Inscribe and Amazon Bedrock: stop document fraud in 90 seconds

Inscribe now detects tampered, fabricated, and AI-generated financial documents in under 90 seconds using an agentic AI pipeline built on Amazon Bedrock. That timing is a 20x improvement over traditional manual review, which takes 30 minutes per application, and addresses a surge in deception: fraud now appears in 1 of every 16 documents, and AI-generated forgeries grew 5x from April to December 2025.

How does Inscribe’s agentic pipeline process documents so fast?

Inscribe ingests documents into Amazon S3, queues them in Amazon SQS, and processes them with Celery workers running on Amazon EC2, producing a complete fraud assessment in seconds rather than hours. The web application is served through Amazon CloudFront and an Application Load Balancer, uploads land in Amazon S3, a processing job is created and queued in Amazon SQS, Celery workers pull jobs from the queue, and the multi-model pipeline analyzes the documents before storing results in Amazon RDS.

After upload and queuing, baseline OCR and text extraction is provided by Amazon Textract, though Inscribe increasingly shifts parsing workloads directly to foundation models on Amazon Bedrock for more accurate extraction from complex financial documents. The pipeline then routes work to specialized models on Bedrock and to proprietary models on Amazon SageMaker AI. Intermediate reasoning traces and final fraud reports are stored in Amazon RDS for an audit trail while Amazon ElastiCache for Valkey and Amazon MemoryDB support short-lived caching and a vector database layer for transaction embeddings. Amazon CloudWatch monitors inference latency, error rates, token usage, and cost per request for observability.

Which models does Inscribe use on Amazon Bedrock, and for what tasks?

Inscribe matches specific foundation models to discrete tasks: Anthropic Claude Haiku 4.5 for high-volume parsing and pre-screening, Meta Llama 3.1 70B and Meta Llama 4 for transaction enrichment and entity extraction, and Anthropic Claude Sonnet 4/4.5 as the coordination and cross-document analysis layer. Claude Haiku 4.5 delivered roughly 40% lower inference costs compared with using Claude Sonnet for routine tasks while maintaining speed and improved accuracy. Meta Llama models handled transaction analysis with comparable quality at lower cost. As Ivo, Engineering Manager at Inscribe, put it, "We didn’t see much of a difference in terms of performance. The quality of Llama for those tasks was on par with what we wanted, so choosing Llama allowed us to reduce costs without sacrificing quality. Price definitely had a role to play in the final decision."

Claude Sonnet serves as the coordination layer, using an extended context window to reason across full document sets, integrate web searches for employer and address verification, and generate audit-ready fraud reports. Proprietary models hosted on Amazon SageMaker AI run in parallel to catch signals general-purpose foundation models can miss, for example pixel-level forensics to detect digitally altered images, network pattern detection against a library of known fraud templates, and visual anomaly detection for manipulated signatures or layouts.

Why does this matter for financial services?

Faster, multi-model analysis addresses three compounding challenges Inscribe identifies: scale, adaptability, and consistency. Manual review cannot scale cost-effectively as volumes rise, static rule-based systems miss sophisticated deepfakes and coordinated identity rings, and different analysts produce inconsistent outcomes. By reducing review time from 30 minutes to under 90 seconds and coordinating specialized models for discrete tasks, Inscribe aims to lower abandonment rates caused by slow approvals while preserving the explainability and audit trails regulators require.

The company points to concrete risk metrics: a single missed fraud case can mean millions in direct losses and regulatory exposure, and the rise in AI-generated forgeries makes adaptive, model-led detection essential. Bedrock’s model selection, serverless scaling, encryption, IAM controls, and model versioning and governance enable Inscribe to treat model choice as configuration and to test new releases in staging before production promotion.

What to watch

Watch whether Inscribe maintains detection accuracy as it promotes new Bedrock model releases from staging to production and how Amazon CloudWatch metrics behave: look for changes in inference latency, error rates, token usage, or cost per request that would signal model drift or the need to retrain. Also watch whether the roughly 40% inference cost advantage of Anthropic Claude Haiku 4.5 for routine tasks holds as models and workloads evolve.