From Days to Minutes: How we built Multi-Agent KYC/KYB on AWS

The Know Your Customer/Business (KYC/KYB) space has been a target of innovation for over 15 years. The challenge is binary. Too much friction loses customers to competitors. Too little validation adds risk to the institution. The balance is rigor internally, ease externally.

For personal and small-business accounts, the industry has largely solved the problem. Straight-through processing via rules-engines handles the “happy path” at scale. But anything outside that path falls into a manual, human-driven exception space. For complex individuals (ultra-high net worth), complex businesses (large corporates, multi-entity ownership structures, multi-business lines), the checks become multi-dimensional. Nature of business, source of wealth, adverse media, corporate structure. No two cases look alike. The work is reading, understanding, cross-validating across disparate sources, and applying genuine reasoning to one question: “does it make sense?”

This exception space is expensive. A single enhanced due diligence case can consume days of senior analyst time. Multiply by thousands of cases per year, factor in the scarcity of qualified compliance professionals, and institutions face a structural problem. The cost of doing it properly makes entire customer segments commercially unviable. Some institutions cap how deeply they investigate. Others decline customers they cannot cost-justify reviewing. Much has been written about applying AI to KYC/KYB. Most of it stops at the conceptual. This post documents a working system. Five KYC/KYB checks, each with its own agent architecture, prompt methodology, and failure modes. We built it, ran it, broke it, fixed it, and are publishing the full method so you can build your own.

The Capabilities That Make This Possible

Agentic AI orchestrates multi-step processes through iterative reasoning. It plans, interacts, adapts, and continuously advances toward a defined objective.

For example, a human analyst doing adverse media screening. They run a web search, review the first page, read a couple of articles, write a report. Good enough. Time-boxed by necessity.

An AI agent does the same task deeper, wider, and completes in minutes. It searches across multiple risk topics in parallel. For every matching result, it reads the full article. It cross-checks findings against institutional policy. It produces a comprehensive, evidenced report with citations. The human reviewer then spends 5 minutes reviewing a thorough, pre-reasoned recommendation rather than the 45–60 minutes a thorough manual review would demand.

Four capabilities make Agentic AI well suited for the KYC/KYB exception space:

Cross-source data comprehension. KYC/KYB is cross-validation. A customer tells us about themselves. We check that against documents, registries, third-party data, and research. Large Language Models understand language, and most data sources are language. A JSON payload from Companies House. A free-text business description from an applicant. A table in a PDF bank statement. Frontier Foundation Models process all these natively. They do not need separate parsers for each format. They read.

Contextual document intelligence. Traditional Optical Character Recognition (OCR) identified letters. It never understood words. If you wanted an address, you taught the system exactly where on the page to look. Different format, different rules; failure. Large Language Models (LLMs) analyze the full document and extract meaning through context. Format is irrelevant. Handwritten, scanned, digital. Even when individual characters are illegible, context fills the gap. “Hello, how a[] you?” The answer is obvious to a human. It is equally obvious to a Frontier Foundation Model.

Adaptive research. KYC/KYB requires publicly available research. Company websites, news searches, registry lookups. This research must be adaptive. Ignore noise, dive deeper into relevant findings, follow threads. An AI agent operates with an objective, a set of tools, and boundaries. It adapts its path based on what it discovers. The output is contextualized findings with reasoning attached. A structured assessment, not a raw list of hundreds of results for a human to triage.

Multi-dimensional reasoning. In complex cases, the real work is weighing an entire profile simultaneously. A sanctions alert matches on a name. “John Smith” matching a sanctioned “John Smith” is statistically unremarkable. “Kael Dorian Vronsky” matching another is extraordinary. Humans weigh name commonality, geography, date of birth, business sector in parallel and apply judgment. AI agents can do the same. We teach them a framework, a persona, a mindset. They apply it consistently while remaining dynamic in approach. The reasoning trail is produced natively. Regulators get the kind of transparency they expect.

KYC/KYB is multiple problems running in parallel, each requiring different skills, different data sources, different reasoning. A single agent doing all five will do all five poorly. A multi-agent system gives each check its own agent, its own reasoning trail, and its own success criterion. The orchestration layer is deterministic code. AI never unilaterally approves a case.

