AI-Powered Content Curation for Financial Crime Detection

lseg · Finance · 2024 · source Agentic: yes · promotional bias: med · source-backed · extracted with prompt 0.5.4 (schema v2)

Generated from extraction.json — do not edit by hand. Every claim cites verbatim source evidence; stated means the quote directly asserts the claim, implied means it supports it indirectly.

Problem

Domain: Financial-crime risk intelligence

LSEG Risk Intelligence needed to accelerate World-Check content curation across high-volume multilingual adverse-media and sanctions-related data without reducing regulatory accuracy or analyst accountability.

Constraints:

World-Check data is used by banks, insurers, governments, and regulated customers
Financial crime moves quickly, so stale updates can create customer risk
Human domain expertise remains necessary for high-risk decisions and compliance

Architecture

LSEG's World-Check GenAI program is organized as a maturity ladder on AWS Bedrock: prompt-only article analysis, single-agent actions, RAG over vectorized content, multi-source/graph-style RAG, and agent orchestration with human quality gates. News/content enters queues, LLMs extract entities and facts, RAG cross-references World-Check data, agents draft/enrich records, and analysts validate outputs before publication.

Diagram confidence: reconstructed (reconstructed from prose).

flowchart LR
  world-check-curation-platform["world-check-curation-platform"]
  sns-content-queue["sns-content-queue"]
  bedrock-llm-processing["bedrock-llm-processing"]
  single-agent-actions["single-agent-actions"]
  rag-vector-search["rag-vector-search"]
  agent-orchestration["agent-orchestration"]
  human-quality-review["human-quality-review"]
  world-check-curation-platform --> sns-content-queue
  sns-content-queue --> bedrock-llm-processing
  bedrock-llm-processing --> single-agent-actions
  bedrock-llm-processing --> rag-vector-search
  single-agent-actions --> agent-orchestration
  rag-vector-search --> agent-orchestration
  agent-orchestration --> human-quality-review
  human-quality-review --> world-check-curation-platform

Drag to pan · scroll or use +/− to zoom · reconstructed from the source's described component topology

Components

Component	Role	Interacts with
`world-check-curation-platform`	Risk-intelligence workflow where analysts and AI curate high-risk individuals, entities, adverse media, and related records.	`sns-content-queue`, `human-quality-review`
`sns-content-queue`	Queues acquired content for downstream LLM extraction and curation processing.	`world-check-curation-platform`, `bedrock-llm-processing`
`bedrock-llm-processing`	AWS Bedrock-backed prompt and agent calls for summarization, entity extraction, classification, and draft generation.	`sns-content-queue`, `single-agent-actions`, `rag-vector-search`
`single-agent-actions`	Narrow agents for one bounded task such as article classification or event extraction.	`bedrock-llm-processing`, `agent-orchestration`
`rag-vector-search`	Titan embeddings and vector databases used to semantically cross-reference articles and World-Check records.	`bedrock-llm-processing`, `agent-orchestration`
`agent-orchestration`	Coordinates search, identification, extraction, draft proposal, quality checks, token budgets, and escalation decisions across business-process stages.	`single-agent-actions`, `rag-vector-search`, `human-quality-review`
`human-quality-review`	Analyst and QA review layer that preserves accountability for regulated risk intelligence before publication.	`agent-orchestration`, `world-check-curation-platform`

Data flow

Content is acquired and placed onto an SNS queue, then Bedrock-backed processing extracts names, relationships, facts, events, and relevance signals. As maturity grows, embeddings and vector search cross-reference articles and existing World-Check records, orchestration routes search/extraction/drafting/QA steps, and humans review risky or unresolved items before customer-facing data updates are published.

“it does the identification, it does the extraction, it potentially proposes a draft, sends it through to quality assurance to go and review that particular process”transcript, stated

“in through the platform, it's being acquired, it's putting it onto an SNS queue there, and then we can start processing that data to be able to step through it”transcript, stated

Why this architecture

LSEG deliberately stages capability maturity instead of jumping into complex RAG or straight-through automation. Human review stays in the process because financial-crime curation is regulated and accuracy-sensitive, while the AI layers are introduced as augmentation that saves analyst time and earns trust one production slice at a time.

“This is very much an augmentation solution, not an automation solution, okay? And it's key and important to understand how that ties back together.”transcript, stated

“We didn't start straight with fine-tuned models, as Chris said, "We started with simple prompts, validated the results, and we progressively added capabilities."”transcript, stated

Models used: AWS Bedrock models · Titan Embeddings

Scale

World-Check supports more than 10,000 customers.
“We have 10,000 customers.”transcript, stated
World-Check operates across 60-plus languages and is used by banks, insurance providers, and governments worldwide.
“60 plus languages, rigorous quality control, and it's used by banks, insurance providers, and governments worldwide.”transcript, stated
LSEG has more than 200 analysts curating World-Check information.
“we've got over 200 analysts who are curating data from thousands of sources.”transcript, stated
The financial ecosystem use case touches 260 million daily transactions.
“that 260 million daily transactions depend upon.”transcript, stated
The AI curation process moved updates from hours to minutes.
“On speed and quality, we are moving from hours to minutes for updates while maintaining accuracy and quality.”transcript, stated

Alternatives considered (roads not taken)

Starting GenAI adoption with full-blown RAG and large vector databases

Rejected because: LSEG argues teams should prove value with simpler prompt and single-agent slices before adding RAG complexity.

