Deal Guardian
Catch deals slipping while there's still time to act
A workflow that reads sales call transcripts and flags the deals that look like they're quietly slipping — so a review becomes an early warning instead of a post-mortem.
Problem
Pipeline reviews are mostly post-mortems. By the time a deal shows up as 'at risk' in the CRM, whatever caused it — a champion going quiet, a competitor showing up, a requirement changing — usually happened weeks earlier, in a call nobody went back and re-listened to.
The warning signs existed. They were just buried in transcripts no one had time to read.
Insight
Risk gets said out loud before it gets logged. The early signs of a slipping deal live in the language of a call — hedging, next steps that never get scheduled, new people suddenly in the room, pushback on price — not in a CRM stage field.
If you treat transcripts as something to listen to continuously rather than archive, you can spot trouble while it's still fixable.
What I built
Deal Guardian reads call transcripts with Gemini, pulls out specific risk signals, scores deals against what slipping deals have tended to look like, and surfaces a short, ranked list of ones worth a closer look.
Every flag comes with the why — the actual moments in the call that triggered it — so it prompts a follow-up instead of just anxiety.
How it fits together
Transcripts come in → Gemini pulls out structured risk signals → a scoring step combines those with deal details → the result shows up in the workflow reps and managers already use.
Pulling the signals out is kept separate from scoring them, so the risk logic can change without re-reading every transcript from scratch.
Calls I made
Show evidence for every flag. A bare risk score is noise people learn to ignore; a quoted line from the call earns a second look.
Rank a few, don't alarm on everything. Surfacing the handful that actually need attention beat flagging everything and training people to dismiss it.
Tune for the cost of being wrong. A falsely flagged deal wastes a rep's time and trust, so being precise mattered more than catching every single risk.
What happened
It surfaces 100+ at-risk deals a week, cut the manual review effort by around 90%, and contributed to roughly a 10–15% lift in conversion. Managers walked into reviews already knowing where to focus.
The unit of attention shifted from 'the whole pipeline' to 'the handful of conversations that actually matter this week.'
What I learned / would do differently
A tool that nudges people only works if it's right often enough to be believed. Calibration mattered more than raw cleverness.
Evidence beats confidence — the credibility came from showing the source, not the score.
The best version of this reduces what people have to pay attention to. Adding more alerts would have made it worse.