Turning Data Requests Into Self-Service: Building an AI-Powered Analytics Product Users Love

Built on Genie, it delivers trusted AI-powered self-service analytics to the people closest to the moment, without pulling engineers off core work. For teams evaluating how conversational interfaces change the analyst's role, AI is rewriting the analyst's job description in ways that make this kind of embedded tooling increasingly central to modern operations.

The result is not a flashy launch but a disciplined product pattern: meet users where they already operate, inject real business context into the model, guide scope with the UI, and run a lightweight improvement loop so quality compounds over time.

Dashboards hit their limit in the live moment

Core KPIs already appeared in real time, yet the highest value questions surfaced mid‑event and demanded context, comparison, and nuance. Routing each of those to the handful of SQL‑literate owners created a people bottleneck, not a data bottleneck. The team drew a line, committed durable metrics to the dashboard, and gave every other question a conversational surface that could translate intent into SQL and return consistent answers.

Conversational analytics also matched how operations think. Insight unfolds as a thread, for example ticket sales lead to show time, then to weekday effects, and users need to follow that thread while the answer still matters.

Building a context‑aware assistant inside existing workflows

The Wizard is a Genie‑powered text‑to‑SQL chat embedded in the on‑sale tool, later extended to Slack. The heavy lift was product work, not application code. The team labeled tables and columns in domain language, wrote instructions that mapped vague prompts like "how are we doing today" to concrete metrics such as tickets sold, redirects, and regions, and validated queries so repeat questions returned the same trusted SQL.

A PM took the Genie workspace about 90 percent of the way with curated context and a sandboxed power‑user group. Engineers then wired the API and built a thin React UI. Slack brought the same surface to mobile and public channels, which unexpectedly turned open threads into living examples where colleagues learned by watching each other query.

Design choices that accelerated trust and adoption

Two choices carried outsize weight. Suggested questions did more than seed usage, they implicitly defined scope and data recency, for example multi‑month averages signaled available history. Polling responses that showed step‑by‑step status, wrapped in playful wizard language, made latency legible and reduced uncertainty.

A staged rollout reinforced confidence. The team launched to a focused beta, added surface‑level permissions, and deferred real‑time coverage initially because upstream ETLs ran hourly or overnight, while the surrounding tool already showed live metrics. Historical context drove early value, and real‑time coverage followed once pipelines caught up.

A small maintenance loop that keeps getting smarter

Going from version one to version two is where many tools stall. The team built a loosely auto‑research‑inspired loop that uses conversation logs to spot dissatisfaction signals, clones the Genie workspace, benchmarks a proposed change, then surfaces a diff for approval before promotion. Run weekly, it catches issues no one would manually find and fixes many within minutes, which quickly rebuilds trust.

Adoption is tangible: 83 unique users across 28 teams, roughly 10 percent of monthly actives, and more than 760 questions answered that would have interrupted engineers. Internal polls suggest engineers reclaimed significant time previously spent on ad hoc SQL.

Treat AI self‑service as a compounding flywheel, not a feature

Meeting users on their native surfaces, grounding the model in business language, and investing early in a maintenance loop turned a one‑off launch into a compounding capability. Conversational analytics covers the long tail of questions, frees dashboards to focus on durable signals, and lets domain experts pull the thread in the heat of live operations.