365 Business Finance — AI transaction categorisation that learns from underwriter decisions — deterministic rules plus Amazon Bedrock for accuracy at scale.

Transaction categorisation at scale requires underwriters to read, interpret, and classify thousands of records manually. Manual overrides are common, time-consuming, and represent valuable institutional knowledge that is lost once the decision is made. The business needed automation that could replicate that knowledge, process transactions reliably and in order, and improve as it received feedback.

• Two-stage architecture: a deterministic rules pass handles high-confidence categories first, with Amazon Bedrock classifying the remainder
• Anthropic Claude on Amazon Bedrock interpreting free-text transaction descriptions for nuanced, context-aware categorisation
• Amazon SQS FIFO queues ensuring transactions are processed in order and exactly once — critical for financial accuracy
• Underwriter review and override interface — every correction is captured and feeds back into rules and prompts to improve future accuracy
• Secure data ingestion via API or batch upload to Amazon S3 with encryption in transit and at rest
• FCA-aligned data governance maintained throughout processing and storage

• Built the deterministic rules engine as the first categorisation pass — Lambda-based matching applied to transaction descriptions for high-confidence categories.
• Integrated Amazon Bedrock with Anthropic Claude as the second pass — classifying transactions that rules alone could not resolve.
• Implemented Amazon SQS FIFO queues to guarantee ordered, exactly-once processing of every transaction.
• Built the underwriter override interface — corrections are captured and fed back to refine rules and model prompts over time.
• Provisioned secure infrastructure: encrypted storage, encrypted transmission, IAM access controls, and FCA-aligned data governance.
• Validated accuracy and performance under production-representative transaction volumes.

• Transaction categorisation automated — underwriters review exceptions rather than processing every record manually.
• Categorisation accuracy improves over time — every underwriter correction sharpens the rules and model prompts.
• Ordered, exactly-once processing in place via Amazon SQS FIFO — no duplicated or out-of-sequence transactions.
• FCA-aligned data handling throughout — secure, encrypted, and auditable.
• Underwriters freed from routine categorisation to focus on complex decisions that require human judgement.