AI-powered call classification for a UK police constabulary

A UK police constabulary needed to understand the thematic content of their 101 (non-emergency) call volume at scale. Calls were unstructured audio with no systematic classification, making it impossible to identify patterns, allocate resources effectively, or measure changes in public demand over time.

The data was highly sensitive — containing personal details, crime reports, and vulnerable caller information — requiring robust PII handling before any analysis could begin.

We built a layered PII redaction pipeline combining Deepgram transcription-level redaction with AWS Comprehend as a secondary protection layer, aligned with metadata including call handler teams, call duration, and hashed caller IDs.

For classification, we developed a hierarchical taxonomy (sub-themes, themes, categories) and built an LLM-as-a-Judge classifier that could select from the established taxonomy or propose additions via a self-healing mechanism with schema-validated suggestions. To bootstrap classification at scale without exhaustive manual labelling, we generated representative call language for each theme and applied cosine similarity matching via vector embeddings.

The system achieved 60% accuracy on primary theme classification and 85% including secondary classifications (top-3), with category-level accuracy significantly higher.

The bootstrapping approach was extended to detect crime scene preservation language patterns and call handler abuse indicators. Findings were presented to constabulary leadership. A thought leadership paper was co-authored with PwC and presented at the Transforming Police Public Contact conference.