A chatbot told a military analyst the ship's cargo was weapons-grade.
The analyst's next move was to have the same chatbot format the findings into an official-looking summary and send it up the chain.
Polished presentation is not verification.
We keep finding the same pattern: the more AI reasoning someone is shown, the more they trust it... the less they actually read it!
At Count we've been taking some time to work out what that means and how we solve it. Follow us for future announcements.
Python in Count just got a major upgrade:
✅ Install almost any package
✅ Work with as much data as the job needs — not just what a browser tab can hold
✅ Let the Count agent write, run, and review the Python for you
✅ Schedule it, refresh it, and fire alerts
✅
Managing agents and context across a whole org is hard. This week we shipped three ways to make Count's agent your own ↓
→ Context for Connections — teach the agent about a specific data source. No semantic layer required.
→ Context via API — manage context from dbt, git,
Count now connects to any #MCP server as a data source — same as plugging in a database.
That means your support calls, docs, and the context scattered across your stack are suddenly things you can actually analyse. Slack, HubSpot, Stripe, Linear, Notion, GitHub, Intercom,
