Medical supplies distributor proves AI email answering in five weeks

Situation

A distributor of medical consumables in Mainz, 50 staff and about €12 million in annual revenue, supplies practices and care homes. Its customer service team of six receives about 400 emails a week: is this in stock, what does it cost, where is my order, which pack size fits. Each one was answered by hand after looking up the catalogue and the order system. First response took about one working day, and the team knew it was losing orders to faster competitors. Management had a vendor quote for a €60,000 customer service platform and wanted proof before spending that.

Approach

We agreed the narrowest useful pilot: one workflow, one data source, an off-the-shelf model. The workflow was the incoming customer email. The data source was the product catalogue, about 3,000 items with prices, pack sizes and stock flags, exported every night.

We built a small service that reads each new email, sorts it into one of five types, looks up the product in the catalogue, and writes a suggested reply into the team’s existing mailbox as a draft. Nothing is sent until a person reads it and clicks. Order status questions need the order system, a second data source, so they were kept out of scope and routed to a person with a label. That kept the pilot within budget and made the result easy to read.

In week one we checked the data protection position: the emails contain customer contact details and product questions, no patient data. The model runs in the EU under a data processing agreement. In weeks one and two we tuned the prompts on 200 historical emails. Week three ran in shadow mode. Weeks four and five went live, with a shared scoreboard the team could see.

Result

In this scenario the team handled about 1,200 emails over three weeks of live operation.

About 70% received a draft that was sent with small or no edits. About 20% had a draft that needed rework, and about 10% were routed straight to a person. First response time fell from about one working day to under two hours, and the team estimated 12 to 15 hours a week freed across six people.

Management approved the rollout on those figures. They chose to extend the pilot to order status questions rather than buy the €60,000 platform. The pilot closed with a written report: the numbers, the failure cases, and what the next step needs.

What it cost

The AI pilot cost €9,500, quoted as a fixed price before work started. That bought the workflow design, the data protection check, the service built on the catalogue, prompt tuning, shadow mode, three weeks of measured live operation, and the closing report. The model usage of about €40 per month is a third-party cost and is excluded. The work was covered by our 100% money-back guarantee: full refund on request within 14 days of delivery.

Every fixed-price engagement carries the 100% money-back guarantee. How the guarantee works →