How we built Foliage's 0 to 1 sales-research agent from scratch and launched it in nine weeks, cutting research time from 45 minutes per prospect to 11.
Foliage's sales reps were spending 45 minutes per prospect on research before writing a single word of outreach. Multiply that across twelve reps running twenty prospects a week each, and you're looking at 400+ hours a week on research. Not selling. Research.
Build a system that takes a company name and some targeting parameters, then produces a research brief sharp enough that a rep can write their first message in under ten minutes. Not a Wikipedia summary. Not a generic company overview. The kind of context a good SDR would piece together after an hour of focused digging, delivered in thirty seconds.
The agent runs in three stages. Stage one pulls structured data: company size, funding history, tech stack, recent hires in roles that matter. Stage two processes unstructured sources: recent press, job postings (which are a surprisingly honest window into strategic priorities), and relevant LinkedIn activity from key people at the account.
Stage three is the synthesis layer, and it's where we spent most of our iteration time. Feed a pile of research to a generic AI model and you get a generic summary. Foliage needed something opinionated: a brief that identified the highest-probability angle for their specific product. That takes prompt engineering, evaluation frameworks, and a lot of reps giving feedback on outputs that were almost right but not quite.
The difference between a useful AI output and a generic one is almost always in how you structure the problem before the model ever sees it.
Week one: embedded research. We sat with four reps for two days and just watched them work. We collected thirty examples of research that had actually led to booked meetings. We built a rubric for what "good" looked like before we wrote a single line of code.
By end of week two we had a rough prototype in front of real reps. Not polished. Not production-ready. Useful enough to get honest feedback, which is the only kind that matters.
The number that closed the deal: time from prospect to ready-to-send message dropped from 45 minutes to 11. A 75% reduction. And reps said the quality of the research was actually higher than what they'd been producing themselves.
How we redesigned and rebuilt Northwind's core advisor platform: 70 screens, active users, live roadmap, zero production incidents.
A case study in moving fast with AI in a high-stakes environment. From brief to five hospitals piloting in nine weeks, and what it actually took to get there.