Technology

How Adient is built

A multi-agent system on AWS in the EU. Each step of a KYC case is a specialised agent; a deterministic orchestrator runs them in order and an analyst makes the decision.
Analyst app and API send a case to the orchestrator, which runs six agents in order; agents use Amazon Bedrock and external data through MCP servers; case data is stored in DynamoDB and S3 in eu-central-1.Analyst appNext.js + APIOrchestratorEventBridge + SQS1Document processing2Company structure3Registry check4Sanctions screening5Financial AML6KYC reportEach agent is an AWS Lambda functionAmazon BedrockAnthropic Claude modelsMCP serversRegistries · sanctions lists · newsCase dataDynamoDB + S3, eu-central-1
A case flows from the analyst app through the orchestrator to six agents. Agents call language models on Amazon Bedrock and reach external data through MCP servers. Case data is stored in AWS Frankfurt.

The six agents

1. Document processing
Reads every document with OCR, classifies it, extracts the people, companies and relationships in it, and flags documents that are missing.
2. Company structure
Builds the ownership graph, calculates effective ownership through each layer and identifies the ultimate beneficial owners.
3. Registry check
Verifies the companies against public registries such as KVK, Companies House, Handelsregister and Zefix, and lists any differences.
4. Sanctions screening
Screens every person and company in the structure against OFAC, EU, UN and other sanctions lists.
5. Financial AML
Analyses the financial statements and bank statements for anti-money-laundering red flags.
6. KYC report
Writes the risk report with an overall score and a breakdown: a draft after the analysis, the final report after the analyst’s review.

Orchestration, not a single prompt

The orchestrator is deterministic: it runs the agents in a fixed sequence over Amazon EventBridge and SQS, each agent as an AWS Lambda function. When the analysis is done it asks the report agent for a draft and stops for the analyst’s review. The final report is generated only when the analyst asks for it.

Language models

Agents use Anthropic Claude models through Amazon Bedrock for reading documents, extracting entities, analysing financial statements and writing the report. Calls go through Bedrock cross-region inference, which can process a request in another AWS region; case data itself is stored in eu-central-1 (Frankfurt).

External data through MCP

Agents reach external data through Model Context Protocol (MCP) servers: corporate registries (such as KVK and Companies House), sanctions lists (OFAC, EU, UN and others) and news sources for adverse media.

Fits into a bank’s own systems

Cases can be handed to Adient and results returned through the Agent2Agent (A2A) protocol, so an existing onboarding workflow can call it like any other service.

Audit trail

Every agent action, every analyst note and document decision, and every report version is stored on the case, so a reviewer can see what was found, by whom and when.

Stack

Next.js analyst app, FastAPI service, AWS Lambda, EventBridge, SQS, DynamoDB, S3, KMS and Cognito, Amazon Bedrock.

See the agents work on a real case

We walk you through a case from start to finish: documents in, ownership mapped, registries and sanctions checked, draft report, analyst review and final report.

Book a demoSee the product