What We Built

A reference proof of concept that automates KYC/KYB due diligence with AI agents. Submit one applicant, and five independent modules investigate it in parallel. Each produces a structured, evidenced verdict. A unified workbench streams every module’s progress live to a compliance analyst who retains the final decision.

This is a reference implementation. We built it to prove the patterns, learn from the failure modes, and teach the method. The agents, prompts, pipelines, and architecture are real. The data used for this blog is synthetic (mock UK Companies House responses). You take the method, plug in your data sources, plug in your policies, and build your own.

Each module owns its own storage, API, and agent runtime.

Tools and data connections are pluggable. The architecture is jurisdiction-agnostic. A UK Companies House client swaps for any corporate registry globally. Only the evidence-gathering layer adapts. Analysis and judgment remain unchanged.

Document processing happens once, up front. A specialist agent processes all submitted documents and produces structured JSON outputs available to every module. Downstream modules receive clean, pre-parsed evidence and focus purely on analysis.

Institutional policies live in Knowledge Bases. When an agent needs to apply policy, it retrieves the relevant passage dynamically and cites it in its reasoning. Because policies are read from the knowledge base at runtime, updates take effect immediately without redeployment.

Solution Architecture

The solution runs on AWS. The service choices are deliberate for a regulated, long-running agentic workload (Figure 1).

Amazon Bedrock AgentCore runs every module. Each of the five KYC/KYB checks deploys on its own AgentCore Runtime deployment, built for sustained multi-step agent workflows. A case runs for minutes across dozens of LLM calls, parallel fan-outs, and tool invocations. It handles session isolation, tracing, timeout management, and scaling to zero natively. Amazon Bedrock is the LLM access layer. Every model call routes through Bedrock, giving you a single API across foundation models from multiple providers with guardrails and invocation logging.

Agents are written with the Strands Agents SDK, an open-source framework that provides multi-agent graph structure, structured output parsing, tool invocation, and a single model-call chokepoint that makes swapping models or adding guardrails a one-line change.

AWS Step Functions coordinates the case lifecycle: accept submission, parse documents, activate modules, hold a task token until all modules complete. Amazon EventBridge is the platform bus for fan-out, progress streaming, and completion signaling. Modules never call each other. AWS AppSync Events pushes real-time progress to the browser without polling. Amazon DynamoDB and Amazon S3 store case state, findings, and events per module, retained independently of compute – becoming your audit trail.

Amazon Bedrock Knowledge Bases, a managed retrieval-augmented generation (RAG) service that connects foundation models to your data sources, stores institutional KYC/KYB policy, retrievable and citable by agents at runtime. Policies update without redeploying code.

Data sources connect through AWS Lambda-based tools fronted by the Model Context Protocol (MCP) Gateway. These are pluggable. Swap the source, agents keep working.

Figure 1: Platform architecture – five independent KYC/KYB modules orchestrated by AWS Step Functions on Amazon Bedrock AgentCore

The Five KYC/KYB Checks and How Each Works

Each KYC/KYB check forms a module, a multi-agent system in itself:

Adverse Media Screening

Searches the open web across configurable risk topics (fraud, money laundering, terrorism financing, bribery, sanctions evasion). For every relevant finding, it reads the full article, applies policy, and determines whether the content constitutes true adverse media.

Figure 2: AI Agents Powering Adverse Media Scanning

• Smart Web Searcher Agent builds and executes targeted queries for each risk topic in parallel. It uses topic-specific anchor words, fetches full page content from promising URLs, and filters for relevance before anything moves downstream. Running searches in parallel collapses a 30-minute sweep into roughly 3 minutes.
• Research Agent reads the merged corpus from all searches and synthesizes it into findings grouped by underlying issue. Five outlets reporting the same conviction becomes one finding. It assigns severity, tags across multiple risk topics, and discards noise with documented reasoning. Typical output: 3-8 distinct findings from 50+ raw URLs.
• AMS Analyst Agent applies the institution’s own KYC/KYB eligibility policy to each finding. It retrieves specific policy sections from a Bedrock Knowledge Base and produces an auditable verdict: CLEAR, ESCALATE, or INCOMPLETE. Every judgment is grounded in institutional rules, with cited references and recommended actions.

