CDD Agent — How It Works

The analyst and customer journey from intake through approval

📋 The CDD Journey CDD Agent takes a corporate customer from initial intake through evidence gathering, AI-assisted report drafting, analyst review, and manager sign-off. At each stage, the analyst remains in control — the AI does the research and writing, but every finding must pass human review before approval.
1

Intake & Case Setup

Analyst creates the case, adds evidence, and sends the initial questionnaire

▼
1.1

Create the case

The analyst creates a new CDD case via the New Case wizard. They enter the company name, registration number, country, and website URL.

Readiness tip:

  • A website-only case is a preliminary case — it will produce a first draft, but coverage will be limited.
  • Adding an annual report (strongly recommended) and sending the initial questionnaire makes the case ready for a full first run with strong financial and UBO coverage.
1.2

Add evidence documents

The analyst uploads key documents: annual report, registry extract, and/or ownership documentation. More evidence means better analysis quality and fewer open questions after the first run.

Evidence priority:

  • Annual report — improves financial profiling and revenue data
  • Registry extract — confirms legal registration, address, and directors
  • Ownership / shareholder list — required for UBO identification when structure is layered
1.3

Send initial questionnaire to customer

The analyst sends the initial CDD questionnaire directly to the customer via email or a secure link. The questionnaire captures information the AI cannot determine from public evidence alone.

Questionnaire covers:

  • Purpose and nature of the banking relationship
  • UBO declarations (beneficial owners above threshold)
  • Source of funds and source of wealth
  • Expected transaction patterns and payment geographies
  • PEP and sanctions self-declaration

Why send before the first run:

  • If the customer responds before the workflow runs, their answers are incorporated as evidence and the draft will be more complete.
  • If they respond after, the analyst can trigger a rerun to incorporate the new information.
2

First Workflow Run (AI Processing)

The AI gathers evidence, screens for risks, and drafts the CDD report

▼
2.1

Evidence gathering & screening

The system automatically fetches and processes all evidence sources — websites, uploaded PDFs, and questionnaire responses — and performs automated sanctions and PEP screening.

What happens automatically:

  • All websites are fetched and chunked for analysis
  • All PDFs are parsed and chunked page-by-page
  • Geographic risk is assessed against EU and FATF policy lists
  • Sanctions, PEP, and adverse media screening runs on company names and identified individuals
  • Questionnaire responses (if received) are incorporated as evidence
2.2

AI drafts the CDD report

Three AI agents work in sequence to produce a compliance-grade CDD draft: Writer, Critic, and Updater.

Agent sequence:

  • Writer — produces Draft v1 with inline citations from the evidence gathered
  • Critic — reviews Draft v1 for compliance gaps, missing sections, and unsupported claims
  • Updater — applies the Critic's findings to produce Draft v2 with a delta log and open questions

What the Gatekeeper checks:

  • All required CDD sections are present
  • Citations are grounded in actual evidence (no fabricated figures)
  • Screening and geographic risk are addressed
  • Conclusions are consistent with the evidence
2.3

Evaluation & quality scoring

After the draft is complete, an evaluator scores each section for evidence coverage, data completeness, and critical gaps. The result determines whether the case can proceed to sign-off.

Quality gates before approval:

  • Gatekeeper must PASS (no blocked sections)
  • Evaluation must show no critical-section data gaps
  • Open questions must be resolved or incorporated
  • Manual review items must be decided
3

Analyst Review & Resolution

Analyst resolves open questions, manual review items, and quality blockers

▼
3.1

Workbench: resolve open questions

The Analyst Workbench is the primary in-case workspace. All open questions, manual review items, pending-incorporation tasks, and quality blockers surface here as actionable task cards.

Types of tasks in the workbench:

  • Open questions — tied to specific report sections; analyst can request customer info or resolve directly
  • Pending incorporation — new evidence or manual input that needs to be incorporated into a section
  • Manual review items — sanctions/PEP/adverse media hits requiring a human decision
  • Quality blockers — Gatekeeper failures identifying specific sections to fix
3.2

Customer follow-up (if needed)

If the AI identified open questions that cannot be answered from public evidence, the analyst sends a follow-up questionnaire to the customer. The case enters a "Waiting for Customer" state while the response is pending.

Flow:

  • Analyst sends follow-up questionnaire via the Questionnaire tab or Workbench CTA
  • Customer receives a secure link and submits answers
  • Case shows "Waiting for Customer" on the dashboard until response arrives
  • When response arrives, the Workbench prompts the analyst to rerun the workflow to incorporate the answers
3.3

Rerun after new evidence (if needed)

If the customer responded, new documents were added, or quality blockers need a fresh AI pass, the analyst can trigger a full rerun. The system explains what will change before the rerun starts.

Before rerun, the analyst sees:

  • Whether the rerun will revoke existing approval (if the case was already approved)
  • What changed since the last run (new evidence, questionnaire response)
  • The scope: rerun is always a full rerun — partial reruns are not supported

After rerun:

  • A new Draft v2 is produced incorporating all new evidence
  • Evaluation scores are refreshed
  • Any previous approval is revoked and must be re-granted
4

Manager Sign-off & Approval

Manager reviews the completed case and grants approval

▼
4.1

Analyst hands off for approval

Once all blockers are resolved, the case reaches "Ready for Approval" status. The analyst's role is to prepare the case — they cannot approve their own work. Approval is reserved for managers and admins.

Analyst responsibilities at handoff:

  • All open questions resolved or accepted
  • All manual review items decided
  • No Gatekeeper blockers remaining
  • Draft v2 reviewed and confirmed as final
4.2

Manager reviews and approves

The manager opens the case directly from the Cases Work Queue (no summary-only detour). They review Draft v2, check the blocker rail, and approve when satisfied. A blocked case cannot be approved.

Approval gate:

  • The system re-verifies all approval blockers at the moment of approval
  • If any blocker remains, approval is rejected with a specific reason
  • Approval is recorded in the audit trail with the approver identity and timestamp
4.3

Approved case

Once approved, the case is locked from further modifications. The approval is recorded in the approval history with the approver's identity and timestamp. If a rerun is triggered later, the approval is automatically revoked.

Audit trail includes:

  • Approval action with user identity and timestamp
  • All prior review and decision history
  • Any approval revocations (e.g., triggered by rerun)
  • Evidence provenance: all sources, citations, and screening results
👥 Who does what
Role What they do
Analyst Creates cases, adds evidence, runs workflow, resolves open questions and manual review items, prepares the case for sign-off
Manager Reviews and approves completed cases; can also work through analyst tasks when needed
Viewer Read-only access — can view cases, reports, and evidence but cannot modify or approve