Orders & Analyst agent

A mid-market manufacturer and distributor running its orders and customer records in a CRM/ERP. The data was all there and almost none of it was reachable.

Answering "who are our biggest clients", "where does most of our business come from", "what did we quote this customer last year" meant someone building a report.

Producing an estimate meant reading a customer's document, hunting for the matching spec, checking pricing, working out landed cost, and applying margin by hand.

Specs and documents were scattered across drives and inboxes.

An agent with Slack as its interface and the CRM/ERP as its backend, so there was no new software for anyone to learn.

Sales now forwards vendor quotes to the agent who drafts the full customer estimates with margin and related operational costs.

It compares a customer's order against the spec and reports the gaps. It queries the web for better pricing and answers logistics questions including country import duties.

And it converts a document in any format into a complete estimate with margin applied.

The measure that matters is not hours saved but share of the work: how much of the live workflow the agent now carries inside the client's own system of record.

In its first eleven months writing estimates, the agent created 69% of every estimate in the system — 786 of 1,135. Part and item records, previously typed in one at a time, run at 89% across the first fourteen months — 2,596 of 2,930 — and have held in a 89–97% band every month since the third. These are not survey responses or time-and-motion estimates. They are counts of every record created in the live ERP, agent and human alike.

Over the same period the business grew revenue by around half year on year, with headcount growth well below what that volume would ordinarily have required. The same commercial team absorbed the growth.

Not every path transferred. Opportunity creation was taken up in the first six months and handed back in the second, while opportunity volume held steady — the manual task was already cheap, and the agent did not earn the work. Adoption tracks the toil. We instrument for that rather than assume it.