Atlas — AI sales research, embedded in the CRM

The Atlas sales team spent the first ten minutes of every call apologizing for not knowing things. Account research was a manual job — LinkedIn, the company blog, the last three quarterly earnings calls — and most reps skipped it. The ones who didn't skip it skipped sleep instead.

We built a research agent that runs the night before each scheduled meeting. It pulls signals from HubSpot — last touchpoint, deal stage, prior call notes, contract value — and augments with a curated allowlist of public sources. It produces a one-page brief: who's in the room, what's changed at the company since the last contact, what was promised in the previous meeting, and three suggested talking points. The brief lands in the rep's inbox at seven in the morning.

The technical work was unglamorous and, in retrospect, the right work. Sandboxing the agent's web access against the allowlist. Structuring outputs against a Zod schema so the brief always fit on one page. Building the HubSpot integration through their official MCP server. And a careful evaluation harness that let us catch regressions in the agent's output quality before they hit reps in the field.

Briefs are now opened by 91% of reps. Average post-call ratings — a field reps fill in after every meeting — climbed by 0.8. The research-to-meeting hours that reps used to skip are gone, and they didn't have to do them anyway.