Day 37: Three Calls, Three Countries
Three calls today. Austria, Ireland, Germany. Three completely different industries. Every single one confirmed the same core thesis: companies want AI agents, but they need guardrails before they'll deploy them.
The Quote Generator
First call: an Austrian manufacturing company that sells fencing products. They run everything on Odoo — sales, inventory, helpdesk, purchasing. Their sales team writes 10-15 quotes per day, each one requiring: look up the right products, figure out quantities, check variants, calculate totals, create the quote in Odoo.
The ask: "Customer needs 40 meters of fence, 1.23m high, in green" → agent looks up matching products in Odoo, selects the right variants, calculates quantities (panels, posts, fixings), and creates a draft quote. Human reviews, adjusts if needed, sends.
This is exactly the kind of use case where AI agents shine. Not because it's intellectually complex, but because it's repetitive, time-consuming, and follows a pattern. The agent doesn't need to be creative — it needs to be thorough and fast.
What makes this particularly clean: everything lives in Odoo. No cross-system integration needed for the first version. Products, prices, customer data, quotes — all accessible through one API. I'm building a Pinchy plugin with seven Odoo tools that cover the basics: read products, search orders, create quotes, manage helpdesk tickets.
They also want support ticket drafts — incoming customer emails get a suggested response based on product documentation and order history. But quotes come first. That's where the time savings are biggest.
The Fleet Company
Second call: a fleet management company in Ireland. They build a platform where transport companies see all their vehicle data — telematics, compliance, efficiency, driver performance. Their solution engineering team spends hours each morning pulling context together for customer requests and upcoming meetings.
The interesting part: they're already using NanoClaw individually. Their CEO has a custom agent that prepares meeting briefings — pulling company info, LinkedIn profiles, fleet size, compliance records. But it doesn't scale. Individual NanoClaw instances can't share skills, can't share memory, and there's no access control.
Their questions were exactly the ones Pinchy answers: How do we harden this for multiple users? How do we share skills across the team? How do we make sure agents can't access data they shouldn't? How do we track what they're doing?
They also raised a smart point about webhooks: instead of sending the full payload to the agent (risky — prompt injection), send only an event ID, then let the agent pull the data back from a trusted source. Like how Stripe webhooks work. I'm going to steal this pattern.
The Pattern
Three calls, three industries, same needs:
- Connect to the systems we already use — Odoo, Outlook, internal APIs. Nobody wants to migrate data.
- Draft, don't execute — every single company wants human-in-the-loop. Create the draft, suggest the response, prepare the quote. Let a human press send.
- Show your work — the audit trail isn't a compliance checkbox. People actually want to see what the agent did and why.
- Start small, prove value, expand — nobody wants a three-month integration project. They want one agent doing one thing well, this week.
I also built the integrations UI today — a new settings tab where admins configure external connections (starting with Odoo) and a per-agent permission matrix for which connections each agent can use. Same allow-list philosophy as the tool permissions, applied to external systems.
Day 37
An Austrian manufacturer. An Irish fleet company. A German corporate startup (from yesterday). Three countries, three industries, one product. The use cases are different, but the requirements are the same: connect to what we have, don't break anything, and let us see what's happening.
That's the product I'm building.