Client case study
A conversational AI cockpit for B2B steel sales
Built for ArcelorMittal Distribution Solutions: a conversation wired to product data, configuration, quoting and after-sales follow-up.
−70%
measured processing time
The starting problem
In steel distribution, a customer request rarely maps to a single reference. You have to interpret a need expressed in everyday language, find the matching grade and dimensions, check whether a cut is feasible, then produce a quote. Each step lives in a different tool.
The cost is not in any one step, it is in the back and forth: switching screens, re-typing, double-checking. That is the time that was measured, and that is what we targeted.
What was built
A conversational agent that understands a complex request, identifies the exact reference and even configures an angled cut, in 3D, inside the conversation. The user states their need the way they would to a colleague; the agent helps them build a workable answer without switching tools.
The same cockpit then covers quoting, ordering, delivery and invoice lookup. This is not an assistant sitting next to the business process, it is the business process.
The demonstration
The technical choices, and why
What separates a demo agent from one teams actually use.
MCP-UI interface inside the chat
Conversation alone is not enough to pick a grade, dimensions and a finish. An interactive interface appears right in the thread and guides the configuration, without pushing the user to an external form.
Wired to the real data
The agent queries the existing product catalogue and configuration rules. It does not guess a reference, it retrieves one, which is the condition for a sales rep to trust it on a quote.
Full cycle coverage
Quotes, orders, delivery and invoices go through the same entry point. That is what removes the back and forth between tools, and therefore most of the time saved.
The result
Measured processing time on the covered journey dropped by 70%. The gain does not come from one magic step but from removing the breaks between tools.
A side effect that matters just as much: less re-typing means fewer human errors on quotes where a wrong grade or dimension is expensive.
A need close to yours?
This project is industrial, the mechanics are not: wire a conversation to real business data, and let the user state their need rather than fill in fields.
We start from the task costing you the most time, scope it together, and you receive a proposal before any commitment.
- Search across a complex catalogue
- Assisted product configuration
- Quote generation from a free-form request
- Unified customer follow-up after the sale
Frequently asked questions
- The reduction in processing time measured on the journey covered by the agent, from the initial request through to the quote. It is a measurement taken on the engagement, not a sales projection.
- The risk exists as soon as a model answers without a source. That is why the agent is wired to the real catalogue and configuration rules: it retrieves an existing reference instead of producing one.
- No, the approach connects what exists rather than replacing it. That is usually faster, less risky, and it avoids replaying a migration.
- A targeted automation ships in a few days, a complete application takes several weeks. The schedule is set during scoping, with delivery in useful stages rather than one big switch-over.