What we've built, what it did, and what we haven't measured yet.

The ERP held the answers — order status, open invoices, what's in stock — and the people who needed them asked finance.

A connector that runs in the client's own Azure tenant. Each person signs in with their own work account, Business Central applies that person's own permissions, and every question leaves a line in the client's own logs: who asked, which tool, whether it worked. Never the data that came back. Nobody sees through Claude what they couldn't already see in the ERP. It reads; it cannot write.

We built it with their finance lead at the keyboard, not for him behind a curtain. We wrote the server and the step-by-step runbook; he ran every step in his own tenant, so the person who'd own it understood how sign-in, permissions and the audit trail fit together.

Live since September 2026 with a small pilot group. Their finance lead now adds new tools himself, inside guardrails we wrote down — read-only, no cost fields where sales can reach them, don't touch sign-in — and we review what he ships.

The owner answered support email himself, when he got to it. No helpdesk, no order lookup beside the inbox, nothing behind the address but him.

An agent that checks the inbox every two minutes, looks the order up live in the store, and answers what store policy covers. Anything involving money — a refund, a return — goes to the owner as an email he answers with one word, YES or NO, and the agent takes it from there. Anything it shouldn't touch — a legal threat, a chargeback, abuse — stays internal. When it hands a customer over, it writes the introduction and copies the owner in.

Before launch it had to pass thirty test emails: order questions, returns, edge cases, and three attempts to talk it out of its instructions. It passed all thirty. That is a test suite on mock emails, not a production score.

We also built him a dashboard that shows the inbox and the orders at a glance.

Live since July 2026 and still checking the inbox every two minutes. Most of what lands there was never a customer: in the seven weeks the event log covers, more than nine emails in ten were bulk mail or sales pitches, and the agent filed them without anyone looking. The owner answers YES/NO emails, picks up the conversations he's copied on, and otherwise rarely touches it.

Every pitch started with a deck a rep built by hand: clicking through the prospect's website, map listings, reviews and social pages, screenshotting as they went. One audit took a rep between one and four hours, and quality depended on who built it and how much time they had.

An agent that gathers the same evidence itself — site, page speed, Google and Apple map listings, social pages, a local-search grid, whether the business's details match across all of them — and lays it out as a deck in the prospect's own colours.

The rule that shaped it came from the reps: a finding they can't point at on a screenshot doesn't sell. So every claim in the deck traces to a stored capture, and anything the agent can't back is held off the deck and listed for the rep to check instead. In one live run it caught itself about to print a follower count that appeared nowhere in what it had collected.

Built, shaped by two rounds of review with their sales team, and bought outright: their code, their keys, their operating guide. The handover happens with their next audit; no rep has pitched from one of its decks yet.