A real-time sales CRM that people and AI share

The problem

Prospecting rarely fails for lack of effort. It fails because nobody has the full picture. Leads are scattered across spreadsheets, sticky notes, and people’s heads. Two people end up with two versions of the truth. A promising contact slips through the cracks because nobody was watching the follow-up date. A small team working the same accounts needs a single shared source of truth that’s always up to date. And since AI assistants work the pipeline here too, not just people, that same source had to be accessible to both.

What we built

A login-protected, real-time CRM, built as a lean web app. It has a kanban board with a column for each sales stage and a follow-up radar that pulls due and overdue contacts to the top. There’s a probability-weighted revenue funnel, a fit score for every lead, and letter and email generators that prepare first drafts, which a person reviews before anything is sent. Every lead record holds the research findings, a specific pitch angle, and an editable email draft. You open the draft in your mail app with one click, then log it as sent. Contacts who said no can be marked “dead,” “on ice,” or “try again in two months.” That last one automatically puts them back on the follow-up radar 60 days later.

Under the hood

The backend is a Postgres database on Supabase, hosted in the EU. Access is locked down at the database level with Row Level Security and an email allowlist: only approved accounts can see any data, even if a stranger somehow managed to create an account. The frontend is deliberately a single file of plain JavaScript, with no framework and just one small library loaded at runtime (supabase-js). It is generated deterministically from one source file, so every version can be reproduced exactly. A git push takes it live in seconds. The app was built with Claude Code.

The trickiest part of any shared tool is people working at the same time. Changes sync live across every open window. An echo guard stops your own save from coming back as someone else’s change, and a stale-data guard catches outdated form data, so two people don’t overwrite each other’s changes. A failed save isn’t lost, either: it’s retried automatically.

AI works the pipeline too

What sets it apart is who has access. The team’s AI assistants work on the database through the same API the app uses. They research a lead on the web, write the sourced findings and a tailored pitch angle straight into the record, and drop a complete email draft into the lead’s draft field. A human checks it, edits it, and sends it. That makes the CRM less of a filing cabinet and more of a pipeline where people and AI split the work, and a human always has the final say.

Before we trusted it, we tried to break it

A tool that handles real contact data and runs cold outreach can’t afford embarrassing mistakes. Before going live, we set a swarm of independent AI reviewers loose on it, each tasked with breaking a different workflow. Then a second, adversarial wave tried to disprove every finding. Sixteen real bugs survived, including text that silently vanished when you switched panels, edits lost when a sync failed, and false alarms in the live sync. We fixed all sixteen, then verified the fixes end to end in a real browser on a throwaway record.

Why we built it this way

• EU hosting plus an allowlist: The CRM manages personal contact data. Data in the EU and access control at the database level are a must here, not a nice-to-have.
• One file instead of a framework: No build toolchain to babysit, no dependencies going stale, and still easy to maintain years from now. The deterministic build makes every version reproducible.
• Real time with guardrails: Shared tools break down when people overwrite each other’s changes. The echo guard and the stale-data guard prevent exactly that.
• Human sign-off: AI can research and draft, but nothing goes out until a human has checked it.

Where it stands today, and what’s still missing

The CRM is running in production and has grown from around eighty to several hundred leads, all maintained by people and AI together. Still to come: incoming emails will show up as a timeline right on each lead, not just in the inbox. Everything else on this page is real and running today. If you’d like to see it in action, we’re happy to show you how it works on a call.