Every sales manager I've ever talked to has the same quiet frustration: the CRM has all the answers, but getting to them takes forever. You want to know which deals are actually at risk this week, and instead you're pivoting spreadsheets, cross-referencing stage history against activity logs, and squinting at a dashboard that can tell you *what* is happening but never *why*. That gap โ between having the data and actually understanding it โ is what I set out to close with Sales Pipeline Coach, a Databricks App built around a Genie Agent for this year's build challenge.
If you haven't worked with one before, a Genie Agent is essentially an AI you point at your data tables. Instead of writing SQL queries or building dashboards, you just ask it a question in plain English โ "which deals are at risk?" โ and it figures out how to answer that from the underlying data, showing its reasoning along the way. That's the whole premise this app is built on: what if a rep or a manager could just ask, instead of digging?
Building the foundation
Every app needs data to work with, so I started there. I set up four tables โ `deals`, `reps`, `stage_history`, and `activities` โ inside Databricks, modeling a realistic B2B sales pipeline: deal stages, quota targets, activity logs, the works. All of it is made-up data, generated with a tool called Faker and checked for consistency (no orphaned records, no missing fields), so there's no risk of any real person or sensitive information showing up anywhere in the system.
On top of that data sits the part that actually does the thinking: the Genie Agent. Setting one up isn't just pointing it at your tables and hoping for the best โ I found it worked much better once I gave it a short "glossary" of definitions specific to this data: what counts as an "at-risk" deal, how a quarterly quota should be calculated from an annual number, how to work out a win rate. I also gave it a few example question-and-answer pairs to learn from. Those two things together โ instructions plus examples โ made a noticeable difference in how consistently accurate its answers were.
The app itself is built with Streamlit (a simple Python tool for building web apps) and kept deliberately basic: one tab shows a dashboard with the big-picture numbers โ open pipeline, closed revenue, a chart of deals by stage โ and a second tab is where the actual conversation with Genie happens. The two are connected, so clicking on a rep's name in the dashboard doesn't just filter a table โ it kicks off a live, specific conversation with Genie about that person's pipeline.
The moment it clicked
The most convincing proof this was actually working came almost by accident. Once the basics were running, I wanted to push it further than the questions I'd already tested, so I asked it to compare my best-performing rep against my worst. I expected the obvious story: top performer, clean pipeline; struggling rep, messy pipeline. Instead, Genie came back with something I hadn't anticipated โ my best-performing rep, sitting at 95% of her quota, actually had *more* dollar value sitting at risk in her open pipeline than my lowest performer. Close to $1.9M in stale deals, compared to roughly $300K for the rep who was supposedly struggling.
It took a second to make sense of, and then it clicked: hitting quota measures what's already closed, but it says nothing about what's quietly going stale behind the scenes. A rep can be crushing this quarter on the strength of a few big wins while an entire backlog of neglected deals piles up, invisible to anyone just watching a quota number. That's exactly the kind of thing a static dashboard would never catch โ and it's the moment this stopped feeling like a demo project and started feeling like something a manager would actually want to use.
What I took away from it
The biggest lesson wasn't really technical โ it was about how to test something like this properly. Every question I'd planned for in advance came back looking clean, which felt reassuring but didn't actually prove much. The real insight only showed up once I asked something unplanned and a little adversarial. If you're building with a Genie Agent yourself, that's probably the most useful thing I can pass on: don't just check that it answers the questions you expect โ try to catch it off guard, because that's where you learn whether it's actually reasoning or just performing.
The other lesson was about trust. A Genie Agent's reasoning is only as good as the context you give it. Left to guess, it will make reasonable-sounding assumptions about vague terms โ but writing down clear definitions myself, in plain language, was what turned "plausible-sounding answers" into answers I could actually rely on. In the end, Genie isn't a feature bolted onto this app โ take it away, and Sales Pipeline Coach is just a chart. With it, the app does something closer to what a good sales coach actually does: notice the thing you weren't looking for.
Feel free to check the App : App Link
#Genie #GenieAgent #DatabricksApps #GenerativeAI #SalesAnalytics
