Telecommunication · Call centre operations
Call notes written by the system, not by the agent
arvato’s call centre agents spent a significant part of their day writing up a summary after every customer call. As an MVP, we built a speech-to-text and summarisation pipeline that produces a standardised note for each call automatically, in Polish, so the agent does not have to write it by hand.
Key results
Client: arvato
Part of the Bertelsmann group, a leading international service provider with call centre operations across several markets.
- Industry
- Telecommunication
- Use case
- Automated call summarisation
- AI approach
- NLP with custom ML models
- Systems
- Call recording and CRM
- Language
- Polish
- Engagement
- Minimum viable product (MVP)
In short
Hundreds of hours of calls, and nothing usable came out of them
- After every call an agent wrote up the summary by hand: the customer’s issue, what was discussed, what happens next.
- That was slow, and the quality varied from agent to agent, which made it hard to track trends, hand context between teams or use the data for anything.
- The pipeline transcribes the call and writes the summary in arvato’s own note format: customer intent, issue category, actions taken, next steps.
- Spoken Polish is the hard part. Out-of-the-box language support is thinner than for English or German, so the transcription needed custom post-processing to hold up.
The starting point
The challenge
Agents spent a significant share of their time on administrative note-writing rather than on the next customer, and the summaries that resulted were inconsistent enough to limit their value downstream.
arvato’s call centre handles a high volume of customer interactions every day. After each one, an agent had to write up a summary capturing the issue, the discussion and any follow-up action. It was slow, it was inconsistent, and it created an information bottleneck.
Hundreds of hours of call data were generated daily, and getting structured, useful information out of it was entirely manual.
Because the quality varied from agent to agent, it was difficult to track trends, share context between teams, or use call data for operational improvement at all.
The build
What we built
We developed an end-to-end pipeline that takes a raw call recording and produces a structured, standardised summary, with no intervention from the agent.
Getting spoken Polish into text
The first stage turns the audio into text, and spoken Polish brings its own problems: varying audio quality, interruptions, colloquialisms, and thinner out-of-the-box language support than English or German. We built custom post-processing to raise transcription accuracy and handle those edge cases reliably.
A summary in their own format
Once transcribed, the system writes a concise summary following arvato’s internal procedures and note-taking format. Custom models extract customer intent, issue category, actions taken and next steps and present them in the same structure no matter who took the call.
Built for peak hours
The pipeline runs on cloud infrastructure so it scales with demand, processing more calls during the busiest hours without slowing down.
What changed
The results
Before
Agents wrote a summary by hand after every interaction. Quality varied by person, and the call data was barely used for analysis or process improvement.
After
Summaries generated automatically in one standard format. Agents move to the next call immediately, and the call data became a structured, searchable asset.
The MVP shows that the manual note can be replaced: every call produces a consistent, structured note regardless of agent, call complexity or time of day.
That puts agents back on customer interactions instead of administration, which is the part of the job that only a person can do.
Beyond the time saved, the standardised summaries opened up analysis that was not possible before: identifying recurring issues, tracking resolution patterns and improving processes from actual call data rather than fragmented manual notes.
Questions about this project
Automated call summarisation. Agents spent a significant share of their time on administrative note-writing rather than on the next customer, and the summaries that resulted were inconsistent enough to limit their value downstream.
Call summaries written automatically, built and tested as a minimum viable product: MVP. Custom models trained for Polish language processing: Polish. Scalable infrastructure for high call volumes: Cloud. One note format across every agent and team: Uniform.
NLP with custom ML models. Technology used: Natural Language Processing, Machine learning, Speech-to-text, Custom Polish language models, Cloud computing, Post-processing pipelines.
Technology used
Talk to the people who built this
theBlue.ai comes back within one business day.
