Multi-Agent M&A Due Diligence: 30% Faster Document Review with 22% More Compliance Flags

What Is New in Multi-Agent M&A Due Diligence? As of May 23, 2026, a consortium of 12 investment banks and law firms has publicly shared results from the first known multi-agent M&A due diligence pilot running entirely on AWS Bedrock. The project combined two specialized foundation models—Qwen 3.8 Max for document extraction and Llama 5 for risk scoring—within a three-agent orchestration framework. The outcomes are the first vendor-neutral, peer-validated benchmarks showing how agentic document review for M&A can accelerate deal workflows without sacrificing accuracy. Why Multi-Agent Architecture Fits M&A Due Diligence Modern M&A due diligence involves hundreds of contracts, financial statements, regulatory filings, and internal communications—often in multiple languages and formats. A single large language model (LLM) may choke on the sheer volume or miss context-switching between extrac

tion and evaluation. Multi-agent architectures break the work into specialized subtasks: Document extraction agent : reads, chunks, and normalizes content Risk scoring agent : applies domain-specific compliance rules and flags anomalies Orchestration agent : manages handoffs, merges outputs, and ensures audit trails By decoupling these functions, the consortium achieved both speed and precision—something a monolithic prompt pipeline could not replicate at scale. The Consortium Pilot: 12 Banks, Two Foundation Models, One AWS Bedrock Backbone In Q1 2026, a group of 12 major investment banks (including two of the top five by M&A volume) and three law firms partnered with an independent AI research lab to run a controlled due diligence pilot. The scope covered 2,500+ documents typical of a mid-market acquisition: purchase agreements, NDAs, employment contracts, intellectual property filings,

and regulatory correspondence. The infrastructure was entirely on AWS Bedrock, chosen for its enterprise-grade security, model flexibility, and built-in data governance. Model selection: Qwen 3.8 Max (Alibaba Cloud)—tasked with document extraction. Its ability to process long contexts (up to 128K tokens with sliding window) made it ideal for chunking and structuring dense legal language into machine-readable entities (parties, dates, obligations, termination clauses). Llama 5 (Meta)—tasked with risk scoring. Its instruction-following reliability and smaller inference footprint allowed the consortium to run compliance checks without over-provisioning compute. Both models were accessed via AWS Bedrock’s marketplace, with data never leaving the VPC. Three-Agent Architecture: Document Extraction, Risk Scoring, and Orchestration The pilot deployed a three-agent topology: Agent 1: Qwen Docume

nt Extraction Agent Ingests raw PDFs, Word docs, and scanned images (via AWS Textract) Extracts structured fields: parties involved, effective dates, renewal terms, restrictive covenants Outputs JSON schemas per document into Amazon S3 Agent 2: Llama 5 Risk Scoring Agent Takes structured document schemas plus user-defined compliance rules (from the banks’ internal policies) Scores each clause on a 0–100 risk scale, highlighting gaps in standard anti-money laundering (AML), data privacy (GDPR, CCPA), and anti-trust clauses Outputs scored tables and free-text flags with rationale Agent 3: Orchestration Agent (custom lightweight model via Bedrock) Manages task queues between extraction and scoring Checks for hallucinations: if Llama flags a risk that contradicts extraction data, it triggers a re-read cycle Generates the final consolidated report with an audit trail of every decision This se

paration of concerns allowed each model to stay within its optimized domain—no extraction model needed to understand risk scoring, and vice versa. Data Pipeline Design for Secure Deal Documents Security was a non-negotiable constraint. The consortium enforced the following pipeline: 1. Ingestion : Documents uploaded into an encrypted S3 bucket with server-side KMS keys scoped per deal. 2. Preprocessing : AWS Lambda functions convert to plain text, detect language, and redact PII using Bedrock’s Guardrails for AI. 3. Chunking : Documents are split into 4K-token chunks (with 10% overlap) to fit Llama’s context window while preserving clause boundaries. 4. Agent calls : Each chunk triggers the Qwen and Llama agents asynchronously via Bedrock Converse API. 5. Result merging : Orchestration agent re-joins chunks, deduplicates flagged risks, and writes the final report back to S3 with versioni

ng. All data stayed within the AWS region (us-east-1), with network-level isolation via private subnets and VPC endpoints. Measured Outcomes: 30% Faster First-Pass Review and 22% Better Compliance Risk Detection The pilot compared multi-agent results against both fully manual review and a single-mod