GrowthScript AI
    Case Study · 300-person infra SaaS · Series C

    From a stalled AI stack to 3× more sales calls.

    A developer-facing observability company had every modern AI tool you can name and almost nothing to show for it. Five tools, no shared brain, pipeline flat for two quarters. We rebuilt the motion as one operating system in Claude Code. The numbers moved in the first month — and kept moving.

    Headline outcomes · first full month
    Sales calls booked per month
    14 → 41
    Daily research + list build
    6 hrs → 25 min
    Reply rate on LinkedIn intent DMs
    <1% → 31%
    01

    The stack that wasn’t working

    The GTM team had bought into AI hard: an intent-data tool, an AI SDR autopilot, a research assistant, a content tool, plus Claude Cowork with connectors to CRM, Gmail, Slack, and a data provider. On paper, modern. In practice, the numbers hadn’t moved in two quarters. Sales calls booked sat at 13–15 a month.

    A modern AI stack and a flat pipeline. The board wasn’t buying it twice.

    02

    Why it stalled — three compounding failures

    The tools didn’t share a brain. The intent tool knew which accounts were hot. The research assistant knew the prospect’s background. The AI SDR knew the email templates. None of them knew what the others knew. An SDR rebuilt context by hand at every step.

    Cowork sessions reset. Good workflows, but every session started cold. The connectors pulled live data fine, but there was no persistent place for the ICP, the voice, or what had already been tried. Nothing compounded.

    The AI SDR sounded like an AI SDR. Grammatically perfect, completely generic. Reply rates under 1%. Two domains started flagging the sends as spam. The team had turned it off but kept paying for it.

    “We had spent a year buying AI and we were busier than ever, but the calendar wasn’t filling up. It felt like we’d automated the busywork and none of the outcome.”
    — RevOps Lead
    03

    What we built

    Not another tool. One operating system in Claude Code, run from the team’s own repo, holding the ICP, the voice, the signal rules, and the history. Six specialist workflows that feed each other instead of working in isolation: Signal scan, List build, Outreach, Qualify, Deal brief, and Retarget.

    Under all of it: a voice profile + automatic linter so nothing went out sounding like a robot, a do-not-contact guard that blocked any draft to someone who opted out, and a long-term memory file so the system remembered what worked last week.

    Cowork is where you do the work. Claude Code is where the system that does the work lives.

    04

    Why Claude Code, not Cowork

    Persistence, visibility, versioning, and enforcement. The team’s system grew past 50 files — knowledge bases, skills, templates, drafts, briefs, a content library. At that point the file tree IS the system, and Cowork’s in-app browser hits a wall. Three things only compound because they live on disk: a memory file of what worked last week, a second brain that decomposes winning posts into reusable patterns, and a component library that assembles new infographics from parts that already performed. Each writes back to a file; each session starts smarter.

    The numbers, first full month

    Before → After
    Sales calls booked / month
    14 → 41
    Daily research + list build
    ~6 hrs → 25 min
    LinkedIn intent DM reply rate
    <1% → 31%
    On-brand infographics / month
    0 → 9
    AI tools in active use
    5 → 1 system

    The lesson

    The team didn’t have a tooling problem. They had a system problem. Five smart tools with no shared brain, no memory, and no enforcement will lose to one operating system that has all three. The AI was never the bottleneck. The glue was. If your stack looks busy but your calendar looks empty, that’s usually the tell.

    Illustrative composite based on common patterns we see, not a single named client. Numbers are representative.