Customer story · Edera

A key role refocused on expertise, not data entry: the Edera case

At Edera, a maker of custom dental prostheses, one person, sometimes more, is assigned to the technical questions raised by the manufacturing labs.

Challenge

Day-to-day technical exchanges with production partners are set up so that each email covers one specific case on one specific topic. That can span a fairly wide range: a missing piece of information, an instruction to clarify, a shape to approve. Some can be handled fairly quickly, without involving the client concerned, through practice and habit alone. The answer is sometimes right in front of us, and that relentless flow of emails over a day can become a burden.

The result: one person tied up in repetitive work whose added value varies from one email to the next, a sizeable share of which “anyone” could handle as long as it is pure data.

The problem is that every question takes the case out of the production flow, so it cannot follow its optimal route. The goal is to filter out a large share of the questions and keep only those with high added value.

What we built

An AI agent connected to the mailbox, running on Claude, with access managed through Azure AD.

It reads every incoming email and classifies it. It fetches context from the related documents: purchase order, patient record, scans. We set out to teach it human reasoning.

  1. The question comes in.
  2. Can I answer it?Yes: as a rule, if a person can answer without arbitration, the AI potentially can too. If not, the question gets a red flag: I have a specific task to do on it.
  3. The red flag turns yellow.If the task involves a validation request or input from the client, I prepare a message and send it to the dentist. Yellow flag: I have asked my question and do not have the answer yet.
  4. The topic is closed.Once the dentist has answered, I reply to the lab and remove the flag.

Without a conclusive result yet, we kept it at this observation stage: we do not want to let the AI loose while its reasoning is not right.

To help the AI improve, we split the screen in two: the human answer and reasoning on one side, the AI’s answer and reasoning on the other, so the agent can learn on its own.

Where the project stands

  • A few daysof development

    The agent was developed over a few days.

  • A few tens of eurosa month for the whole flow

    On cost, splitting triage and drafting across two separate models keeps the monthly bill in the region of a few tens of euros to absorb the entire flow: an order of magnitude that bears no comparison with the cost of a dedicated full-time role.

  • A significant shareof the time redirected to arbitration

    On usage, the person who used to spend most of the day in the mailbox now sees a significant share of that time redirected to arbitrating the cases that genuinely need their expertise.

This is work that keeps being tuned with the business: the confidence threshold and the quality of the drafts improve with feedback from the reviewers, and adoption by the internal teams is still under way. At this stage it is a field report, not a final assessment.

How we keep the cost down

A controlled monthly cost at this volume does not come from a well-written prompt. It comes from three trade-offs.

  • The right model in the right place

    Triage and parsing go through Claude Haiku: fast, economical, good enough to structure incoming mail. Drafting goes through Sonnet, where the quality of the reasoning shows. On this project, triage handles the largest volume of calls for the smallest share of the cost.

  • What should stay code stays code

    Any deterministic logic handed to the LLM is a pointless recurring cost. That call is made at design time, not after the bill arrives.

  • Structuring upstream

    Preparation work is paid for once; tokens are paid for on every call.

The method

Three phases.

  • Scoping

    Mapping the real processes, not the documented ones, and defining the KPIs with the business.

  • Build

    Building the agent, the confidence score and the review interface, designed to learn from the reviewers’ corrections.

  • Monitoring and handover

    Performance tracking over time, regular adjustments, then a skills transfer to the internal teams so the client can run its agent on its own.

Technology

Claude Haiku for triage and parsing, Claude Sonnet for drafting. An agent connected to the mailbox, with access managed through Azure AD.

What this case says about the rest

Edera is not an isolated case. In many companies there is at least one role where a skilled person spends most of the week moving technical information between systems that do not talk to each other: neither an industrialisable process nor a real decision. That is where AI has the most value to add, provided the decision and the accountability stay on the human side.

The question is not whether AI can do this work: on this kind of task, it can. The real question is time to production, recurring cost, and the level of control kept over what goes out. That is the trio this case illustrates, with its encouraging results and the adjustments still to be made.

Do you have a role that looks like the one at Edera? Let’s talk.

Koneetiv is an official Anthropic partner in France, working exclusively with Claude. We help companies adopt AI and transform their business with it.

Get in touch

Humans set the course. AI scales it up.