“But no, you don't need to have a full blown RAG solution where you've got vast amount of data sitting in a vector db, okay? You don't need to start there”transcript, stated

Straight-through AI automation for regulated curation

Rejected because: Humans remain accountable at critical junctions, reducing risk while the agents perform search, extraction, drafting, and QA preparation.

“Why? Because your humans are involved in all the critical junctions through the process.”transcript, stated

Fine-tuning before standard-model value is proven

Rejected because: The presenter says most use cases do not need fine-tuning early, and premature fine-tuning can waste time and make things worse.

“You don't need to fine-tune a model to get value out of it, okay? And especially with the types of use cases”transcript, stated

“It's very easy to fine-tune a model, get it wrong, make things worse.”transcript, stated

Patterns

`multi-agent-orchestration` (regulated-workflow)

The curation workflow chains search, extraction, drafting, QA, enrichment, and human routing agents around business-process state rather than simply passing text between agents.

“having an orchestration of agents step together within a business process allows you to do that, okay? The key thing here is this is not simply agents passing information onto other agents.”transcript, stated

`human-in-the-loop-gate` (regulated-decisioning)

High-risk or low-quality agent outputs are reviewed by analysts, preserving human accountability before data reaches customers.

“If you've exceeded that and you still haven't met the criteria presented up to a human being to go through and review it. All of these steps are critical”transcript, stated

`rag-retrieval` (risk-record-context)

The platform cross-references new articles against existing World-Check records and potential relationships before proposing updates.

“We're cross-referencing data on existing World-Check records, stuff that's already been processed by our system, it's made available to customers.”transcript, stated

`canary-rollout` (ai-adoption)

LSEG describes a staged maturity roadmap: prove value with prompts, then add agents, RAG, orchestration, interoperability, and fine-tuning as confidence grows.

“you're starting at prompt only, you're moving to the single agent action. You're then implementing something like retrieval augmented generation”transcript, stated

`tool-use-orchestration` (curation-pipeline)

The ingestion workflow routes content through queued processing, Bedrock-backed entity extraction, fuzzy matching, inclusion criteria, and query feedback.

“content coming in through the platform, it's being acquired, it's putting it onto an SNS queue there, and then we can start processing that data”transcript, stated

Tradeoffs

How far to automate regulated curation.

Chose augmentation with human accountability at critical junctions over straight-through AI automation.

Rationale: The approach accelerates work while avoiding added regulatory risk and preserving explainability for customer-facing data.

“But one of the key things that you haven't done is you haven't added risk, okay? Why? Because your humans are involved in all the critical junctions”transcript, stated

How to introduce GenAI into production.

Chose start with prompt-only, low-risk single actions over jumping directly to complex RAG and multi-agent orchestration.

Rationale: Small augmentations let teams prove business value quickly while managing risk before adding more complex automation.

“the idea behind a prompt only scenario here, okay? Is it allows you to prove out something very quickly and very easily with relatively no risk.”transcript, stated

Gotchas (what bit them)

Straight-through AI can misclassify context-sensitive media.

Root cause: Flag-based extraction may confuse fictional or irrelevant content with actual financial-crime risk.

Resolution: LSEG combines AI curation with human expertise and review instead of relying on fully automated decisions.

“it's actually a movie, okay? And it's a review about that movie. But when you look at it from an AI perspective, all the flags tick all the boxes”transcript, stated

Scaling curation by hiring alone is unsustainable.

Root cause: News volume, language coverage, and specialized risk knowledge exceed what can be expanded through human staffing alone.

Resolution: AI handles toil such as summarization, extraction, and drafting so analysts can focus on higher-judgment tasks.

“you can't just simply scale that using human beings, it's unsustainable, okay? And at the same time, trying to find individuals with the sort of language knowledge”transcript, stated

Agent interoperability is hard to get right too early.

Root cause: MCP-style integration across many agents adds complexity before the organization has matured simpler AI workflow stages.

Resolution: The speaker recommends leaving MCP/interoperability to the end of the maturity journey.

“The big challenge though is building out MCP in an automated way where you've got a whole series of agents being tied together is actually quite difficult to do”transcript, stated

Evaluation

LSEG emphasizes production value metrics, speed, quality, accuracy, compliance, human validation, and analyst efficiency rather than demo-only AI adoption.

Metrics: hours-to-minutes update speed · accuracy · quality · analyst efficiency · content capacity · regulatory compliance

“Demoing metrics is what really allows AI play adoption to really accelerate 'cause that's the key win that you get.”transcript, implied

“We've maintained our commitment to accuracy and compliance by keeping humans in the loop of that decision-making process”transcript, stated

Outcomes

World-Check updates moved from hours to minutes.
Analysts and SMEs gained more time for deeper judgment work.
Content capacity can expand without proportional headcount growth.
Humans still validate outputs before customer release.

World-Check uses more than 200 analysts, thousands of sources, 60-plus languages, and serves about 10,000 customers across financial-crime and identity-risk workflows.

3 supporting quotes

“On speed and quality, we are moving from hours to minutes for updates while maintaining accuracy and quality.”transcript, stated

“we're looking to scale more effectively and we're seeing this. We're able to expand content capacity without proportional headcount growth”transcript, stated

“AI accelerates our process from hours to minutes, but our analysts still validate every output before it reaches customers.”transcript, stated