Firstcom Europe – AI Call Analysis on AWS

About Firstcom Europe

Firstcom Europe provides telecommunications services to small businesses across the United Kingdom and Germany — cloud-hosted phone systems, business mobile and connectivity. StableLogic designed and operates the AWS platform behind it, handling over 100,000 calls a day.

Every one of those calls carries information Firstcom’s customers would value: why someone rang, how the conversation went, whether the caller left satisfied. Until recently, almost none of it was being captured.

The Challenge

Firstcom operates a large and highly successful voice platform for its customers. Customers depended on listening to calls and manual sampling to measure quality and outcomes – a small handful of calls listened to by a person, standing in for thousands that nobody would ever hear.

That creates a familiar set of problems. Sentiment goes unmeasured, so a deteriorating customer relationship only surfaces when it is already lost. Reasons for calling are guessed at rather than counted, so resourcing and self-service decisions get made on instinct. Quality issues are caught only if they happen to fall inside the sample.

Scaling the manual approach was never realistic. Reviewing a meaningful proportion of a large daily call volume by hand would cost more than the insight is worth. The answer had to process every call automatically, or it was not worth building.

The AWS Solution

StableLogic built an event-driven pipeline that analyses each call as soon as it finishes, and surfaces the results to Firstcom’s customers inside the portal they already use.

How it works

When a call recording completes, the voice platform triggers a Go service running on Amazon ECS over an internal HTTP endpoint. That service transcribes the audio using Amazon Transcribe, stores the transcript in Amazon RDS for PostgreSQL, and then calls Claude on Amazon Bedrock with a structured prompt and a defined output schema.

Claude returns a consistent set of fields for every call — sentiment, reason for the call, a short summary, and the topics discussed. Those results are written back to the database, and Firstcom’s existing customer portal reads them directly. No new interface was introduced, and no change was required to the live voice platform. Recordings are held in Amazon S3.

Why these services

Amazon Transcribe was chosen for speech-to-text because it is a managed service that scales with call volume, with no transcription infrastructure to operate. Claude on Amazon Bedrock was selected for the analysis layer because it handles long transcripts comfortably and returns reliable, schema-conformant structured output that can be stored and displayed without post-processing. Neither requires model hosting, and both are priced per use, so cost tracks call volume directly rather than sitting as fixed infrastructure.

Amazon ECS suits the shape of the work: asynchronous, event-triggered and batch-like, with volume that varies through the day. Because both transcription and inference are fully managed services, the ECS service is a lightweight coordinator rather than a heavy compute workload.

Analysis only, by design

The agent has no tools that can take actions. It cannot send a communication, modify a Firstcom system, or act on what it finds. It reads a transcript and writes structured analysis to a database — nothing else.

That constraint was deliberate. It means the responsible AI questions on this solution are about accuracy and presentation rather than about harmful action, and it removes an entire category of risk at the design stage. The service is not exposed publicly: it is invoked service-to-service through an internal load balancer, and its ECS task role is scoped to only Amazon Transcribe, Amazon Bedrock, the S3 recordings bucket and the RDS database. Analysis shown to customers in the portal is clearly labelled as AI-generated.

Testing and refinement

Before release, sentiment, reason-for-call and summary outputs were validated by human review against the underlying calls. Safety testing specifically covered instructions embedded in call content, confirming the model treats transcript text as data rather than as commands.

In production, processing and model outputs are logged to Amazon CloudWatch. Customer feedback reaching Firstcom’s support team is used to refine the analysis prompt and output schema where categories or wording did not match what customers expected.

The Results

• 100% of completed calls analysed automatically, replacing a small manual sample
• Manual review effort per call reduced by around 80%
• Insight available to Firstcom’s customers shortly after each call ends, inside the portal they already use
• No public endpoint and no action tools, so the solution added no new external attack surface
• No change to the live voice platform, and no new infrastructure to operate — transcription and inference are both managed, pay-per-use services
• EU data residency maintained throughout