mainly as a brief.
Learning through the prototype
The initial build was followed by roughly two weeks of experience refinement. Conversation flow, tone, context, and the balance between suggesting and asking needed deliberate attention.
That work turned an abstract ambition into something people could try, critique, and compare. It made the requirements more tangible while keeping the limitations of a mocked-data POC explicit.
The outcome and handover
The POC was completed, reviewed, and handed over to the product team for further evaluation and ownership.
The outcome was a working reference and a better-informed decision process. This case does not claim a production launch, live payment integration, or measured commercial uplift.
About this case
Generalised from a 2026 delivery summary and my earlier public reflection on the prototype. The early build used mocked data; the later milestone was POC completion and handover.