Use case · AI development, software
Anonymised, like all our use cases
The AI learns from the archive.
For an AI in customer service to give good answers, it needs verified knowledge. For an online retailer with several brands we built an application that gains this knowledge from answered requests. The AI makes proposals and removes customer data, the team lead decides what applies.
Deployed in
Customer service, several brands
Services
AI, software
Source
Answered requests from the archive
Approval
Team lead, with versions
A single case becomes knowledge.
The AI reads an answered request, recognises brand and concern and writes question and answer in general terms. Names, numbers and amounts are removed. If something personal is still there, the system rejects the text.
How new knowledge is created.
- Collect When an answered request is archived, the service desk offers to pass the solution to training. Or the AI works through the whole archive.
- Generalise A model recognises brand and concern and phrases question and answer without reference to the individual case.
- Group Identical concerns end up in one shared proposal, with every source ticket.
- Approve The team lead reviews, edits if needed and accepts or rejects.
- Publish New knowledge only applies once published, for all agents at the same time. Older states can be reactivated.
0
ticket texts are kept, only question and answer
1
knowledge state applies to all agents at once
1 click
back to an older state
2
roles: edit and approve
Where training pays off.
An overview shows per brand and concern how much knowledge there is and how often the AI had to hand over to a person, with the reason. Where there are handovers but no knowledge, training pays off first.
One application, six areas.
For everyone who maintains knowledge, and for the team lead who approves it.
Automation
Proposals
One inbox for new knowledge from the service desk and the archive, with brand, concern, question, answer and source tickets.
AI
Archive evaluation
The AI works through the archive, with progress. If a run stops, the next one continues where it left off.
Software
Questions and answers
Knowledge per brand and concern, linked to the tickets it came from.
Software
Templates
Per brand and concern, a frame the reply is written into.
Platform
Versions
Publish with a note, list of all states, roll back with one click.
Software
Coverage
Knowledge and handovers per brand and concern, gaps highlighted.
What makes the difference.
People decide
The AI proposes, only what the team lead approves is accepted.
Customer data stays out
A check finds names, numbers and amounts, including in the greeting, before a text enters the knowledge.
Robust against interruptions
Every evaluated request is marked. If the model fails, work resumes seamlessly later.
A way back built in
An older knowledge state is active again with one click, the working state is kept.
What we built it with
- .NET 10
- Blazor
- .NET Aspire
- Microsoft.Extensions.AI
- PostgreSQL
- Redis
- Docker
- GitLab CI
- Ansible
- OpenTelemetry
What does your team know that the AI doesn’t yet?
In an intro call we look at what knowledge sits in your mailboxes and how it can help your service.