Fraud is arriving faster and more convincingly than many legacy systems can handle. Inscribe now uses a multi-model, agentic architecture on Amazon Bedrock to analyze entire application document sets and deliver a final fraud assessment in under 90 seconds, a dramatic speedup over manual review.
We looked at how that system is put together, which models handle which tasks, and what measurable results customers have seen since deployment. If you work in risk, underwriting, or platform engineering, this is a useful case study in building fast, auditable AI pipelines for high-stakes decisions.
Why banks and lenders needed a different approach
Manual document review still dominates many underwriting workflows. The old flow is tedious: a human analyst inspects bank statements, pay stubs, tax documents, and ID images, then cross-checks employers, signatures, and addresses. Inscribe’s 2026 State of Document Fraud Report says manual review can take roughly 30 minutes per application, and that AI-generated forgeries increased about fivefold between April and December 2025. That kind of volume breaks traditional processes.
Three problems stack up. First, scale: more applications mean proportionally more analyst hours and rising costs. Second, adaptability: static rule sets miss novel tactics like synthetic identities and AI deepfakes. Third, consistency: different reviewers reach different conclusions, which creates compliance and fairness headaches.
Agentic AI: an expert analyst in software form
Inscribe describes their system as agentic. For us, that means the pipeline takes a high-level goal, breaks it into tasks, picks the right tool for each task, and synthesizes the outputs into an audit-ready report. The whole process runs without a manual handoff in most cases. This is the same basic agent idea larger vendors are positioning for enterprise workflows, see SAP’s take on AI agents for one example of that framing.
The practical benefit is that the system does more than flag a suspicious field. It parses documents, runs parallel forensic checks, cross-references data across every document in the application, and performs external verification when needed. The result is a single, explainable fraud assessment delivered quickly.
How models are matched to tasks
One useful design choice we like: use different foundation models for different jobs rather than trying to make one model do everything. Inscribe evaluated models on performance, latency, and cost and then assigned tasks accordingly. That makes model selection a configuration choice instead of an engineering rewrite when a new model appears.
- Fast parsing and routine field extraction run on a lower-cost Claude model variant optimized for throughput.
- Transaction enrichment and entity extraction use Meta Llama family models where they matched the accuracy requirements at lower cost.
- Cross-document reasoning, extended-context analysis, web search coordination, and final report generation run on a higher-capability Claude variant that supports long context windows.
This layered approach keeps routine work cheap and fast while reserving more expensive resources for complex, multi-step reasoning.
Architecture highlights and the AWS pieces that glue it together
Documents land in object storage immediately on upload. A queue-based job system then hands work to distributed workers that scale with demand. Optical character recognition and baseline extraction have traditionally used Amazon Textract, but Inscribe increasingly pushes parsing workloads directly to foundation models for better results on complex financial layouts. For foundation models and large-scale orchestration, Inscribe uses Amazon Bedrock, which lets them pick models from multiple providers through one API and scale serverless inference to match spikes in traffic.
Proprietary detectors run in parallel. Pixel-level forensic checks, pattern matching against known fraud templates, and visual anomaly detectors are hosted on a managed ML platform so they can scale and be monitored independently of the foundation model pipeline. Results and intermediate reasoning traces are stored in a relational store for auditability, with a fast in-memory cache and a vector store for similarity queries powering lookups across past cases. Observability tracks latency, error rates, and model usage so teams can spot model drift and regressions. For a complementary take on real-time model-powered fraud detection pipelines, see Databricks’ real-time fraud approach.
If you want to explore the specific AWS services mentioned here, start with the Amazon Bedrock product page and the Amazon Textract and Amazon SageMaker pages for the managed model hosting and deployment pieces.
What the numbers show
Speed is not theoretical. Inscribe reports that its agentic system reduces review time dramatically, delivering assessments in under 90 seconds. That is roughly a 20x speed improvement compared with the cited manual process. Customers gave concrete outcomes:
- A mid-sized financial services firm reduced manual review time by more than 90 percent and said that prevented millions in potential losses while standardizing workflows.
- A credit union reported more than $3 million in potential loan fraud savings within eight months of deployment, while cutting individual case review time by up to 99 percent in some workflows.
- Another customer credited the system with detecting coordinated fraud rings and preventing roughly $5.6 million in losses.
Those are operational examples from organizations using the platform, not guarantees for every deployment. They do show the kinds of outcomes a multi-model, instrumented pipeline can produce when tuned for a particular customer profile.
Practical takeaways if we were building this ourselves
- Pick the right tool for each job. Measure latency, accuracy, and cost per task and map models accordingly.
- Keep auditability front and center. Store intermediate reasoning traces alongside final reports so compliance teams can reconstruct decisions.
- Run parallel, specialized detectors for signals that general-purpose models miss. Pixel-level and template matching models add complementary coverage.
- Design for serverless scaling so you can handle peak application windows without overprovisioning overnight capacity.
- Monitor model drift with observable metrics and have a robust staging workflow for promoting model versions into production.
Where this leaves us
We have seen how an agentic approach, combined with careful model selection and cloud-native scaling, can push fraud detection from minutes to seconds while keeping results auditable and defensible. That matters because fraud tactics migrate quickly and because financial institutions must balance speed with regulatory obligations.
If you work in risk or engineering, this is a practical template: coordinate specialized models, instrument everything, and treat model selection as configuration. The results reported by Inscribe’s customers show what is possible when those pieces come together, and the architecture gives us a clear path for iterating as foundation models and fraud techniques change.
What have we missed that matters for your stack? Tell us what your detector pipeline looks like and where it trips up so we can dig into the next problems together.