Sanctions Screening

Sanctions Screening mirrors the human four-eyes (dual independent review) compliance process. One AI analyst evaluates each pre-screened match. A completely independent AI reviewer audits the reasoning. If the reviewer rejects, the finding escalates to a human. Both agents use zero tools. All evidence is embedded directly in the prompt.

Figure 3: AI Agents Powering Sanctions Screening

• Match Investigator Agent takes the compiled evidence and determines whether the case is a true match or a false positive. It draws on policy and operating guidance from a Knowledge Base to produce a clear, reasoned justification. The output must be well-reasoned, policy-aligned, and fully explainable to support auditability and continuous improvement.
• Data Enricher Agent can be used as a tool to further research the individuals and gather more information. For example, does this company have subsidiaries or parent owners? Can we learn more about their dealings in a particular country?
• Outcome Validator Agent independently reviews the Investigator’s decision and reasoning. It confirms the required standard of accuracy and evidentiary certainty is met. No hallucinated or unsupported data. Logical consistency in reasoning. Sufficient verified evidence such that a human reviewer would not question the conclusion. If any of these checks fail, the case is escalated to human review.

Business Details Validation

Business Details Validation cross-references what the customer declared against the public register and their own website. In one test case, the applicant declared they sell alcohol. Their Standard Industrial Classification (SIC) code (unlicensed restaurant) explicitly means they do not have permission to sell alcohol. Their website showed alcohol on the menu. The agent caught the contradiction across three sources and escalated.

Figure 4: AI Agents Powering Business Details Validation

• Core Details Analyst Agent compares the customer’s declared business details (legal name, trading name, address, directors, incorporation date) against Companies House and corroborating sources like company websites. It identifies matches, mismatches, and undisclosed information across multiple data points. This is the factual verification layer. “X Restaurant Café” vs “X RESTAURANT LTD”: is that formatting or fraud? It requires multi-dimensional reasoning, naming convention awareness, address normalization, and fuzzy director matching that deterministic code handles poorly.
• Nature of Business Analyst Agent assesses whether the customer’s declared business activity is coherent across all available evidence: SIC codes, company website, stated profile, and licensing signals. It flags contradictions or undisclosed regulated activities. This is the coherence layer. A company registered as “restaurant” whose website advertises alcohol delivery and late-night entertainment needs specific licenses. Requires cross-referencing multiple sources and applying regulatory knowledge.

Company Structure Resolution

Company Structure Resolution recursively walks corporate ownership through registry APIs, identifies all directors and Ultimate Beneficial Owners via holding companies, calculates whether each person is in-scope (greater than 25% ownership), and validates that required documentation exists. The walk is deterministic. The judgment on documentation sufficiency is where the agent reasons.

Figure 5: AI Agents Powering Company Structure Resolution

• CSR Researcher Agent [used if needed] browses trusted web sources to find ownership structures through government portals, web pages, or documents. Some companies have overseas owners in countries with no API access. This agent performs agile, adaptable agentic web-browsing to handle the human investigation steps programmatically.
• Structure Analyst Agent takes the Researcher’s output and, for each in-scope Ultimate Beneficial Owner (UBO), checks whether the applicant has provided appropriate declarations, Identity & Verification (ID&V) documents (e.g. passport photo), and supporting paperwork as required by KYC/KYB policy. It flags any gaps in structure or missing documents against policy requirements.

Source of Wealth

Source of Wealth reasons across four dimensions: business profile plausibility, bank statement evidence, cash-intensity risk signals, and source-type plausibility. The synthesis rules combine dimensions into a single verdict. In test cases, the agent identified mismatches that went beyond missing documents. It was the sense-check on whether the overall picture was coherent.

Figure 6: AI Agents Powering Company Structure Resolution

• Source of Wealth Analyst determines whether the customer’s declared source of wealth is plausible given their business profile, maturity, sector, and supporting financial documents. It identifies claims that don’t add up. This is the plausibility layer. £1M annual revenue from a 1-year-old barbershop doesn’t pass the smell test. Requires sector-appropriate revenue benchmarks and the ability to assess whether document evidence supports the claim.

Deep Dive: Adverse Media Screening

Adverse media screening is the most research-intensive of the five checks. Unlike sanctions screening, which compares against a fixed database of designated persons, adverse media screening requires actively searching the open web, evaluating what is found, and applying institutional eligibility policy to reach a defensible verdict. We automated that end to end with three agents.

