I worked with Claude to get a rough version of the concept, enough of a working prototype to bring the idea to life and share with the team, rather than talk about it in the abstract. That prototype became the basis for mapping the feature set against a Jobs to be Done framework, breaking the negotiation down into what a seller actually needed to do at each point: understand the price, contest it, see evidence, decide, or get handed off to a person.

For the initial messaging approach, I built built a hybrid model: fixed blocks for anything legal needed word-perfect, data slots filled live from the pricing engine, and flex zones drawing from a deep, pre-approved copy library, specifically to de-risk what the agent could say to a seller. Working with Claude, I turned that approach doc into a full authoring pipeline: 363 pieces of reviewed template copy across openers, closers, core claims and evidence lines, benchmarked against explicit legal and tone checks.

Testing that library against real conversations, though, the team decided the fixed-pool approach was constraining the negotiation more than protecting it. Furthermore, it was deemed simpler just to employ an agent rather than a vast library of randomised messaged. We chose to move to AI messages rendering directly at runtime, grounded in the same rules and objection research the library had captured, rather than continuing to select from a fixed set.

Alongside the messaging work, I mapped the all potential conversations as user flows, from the recommended price through acceptance, rejection reasons, price justification, and the handoff to a support agent, so every branch the agent could take was planned and reviewed before I designed the final UI flows.