Use case · AI development, software, integration
Anonymised, like all our use cases
The AI replies. Critical cases stay with the team.
For an online retailer with several brands we built an AI agent for customer service. It reads every incoming email, looks up order, shipping and payment in the ERP and writes the reply from approved knowledge. If money is involved, the tone is sharp or it is not sure, it hands over to a person.
Deployed in
Online retail, several brands
Services
AI, software, integration
ERP connection
16 queries, read-only
Operation
Shadow mode and live
The agent replies. The team stays in control.
For every request the agent identifies the concern, fetches the matching data from the ERP and writes a draft. Before anything goes out, it checks the draft for unsupported statements and open questions. In shadow mode everything runs, only nothing is sent.
How a reply comes together.
- Understand The AI identifies the concern and checks up front: is the intent clear, does someone want money back, is the tone too sharp? If so, the email goes straight to a person.
- Look up Customer, order, shipping and payment come straight from the ERP, via the Model Context Protocol and read-only.
- Write The reply is built from the brand’s approved knowledge and a template for the concern.
- Check A safeguard looks for unsupported claims, unanswered questions and manipulation attempts. If it finds something, the agent revises the draft.
- Decide The agent sends the reply or hands the ticket to the person responsible, with a note and findings.
16
ERP queries, all read-only
3
checks before every send
3
reasons to hand over to a person at once
0
emails in shadow mode, yet the agent runs end to end
Every run can be traced.
The log shows step by step what the agent recognised, looked up and decided, with the incoming email and the draft alongside. When it hands over to a person, the reason is right there.
One application, six areas.
Everything an agent needs, on the same platform as the service desk, with one login.
AI
Agents
One agent per mailbox, with its own model for every step, thresholds and texts. Shadow or live at the flick of a switch.
Software
Run log
Every run as a timeline with incoming email, draft and raw data, filterable by outcome, mode and date.
Integration
ERP queries
For each concern, administrators decide which ERP queries the agent may use.
AI
Knowledge and templates
Questions and answers per brand and concern, plus templates that frame the reply. Agents only read the published state.
AI
Models per step
Intent detection, reply writing and the safeguard can each use their own model.
Software
Test runs
Test sets with expected outcomes compare settings before they go live.
What makes the difference.
Critical cases stay with people
Refund requests, a sharp tone and uncertain cases always go to the team, with a note and findings.
Shadow first, then live
The agent runs along without sending. It only goes live once the logs are convincing.
Read-only in the ERP
The agent looks up orders, shipping and payments. It cannot change anything there.
Rules per concern
Reply texts, thresholds and permitted ERP queries can be adjusted per agent and per concern.
What we built it with
- .NET 10
- Blazor
- .NET Aspire
- Microsoft Agent Framework
- Microsoft.Extensions.AI
- Model Context Protocol
- PostgreSQL
- Redis
- Docker
- GitLab CI
- Ansible
- OpenTelemetry
Which requests should your AI prepare?
In an intro call we look at which concerns can be answered automatically and which should stay with your team.