AI outbound engine
A system that researches each prospect company, sets aside poor fits, drafts a first message and checks it against that research, and passes every reply to a person in Slack.
- In May 2026 it ran at about $0.18 in AI and data costs per prospect.
What it does
The engine does the slow part of prospecting on a schedule: it researches each company, writes a first message that fits it, and has a review pass check each draft against the research. A person handles every reply, which lands in Slack with talking points from the research. The firm sets the sending limits and can stop all sending with one switch.
The situation
Maple DX, a Salesforce consulting firm, wanted a steady flow of first conversations with companies that run Salesforce. Researching each prospect and writing to them one at a time is slow work, and the firm’s earlier outreach was spread across several separate tools.
What was built
A system that runs on a schedule. For each target company, Claude, Anthropic’s AI model, writes a short research brief that may only state facts with a source. Companies that sell Salesforce services themselves are set aside before anything is written. Claude then drafts a message, and a second review pass checks its facts against the brief.
The system never answers a reply itself. Each reply is labeled and posted to the firm’s Slack with the brief’s talking points, so a person can pick it up. A password-protected dashboard lets the firm run campaigns, set limits, and stop all sending.
The current version has been live since August 2026. It replaced a first version from May 2026.
How it works
Choose target companies
The firm picks the kind of company it wants to reach. Companies that sell Salesforce services themselves, and ones already in progress or recently turned down, are skipped.
Research and fit check
Claude reads public information about each company and writes a short brief that may only state facts with a source. A second check labels the company a likely customer, a services firm, or unclear. Services firms are dropped before any message is written.
Draft and check
Claude drafts the message the prospect reads first. A review pass reads it as the prospect would. A fact the brief does not support gets a small fix and one more check, and a draft that still fails is blocked.
Limits and a stop switch
Sending stays within limits the firm sets, and a stop switch halts all sending. Only a person can turn it back on.
Pass replies to a person
A reply pauses messages to that prospect at once. Claude labels the reply, and the system posts it to Slack with the brief’s talking points. A person decides what to say.
Weekly review
Each week Claude summarizes the replies, with direct quotes, for a person to read. It suggests at most two changes, suggests none when there are fewer than three replies, and never edits its own instructions.
How it is tested
Automated test suites run in GitHub Actions each time code changes on the main branch and on every proposed change. They run offline, with no live accounts, and cover how companies are scored, the drafting rules, the review pass, the sending limits, and the stop switch. Contract tests check that the schedule, the settings, and the database match what the code expects. A weekly rehearsal runs the daily jobs without sending anything.
Inside the system, every AI-drafted message goes through the review pass before it is saved. A draft saved while that pass could not run is marked as unchecked.
Tools, status, and who built it
- Tools
Python, TypeScript, Next.js, React, Tailwind CSS, Supabase (Postgres), GitHub Actions, Vercel, Anthropic Claude, Slack
- Status
Live
- Who built it
Stratyx Ventures
Talk to our team about your project
Whatever you’re building or fixing, you’ll leave the first call with a straight answer on what it would take.
After the call, AI work starts with a Workflow Review for $2,500, and a site or an app with a written scope you approve.