A Smart Search Agent runs per risk topic in parallel. Each one constructs targeted queries using topic-specific anchor words, fetches full page content from promising URLs, and filters for relevance before anything moves downstream. Results from all topics are merged into a URL-deduplicated corpus deterministically. A Research Agent reads that corpus and synthesizes it into structured findings grouped by underlying issue. Five outlets reporting the same conviction becomes one finding, scored and tagged. Finally, an AMS Analyst Agent runs per finding. It retrieves the relevant policy section from a Bedrock Knowledge Base and applies it to produce an auditable verdict: CLEAR, ESCALATE, or INCOMPLETE. The case-level outcome rolls up through a fixed decision table. No AI is involved in the final roll-up.

Figure 7: Adverse Media Screening – sample execution output. Three findings were surfaced across risk topics. One was cleared after policy assessment determined it fell below institutional thresholds. The remaining two were flagged for escalation to a human reviewer.

Figure 8: Adverse Media Screening – detailed audit trail of an escalated finding, showing the evidence, the policy reference used to determine escalation, and the recommended action.

The Decomposition Method

This is the transferable methodology. Regardless of jurisdiction or data sources, the approach to breaking any compliance workflow into agents follows the same steps, underpinned by four principles.

Start with the human process. Sit with an analyst. Watch them work a case. Write down every step they take. This is your raw material.

Separate the deterministic from the judgmental. Fetching a company record. Calculating ownership percentages. Checking a name against a sanctions list. These have one correct answer. They become code. No AI, no tokens, no hallucination risk. Deciding if an article constitutes adverse media. Assessing whether a source of wealth is plausible. These require reasoning against policy. They become agents. Principle: establish facts deterministically, reserve AI for genuine judgment.

Identify the natural fan-out points. Where does the human do the same cognitive process multiple times with a different lens? Adverse media across fraud, money laundering, bribery, sanctions evasion. Same pattern, different input. Each becomes its own agent. Principle: narrow agent mandates. One job, one success criterion, one reasoning trail.

Separate gathering from judging. A single agent that searches the web and then decides whether the content is adverse media is doing two jobs. When it fails, you cannot tell which job failed. Split them. A search agent gathers. An analyst agent judges. Principle: structured, validated output at every boundary. Free-form text never flows downstream as a decision.

This boundary also contains errors. If the search agent returns irrelevant content, the analyst agent rejects it at the validation boundary rather than reasoning over garbage. Each boundary is a circuit breaker. Failures localize to the stage that produced them. A bad search result does not cascade into a bad verdict because the judging agent independently validates every input against its acceptance criteria before reasoning over it.

Minimize the actuation space. If deterministic prework can establish the fact, the judging agent gets no tool. Sanctions analysts use zero tools because all evidence is already assembled in the input. Adverse media search agents have budgeted tool access. Principle: tools are liabilities.

Define success in one sentence. For every agent, write one sentence that describes success. If you cannot write that sentence, decompose further.

Map the data flow. Draw which agents produce outputs that other agents consume. Anything without dependencies runs in parallel. The orchestration layer implements this graph in deterministic code.

Teaching Agents How to Think

Every agent has a dedicated prompt stored as standalone markdown, read at runtime. These teach the agent a methodology with worked examples, outcome rules, and anti-patterns.

Persona and regulatory grounding. Every prompt starts with who the agent is and what the stakes are:

You are a Level 2 sanctions screening analyst at a regulated financial institution.
Sanctions violations carry strict liability. The standard is "reasonable grounds to suspect." False negatives carry criminal consequences. False positives carry operational cost but no regulatory risk. When evidence is ambiguous, the correct answer is always ESCALATE.

This establishes asymmetry of risk. Every subsequent decision is shaped by it.

Multi-dimensional reasoning frameworks. For complex judgment, we teach the agent to assess independent dimensions then synthesize:

DIMENSION 1: Business Profile Plausibility Does the claimed wealth make sense for THIS business, THIS location, THIS many years established?
DIMENSION 2: Bank Statement Evidence Do the parsed documents support the declared source of wealth?
DIMENSION 3: Cash-Intensity Risk Signals Does the business type carry inherent Anti-Money Laundering (AML) risk requiring additional scrutiny?
DIMENSION 4: Source-Type Plausibility Is the declared source type consistent with the business profile and documentary evidence?

