Marcus Holm leads growth at Linear, the issue tracker built for engineering teams that ship every day. By the time he came to Mercator, Linear had reached the founder-led ceiling: a steady inbound flow, a CEO doing every meaningful outbound conversation, and no system underneath that could be handed to anyone else.
The brief was specific. Not a campaign. Not a deck. An engine that the team could own, operate, and extend without Mercator in the room.
A team that scaled product without scaling its outbound motion.
For the first three years, Linear sold the way most product-led companies sell: word of mouth, a viral demo, founders in every important deal. It worked until it stopped scaling. The product was running ahead of the system that supplied it with conversations.
What Marcus needed wasn't more headcount. He needed the kind of infrastructure a 30-person sales team would have built over eighteen months — without the eighteen months. "Velocity is the only thing we hire for," Marcus says. "The system has to keep pace with the people, not the other way around."
The buying signal blind spot
Linear's intent data lived in five places: HubSpot, the product DB, Clearbit, intercom logs, and a Notion table maintained by one operator. Each system held part of the picture. None of them spoke. By the time a signal surfaced, the window to act on it had usually closed.
The problem wasn't the data. It was the absence of a control surface.
Mercator spent the first four days inside Linear's stack, not on a strategy deck. The diagnosis came back narrower than expected: every signal Linear needed was already being captured. None of it was being acted on, because there was no surface that ranked, routed, or remembered.
Most teams think they have a data problem. They almost never do. They have a routing problem — and routing is what compounds. — Mercator diagnosis memo, week one
The architecture call was made in the same week. Three custom AI agents — Scoring, Enrichment, Outreach — sitting on top of Clay as the orchestration layer, with HubSpot as the system of record and n8n handling the wiring between them. No new tools introduced. No license bills added.
Three agents, one orchestrator, zero net-new tools.
The Scoring Agent reads every account against thirty-one weighted criteria pulled from product usage, hiring signals, technographic footprint, and recent funding events. A score that used to take a SDR forty minutes to assemble takes the agent under nine seconds, and the rationale travels with the score so the human reviewing it can trust or correct it in one read.
The Enrichment Agent runs once per ranked account, pulling the surface area a senior SDR would normally collect by hand: team comp, stack signals, the last public artifact each decision-maker shipped. The Outreach Agent drafts the opener referencing the rationale — never a templated line — and the result lands in Marcus's inbox as review-ready, not draft-ready.
The data layer that learned to explain itself
The detail that changed Marcus's mind on AI: when the Scoring Agent gets it wrong, the team corrects it in plain language. Within a week of feedback, the model's confidence calibration shifted measurably. Within a month, the exception list shrank by 64%. The agent didn't just get more accurate — it became legible.
Ninety days in, the engine ran without Mercator in the room.
The hand-off ran for six business days. By day seven, Marcus's team owned the operating cadence: a Monday review of the scoring queue, a Wednesday calibration on whatever the agent missed, a Friday digest pushed into the leadership channel. No Mercator on Slack. No retainer on the meter.
Ninety days from kickoff, qualified pipeline was up 187%. Time spent on account research collapsed from four hours per SDR per day to under twenty minutes. The cost line for the entire system, including tool spend, came in under what a single senior SDR would have cost the company.
What I underwrite now isn't the agency. It's the engine they handed me. I'd rebuild it from scratch tomorrow if I had to. — Marcus Holm, Head of Growth, Linear
What Linear walked away with
An owned engine, not a deliverable. Three custom agents and the orchestration around them — operated end-to-end by Linear's existing team after a six-day handoff.
Routing solved before scale. Every signal Linear was already capturing now lands in front of the right operator within an hour, with the rationale attached.
A model that learned in plain English. Corrections come from Marcus's team in language, not config. The agent's calibration improves with every cycle.
Compounding economics. The system's monthly cost is below one senior SDR fully-loaded — and it gets more accurate the longer it runs.