We built a conversational AI. Our customers built a report generator.
On 30 January 2026 we committed the first line of Ask NiceAi — an in-product AI agent meant to let customer-experience teams have conversations with their survey data. By 31 May we'd exceeded our H1 revenue target. But what users were actually doing with the agent looked very little like the conversations we'd designed for: they wanted reports. HTML first, then PDFs, then PowerPoints.
This talk is about how we noticed that (a team practice of weekly customer-conversation reviews, plus making real prompt traffic legible to engineers), what we changed in response (a late but pivotal Bedrock AgentCore Runtime bolt-on to a PHP-on-Lambda backend so the agent could produce documents), and the small set of habits that let us pivot fast enough to make the year.
Specific tech, specific surprises, specific outcomes. Aimed at engineers and product folks shipping AI agents to non-developer end users.
October 22 — Conference Day
David Sickmiller is Principal Product Engineer at AskNicely in Auckland, where for the past two years he's led the company's move into generative and agentic AI — including Ask NiceAi (the subject of this talk), the AI Themes feature, and AskNicely's practices around AI evaluation (DeepEval in CI) and observability. Across nearly nine years at the company he's also been VP of Engineering and Director of Engineering. He cares most about what AI products actually do once real users get hold of them.