An internal AI assistant running against a live HubSpot CRM: it generates German account briefings via Claude, surfaces stale deals and data gaps, and prepares tasks and follow-up drafts. It prepares the work — a human makes every decision.
A lead comes in and nobody has time to research the company before the first call. A deal goes silent for two weeks and nobody notices. The CRM slowly fills with contacts missing phone numbers, job titles, lifecycle stages — and every gap is a conversation that starts blind.
The client — a small Swiss AI consultancy running on HubSpot — didn't need more automation sending things. They needed something that prepares work: research done before the call, silence made visible, gaps flagged before they hurt.
For every new lead, SalesPilot gathers CRM context plus the company homepage and generates a compact German briefing via Claude — signals, conversation openers, and what's missing. Written back into HubSpot as a note, marked as AI-generated.
Open deals with no activity beyond a configurable threshold surface automatically — deal, stage, days silent, owner. Computed live on every request, never stored.
Contacts and deals missing required fields show up as chips: exactly what's missing, per record. Every gap is a data-quality conversation waiting to happen.
Every element in SalesPilot belongs to one of two worlds, and the design never lets them blur. Machine work is labelled machine work — briefing cards, research traces, suggested actions. Human judgment gets its own moments — review, approve, snooze, create a task.
The system drafts, flags and creates tasks. It sends emails, contacts anyone, changes deal stages — structurally impossible, not switched off. A capability ceiling, stated plainly, builds more trust than any feature list.
The most important engineering decision was making the AI admit ignorance. The briefing prompt uses only the provided CRM data — nothing else — and where information is missing, it must literally write «keine Angabe».
That's not a limitation, it's the product insight: an AI that invents nothing becomes a mirror for the CRM's health. Every «keine Angabe» is a data-quality finding — and the gaps panel turns those findings into a to-do list.
SalesPilot is in active development against a test HubSpot portal, with the client pilot scheduled for Q3 2026: the full briefing pipeline, both detection panels, review actions, and a daily automated scan. Three things this project has already taught me:
"This system technically cannot send email" became the strongest line in the client conversation. A ceiling, stated plainly, beats a feature list.
Marking every AI output as AI-generated costs nothing to build — and answers the hardest question about AI tools, who did this?, on every screen.
The most valuable output of an AI research assistant isn't what it finds. It's how clearly it shows what nobody bothered to write down.