A 24-minute build walkthrough from Ben AI (Ben van Sprundel) on the AI sales system his team runs, to the point where reps "are barely in the CRM anymore." The core argument: most people use AI to do sales work but never give it the intelligence to do that work well. This system automates both, and the intelligence layer is the part everyone skips.
The core idea
Sales comes down to two things: gather intelligence, then do the work off that intelligence. No lead research, no call prep. No ICP, no qualification. No metrics, no strategy. Most AI-for-sales effort goes into the second half (skills that draft emails, prep calls) while starving the AI of the first half. Feed Claude the real, current intelligence and the same skills produce far better output. This system automates both halves, and the infrastructure is a one-time build.
The architecture
Five layers, bottom to top:
| Layer | What it is |
|---|---|
| Raw data | Your sales software (CRM, transcription, proposals) via connectors |
| Intelligence layer | Scheduled routines that filter the raw data and push only what matters up |
| Sales second brain | The intelligence store: a folder of Markdown files |
| Sales dashboard | The rep's control center, built on top of the brain |
| AI agent | Skills and routines that act, then feed results back into the brain |
The second brain sits in the middle. Software data flows up into it daily; the dashboard and the AI both read from it; the AI's output flows back down into it.
The three components
1. The sales second brain
Not fancy: a folder of Markdown files, viewed in Obsidian (a free app that visualizes the folder and the links between notes). Subfolders:
- Calls - each transcript plus a summary and a call-performance score.
- Contacts / core docs - the permanent context: ICP, an "about me and the business" doc, offer, positioning, sales process, tone of voice.
- Deals - a running history per lead. "The CRM, but for AI."
- Intelligence - monthly and quarterly reports.
- Campaigns - every outreach campaign with its metrics, results, and sequence.
- Skills and templates - email templates, NDA, one-pager, pitch deck, proposal.
Two kinds of documents live here: real-time intelligence (updated from your software) and permanent context (ICP, offer, positioning). The context compounds: it starts as a handful of docs and grows to hundreds, so the earlier you start, the smarter it is later.
2. The sales dashboard
The rep's control center, so they never open the CRM. Tabs: Today (upcoming calls with pre-built prep briefs, qualification scores, top tasks), pipeline metrics, CRM snapshot (deals by stage), core documents (offer, ICP, objections, process), the tech stack, the running automations, and a monthly report. Built once with Claude, then rebuilt daily by a routine. Hosted either as a Claude artifact (team-only) or deployed to a public URL through Vercel (free) so anyone can view it.
3. The AI automations
Two ways the brain drives work:
- Skills for repetitive workflows: call prep, client one-pager, lead generation, LinkedIn scraping, outreach, pipeline review, follow-up, proposal generation. Because each skill reads the second brain, a proposal is built from that lead's real history and objections, not generic filler.
- Routines (event-triggered or scheduled): run call prep when a meeting is booked, draft a follow-up and proposal after a meeting ends. For one-off work, open a chat with the second-brain folder selected so Claude has full context.
How to set it up
The build, bottom to top. Budget a couple of hours.
- Connectors. Two types: your existing software, and web scraping for research.
| Category | Example | Why |
|---|---|---|
| CRM | Attio | deal data |
| Meeting transcription | Fireflies | call transcripts (start here if you have nothing) |
| Proposals | PandaDoc | tracks when proposals are signed |
| Web scraping | Apify | scrape the web for lead and company research (the key one) |
| Lead database | Apollo | prospecting data |
| Hosting | Vercel | serve the dashboard on a live URL |
Plus your calendar, email, and chat tools. In Claude Code, a Google Workspace CLI bundles Gmail, Calendar, and Sheets into one connector.
- Create the second-brain folder with the subfolders above. It can be a subfolder of an existing company second brain.
- Write the permanent context docs first (ICP, offer, positioning, about-you). These move the needle most. Have Claude interview you to draft them.
- Add a
CLAUDE.mdat the root as the routing map (what lives where, plus rules), and one in each subfolder. Claude reads it first to know where to pull and save context. - Set up daily routines. Scheduled tasks run only when the desktop app is open; Claude Code routines run in the cloud regardless. Three core ones:
- Morning routine - pulls the last 24h of calendar, email, and calls; preps today's calls; reconciles; logs.
- CRM sync - checks proposal status, cleans deal fields, freezes won/lost, flags cold deals, snapshots metrics.
- Call scoring (optional) - grades call transcripts for coaching.
- Later: monthly report, quarterly report, campaign metrics.
- Build the dashboard in Claude Code with the folder selected ("build a dashboard on this data"), deploy it via Vercel, save the HTML into the brain, and add a routine to rebuild and redeploy it daily.
Key takeaways
- Intelligence first, automation second. Skills without a fed context layer produce generic output. The second brain is the unlock, and it is the step most people skip.
- The second brain is just Markdown in folders. No special tooling. Obsidian only visualizes it.
CLAUDE.mdis the routing table. Root plus per-subfolder maps are what let Claude navigate and update the brain reliably.- Routines keep it alive. Cloud routines (always-on) beat desktop scheduled tasks for a system that must update daily.
- Context compounds, so start now. A few docs today become hundreds of compounding intelligence files in months.
Why it matters for me
This is the exact system I am building, so it is a useful reference architecture. The transferable parts: an intelligence layer that feeds every skill, a CLAUDE.md routing map, daily routines that reconcile software data, and a dashboard that replaces living in the CRM. The main divergence: Ben's second brain is Markdown in Obsidian, while mine already uses SQLite as the source of truth (btob.db, linkedin.db, pipeline data) with Markdown for the KB. So my "second brain" is a hybrid, structured data in SQLite plus prose context in the KB, and my connectors are Salesforce and Outlook rather than Attio and Fireflies. The architecture holds; the storage layer is mine to adapt. LinkedIn data stays local (never scraped), so Apify-style research runs on other sources.
Further Reading
