Challenge
A leading Australian financial services research firm required its analysts to validate managed fund research reports against authoritative source material before publication — a critical compliance and quality control function underpinning the integrity of investment research ratings that inform adviser and investor decisions.
Each assessment required reconciling draft reports against 20–30+ source documents, including Product Disclosure Statements (PDS), FSC questionnaires, due diligence materials, fund manager responses, internal assessment records, fee data, and portfolio disclosures. These documents collectively form the evidentiary basis for published research, and any factual misalignment between a draft report and its source material carries compliance and reputational risk.
In practice, this manual validation process took approximately 8 hours per report. Around 60–70% of that time — roughly 5–6 hours per assessment — was consumed by evidence lookup: locating, reconciling, and tracking documents across disparate sources. Only 30–40% of an analyst’s effort was available for applied research judgement and decisionmaking.
As report volumes increased, the process created structural bottlenecks. It could not scale without proportional headcount increases. Outcomes varied depending on who performed the review and how evidence was tracked, introducing consistency risk. There was no structured record of which source evidence supported or contradicted each claim, creating audit trail gaps. And citation or consistency errors in published research introduced compliance and reputational exposure.
The firm needed a governed, auditable, financial-services-ready workflow that could extract claims automatically, retrieve source evidence, generate transparent verdicts, and let analysts review exceptions — without losing control of final research judgement.
Stack Highlights
The Approach
We designed and implemented an AWS-native, serverless Intelligent Document Review platform — a multiagent AI system that orchestrates foundation models while keeping outputs explainable, auditable, and governed for financial services compliance requirements.
1. Partnered with stakeholders to understand the existing review workflow, clarify accuracy and compliance expectations, and map the end to end document validation lifecycle.
This included identifying the highest value automation opportunities across the 20–30+ document types so effort was focused where it would have the biggest impact on turnaround time and risk reduction.
The engagement established clear KPIs: validation turnaround time, direct cost per report, manual evidence lookup effort, and analyst throughput capacity.
2. Designed a six stage deterministic pipeline orchestrated by AWS Step Functions: ingestion of assessment metadata from the firm’s SQL Server database via ODBC connectivity; input validation confirming assessment integrity and document availability; fact extraction from structured database records and candidate documents; truth document indexing where source documents are vectorised and stored in S3Vectors for semantic retrieval; claim evaluation using a multiagent AI pattern on AWS Bedrock; and analyst review where verdicts are surfaced for human oversight and final judgement.
Seven AWS Lambda functions execute the processing steps, all sharing a single ECR container image with a dispatcher pattern — allowing native dependencies (ODBC drivers, Python packages) to be packaged consistently while each function scales independently.
3. Built a multiagent AI layer on AWS Bedrock (Claude Sonnet 4 and Claude Haiku 4.5) with four specialised agents working in sequence.
The ExtractorAgent converts facts and candidate content into atomic, verifiable claims.
The RetrieverAgent searches S3Vectors for semantically relevant evidence from the truth corpus.
The TrimmerAgent removes lowvalue evidence before judgement — a deliberate cost control mechanism that reduces token consumption at the expensive evaluation stage.
The JudgeAgent determines whether each claim is supported, contradicted, or has insufficient evidence, citing specific source quotes for every verdict. Evidence traceability was built in from the start: every conclusion is backed by inspectable citations, persisted to S3, and can be explained and audited through CloudWatch and Step Functions execution history.
4. Delivered an analyst facing verdict viewer purpose built for the research workflow.
Analysts can filter contradictions by confidence score, review supporting evidence with source quotes, view autohighlighted PDF citations showing exactly where in a source document evidence was found, and track token usage and cost per run.
Financial services governance was embedded throughout: least privilege IAM access, VPC Endpoints for private service connectivity to S3, Secrets Manager, and Bedrock Runtime, secure credential handling via AWS Secrets Manager, and infrastructure as code deployment via AWS CloudFormation for repeatable, auditable environments.
This keeps the analyst in control of final research judgement while eliminating the administrative burden of manual evidence lookup.
The Outcome
Codex delivered a deterministic, AI-driven document intelligence capability that scales across hundreds of assessments while preserving the firm’s compliance requirements, established research tone, and analyst oversight model.
Key results included:
Validation time reduced from approximately 8 hours to under 1 hour per report — an 87–88% reduction — saving around 7 hours per assessment and enabling same day validation turnaround.
Direct validation cost reduced from approximately $480 per report (8 hours at ~$60/hour analyst time) to approximately $4 in inference costs per automated assessment run — a saving of roughly $476 per report, or approximately $47,600 per 100 reports.
Around 5–6 analyst hours are freed per report for higher value review activity, with AI agents automatically retrieving and cross referencing evidence from 20–30+ source documents — eliminating the manual evidence lookup that previously consumed 60–70% of the validation cycle.
Approximately 40% more fund coverage capacity enabled without additional headcount, as analysts can now supervise materially more assessments through automated retrieval, batching, verdict generation, and exception based review.
Full auditability achieved with complete traceability via S3 artifact persistence, CloudWatch logging, and Step Functions execution history — providing a structured, defensible record of every claim, verdict, and source citation.
This resulted in a scalable, serverless foundation processing approximately 80 assessments per month (~150 claims per assessment, ~1,280 Lambda invocations monthly) with no always on compute infrastructure — ready to extend to new document families as requirements evolve.
Overall, the solution streamlined the validation workflow, improved consistency and auditability, and allowed analysts to focus on higher-value oversight rather than manual checking.
Document Review Platform Highlights
Every verdict is backed by inspectable citations across 20–30+ reference sources including PDS, FSC materials, diligence questionnaires, manager responses, and internal assessments. Facts, claims, verdicts, and source quotes are persisted to S3 with full Step Functions execution history, strengthening defensibility and reducing citation risk in regulated financial services publishing.
A standardised sixstage flow for claim capture, evidence verification, relevance cleaning, and assessment produces repeatable, governed decisions. The multiagent pattern ensures every claim follows the same extraction, retrieval, trimming, and judgement sequence — eliminating reviewer/reviewer variability and reducing rework.
The TrimmerAgent deliberately reduces irrelevant evidence before expensive JudgeAgent prompts, controlling token consumption at the most costintensive stage. Combined with S3Vectors (avoiding dedicated vector database infrastructure), batched parallel processing across claim groups, and fully serverless compute with no always-on clusters, the platform maintains approximately $4 per assessment at scale.
The verdict viewer brings contradictions, confidence scores, and supporting evidence into one place — with auto PDF highlighting and per run cost tracking — so analysts spend time on judgement and oversight rather than manual crosschecking and administration. The interface preserves familiar ways of working while surfacing only the exceptions that require human attention.
Built as an AWSnative capability with infrastructure as code deployment via CloudFormation, least privilege IAM access, VPC Endpoints for private connectivity, and Secrets Manager for secure credential handling, the platform extends to new document families and assessment types while maintaining the governance, auditability, and explainability requirements of financial services.
Talk to Us
We would love the opportunity to connect and understand more about the problems you are trying to solve.
Adrian Campbell
Chief AI Officer
Martin Campbell
Managing Partner
Get in touch to coordinate a meeting with one of our technical experts.
Australia: +61 7 3132 3002.
