Case studyActiveCampaign
Teaching the CRM to read the room.
As design lead for the CRM team, I helped turn the sentiment inside customer conversations into a win probability that sales reps could trust and act on.
Deal signallive
As design lead for the CRM team, I helped turn the sentiment inside customer conversations into a win probability that sales reps could trust and act on.
ActiveCampaign is one of the most widely used customer experience automation platforms for small and mid sized businesses, pairing email and automation with a sales CRM.
I led design for the CRM team: the Deals, Accounts, Contacts and pipeline product that sales teams lived in every day. Because the CRM sat right next to automation and email, it already held an unusually complete record of every customer relationship. My focus was turning that raw history into foresight.
Product context: automation and email reporting are the neighbors of the CRM. Because they share the same record, the system already sees how every relationship actually unfolds.
The CRM captured everything: every contact, note, email, conversation, task and stage. And yet the one question a sales leader cares about most still went unanswered.
Without a trustworthy read on a deal's health, teams over invested in deals that were already cold and under served the ones they could still win. Forecasts ran on optimism. The signal that mattered most, how the customer actually felt, was buried in the text of conversations nobody had time to reread.
Win Probability: Not yet calculated. Every field around it was full. The one that would change a rep's day was empty.
A representative rep's open pipeline:
Plenty of numbers about size. Nothing about likelihood.
The strongest predictor of a close was never a form field. It was tone. So the design problem became: how do you turn the emotional temperature of a conversation into a number a busy rep will believe?
The team already had explicit scoring reps understood: Customer Health, Marketing Qualified Lead, Sales Qualified Lead. I treated sentiment as a new input into that same vocabulary, not a black box bolted on the side. New signal, familiar language.
Every signal lived on the contact and account: campaigns sent, sites visited, calls, notes, conversations. Sentiment analysis read that history. Deciding where and how the score should appear on these records, without burying the work reps still had to do, was half the design.
I designed the signal as a clear, legible pipeline. Each step had to make sense to a rep on its own, so the final number never felt like magic.
The win probability did not live in a separate analytics tab. It belonged on the deal itself, beside the contacts, tasks and history a rep was already reading. The same field that once read Not yet calculated became a living signal, with the conversations behind it one click away.
An account brings contacts, deals, notes and activity into a single view. Layering sentiment driven health on top of this meant a rep could feel the temperature of the relationship before scrolling a single note.
Deal value and counts by stage gave leaders the shape of the pipeline. Once win probability fed these views, the forecast stopped being a wish and started being a weighted, defensible number.
The hardest part of AI in a CRM is not the model. It is earning the right to be believed. Three principles guided every decision.
Every sentiment read links back to the conversations that drove it. A rep sees the evidence, not just the verdict, so a surprising score invites a look rather than a shrug.
The model suggests, the rep decides. A score informs the next step. It never quietly moves a deal or changes a stage behind a rep's back.
When the system did not have enough signal to call a deal, it said so, plainly. Faking precision is the fastest way to lose a team's trust for good.
Beyond any single screen, the work reframed how the CRM team thought about AI: not as a feature to add, but as a signal to earn trust in. Sentiment only mattered once it reached a rep at the moment of a decision, in language they already spoke.
In a CRM, trust is the interface. The model can be brilliant and still go unused if a rep cannot see why it said what it said.
Two lessons followed me out of this work. Reuse the vocabulary people already trust before inventing a new one. Sentiment landed because it spoke through scoring reps had relied on for years, not against it.
And place a signal where the decision happens. A prediction in a separate dashboard is trivia. The same prediction on the deal, beside the next call, is leverage. Foresight is only useful when it shows up in the flow of the work.