Each dimension is assessed independently. Synthesis rules combine them. This prevents a strong signal in one dimension from overwhelming a weak signal in another.

What It Takes to Go to Production

Evaluations come first. Score agent outputs against known-good cases. Build a test suite of historical decisions with expected outcomes. Run agents against it. Measure agreement. Where agents diverge from human decisions, investigate whether the agent was wrong or the human was inconsistent. Both happen.

Silent mode proves the system under real conditions. Run agents alongside human analysts without acting on results. Compare outcomes over a sustained period. Every divergence gets investigated. This is where you discover the edge cases your test suite missed and the prompts that need tightening.

Human-in-the-loop is the target operating model. Agents recommend. Humans decide. This is permanent. Regulators require human accountability for KYC/KYB decisions. The agent removes the research burden. The human retains the authority.

Observability starts with tokens. Every model call in every module records input and output token counts. The finalize stage aggregates them per agent and computes estimated cost per case. If a module that used to consume 40k tokens starts consuming 80k, something changed. Prompts were modified, input data grew, or an agent is looping. Pair token monitoring with outcome distribution tracking (what percentage of cases escalate vs. clear over a rolling window) and you have a lightweight quality signal before you need to rerun full evaluations.

Cost profile favors agents at scale. In our reference runs, a full five-module case completes for single-digit dollars in model costs. The exact figure depends on the model, the number of risk topics configured, and how many findings surface. Compare that to the fully loaded cost of a senior compliance analyst spending days on an enhanced due diligence case. The more important question is whether reduced unit cost makes previously unviable customer segments commercially justifiable.

Data privacy requires institutional controls layered on platform defaults. Agents process names, addresses, identity documents, financial statements, ownership records. Amazon Bedrock does not use customer inputs or outputs to train models. Data stays within the AWS region you deploy to. Beyond that baseline, institutions should enforce PII handling through Amazon Bedrock Guardrails, encrypt all case data at rest, scope module access so each agent only sees what it needs, and map storage in DynamoDB and S3 to existing data retention and deletion policies.

Regulatory engagement belongs at the start, not the end. Compliance and legal teams need to see the system before it sees real cases. The questions they will ask are predictable. How are agents tested? Evaluations and silent mode. How are decisions traced? Every agent produces a reasoning chain with policy citations, stored as an immutable audit trail. What happens when the agent cannot conclude? INCOMPLETE outcomes route to human review automatically. How is quality monitored post-deployment? Token-level observability and outcome distribution tracking. Framing the conversation around explainability, auditability, and conservative failure modes (escalate to human, never silently approve) gives compliance teams the vocabulary to assess the system against their regulatory obligations.

Beyond The Reference Implementation

We have already worked with financial institutions to reimagine parts of their KYC/KYB processes using these techniques. Some have reached production in under a year. In one case, our specialists and Professional Services builders partnered with a large European bank, to reimagine and implement agentic Ongoing Due Diligence; case processing time dropped from multiple analyst-days to under two hours per case. The patterns in this post apply to a single bottleneck or the entire lifecycle. Onboarding. Ongoing due diligence. Enhanced due diligence. Periodic reviews. Event-triggered reassessments. Apply one module or all five (or more).

Given agents do not get tired, and run on demand, the end-state is perpetual KYC/KYB. Customer profiles continuously monitored. Changes in behavior or corporate structure detected as they happen. Risk assessments that stay current without waiting for a scheduled review cycle. The shift from periodic to continuous is the structural elimination of the remediation backlog that every compliance team is fighting today.

Next Steps

This post introduced the reasoning frameworks and architecture patterns behind agentic KYC/KYB. Each of the five modules carries its own design decisions, failure modes, and prompt engineering depth that go well beyond what a single blog can cover.

If you are exploring agentic approaches to compliance automation, reach out to your AWS account team or the AWS Financial Services Solutions Architecture team, we can walk you through an in-depth demo, the full architecture of every module and help you design, build and launch your own version of agentic KYC/KYB on AWS.

Disclaimer: This post is for informational purposes and does not constitute legal or compliance advice; consult your compliance and legal teams for your jurisdiction.