Swarm Cuts Invoice Processing Time by 70% with Agentic AI on AWS

The customer is a mid-sized enterprise in the professional services sector, operating across multiple subsidiaries in Southeast Asia with a lean finance team of five to eight people.

It processed around 10,000 supplier invoices a month almost entirely by hand, with every invoice requiring manual data extraction, validation and approval routing. This created processing delays of several days per batch and tied up the equivalent of three full-time staff in repetitive work.

Invoice formats varied widely across vendor templates, mixing digital PDFs, scanned documents and multi-currency invoices with no standard layout, so data inconsistencies carried through into financial reporting and visibility across subsidiaries stayed limited.

Existing third-party OCR tools could not interpret invoice content in context, and routing financial data through external SaaS providers raised data security and sovereignty concerns. Without intervention the customer faced rising operational costs, weaker financial control and an inability to scale finance operations in line with business growth.

Swarm engaged Dorje AI as the specialist Hive team to design and deploy a serverless, agentic invoice automation solution inside the customer's own AWS account in the Asia Pacific region.

Invoices arrive by email through Amazon SES or are uploaded through a single page application served by Amazon CloudFront, and are written to a versioned, KMS-encrypted Amazon S3 bucket. The S3 object event reaches Amazon EventBridge, which starts an AWS Step Functions Standard Workflow orchestrating roughly twelve AWS Lambda task functions per invoice.

Amazon Bedrock Data Automation performs OCR, then Amazon Bedrock extracts structured fields from the OCR'd text, with Claude Haiku 4.5 handling about 95 percent of inference calls and Claude Sonnet 4.6 and Claude Opus 4.7 covering complex documents and manual overrides. A Lambda function resolves vendors and cost codes by fuzzy matching against reference data in Amazon DynamoDB, which also holds extracted fields, confidence scores and workflow state.

Failed and low-confidence extractions route to an Amazon SQS dead-letter queue and surface for human review with a flagged reason rather than failing silently. Step Functions pauses at the approval gate using the waitForTaskToken callback pattern while Amazon SES notifies the approvers defined in each workflow, who respond through the application via AWS AppSync.

On approval, a Lambda function retrieves rotated ERP credentials from AWS Secrets Manager and posts schema-validated invoice data to the customer's ERP over TLS. Amazon Cognito federates the customer's existing corporate SSO and enforces MFA, and Amazon CloudWatch, AWS IAM and AWS KMS provide monitoring, least-privilege access and customer-managed encryption across the workload.

The solution runs in production in the customer's AWS environment, processing around 10,000 invoices a month.

Invoice processing time fell by roughly 70 percent against a baseline of several days per batch and a target of 50 percent, measured as elapsed time from ingestion to approval. Manual effort fell by roughly 60 percent, from three full-time equivalents to about one, against the same 50 percent target, measured as staff hours per invoice batch. Data accuracy improved to above 95 percent and processing cost per invoice fell by roughly 40 percent.

Swarm and Dorje AI have continued to refine the solution against production usage: prompt changes corrected invoice date extraction on vendor templates where a purchase order date appeared ahead of the issue date, and a confidence threshold check now routes below-threshold extractions to human review with a flagged reason instead of allowing them to fail silently.