BotCurrent↗
← Back to examples

Official recipe

Prepare an evidence-checked AML alert review

Referenced transactions may not support the alert’s conclusions.

Read original recipe
Suggested result

A calculation-linked findings sheet, information gaps and a versioned human-review record.

What you need: Source-named agent: OpenAI Agents SDK; Amazon Bedrock. Suggested inputs: Case fixture; Review criteria. Use an offline synthetic worksheet. Keep revision-specific review records. Workflow costs and current compatibility remain unverified.

Proposed: Access to real case data, dissemination of allegations or any filing or account action.

Official recipe. BotCurrent has not independently tested this workflow.

Setup and access

Proposed inputs

  1. Name

    Case fixture

    Owner

    Compliance reviewer

    Minimum scope

    One wholly invented alert and transaction table

    Quality check

    Check unique IDs, time zones and currency consistency.

  2. Name

    Review criteria

    Owner

    Policy owner

    Minimum scope

    Owner-authored fictional pattern definitions

    Quality check

    Separate training conditions from real regulatory requirements.

Access and setup — proposed or unknown

Proposed minimum

  1. Use an offline synthetic worksheet.

  2. Keep revision-specific review records.

Compatibility unknowns

  1. Unknown: deployed model access, production controls and jurisdiction-specific compliance obligations.

External actions requiring approval

  1. Access to real case data, dissemination of allegations or any filing or account action.

Detailed steps and output format

At a glance

Suggested result
A calculation-linked findings sheet, information gaps and a versioned human-review record.Suggested; untested.
Tool
OpenAI Agents SDK; Amazon BedrockSource-named; check current access.
Suggested inputs
Case fixture; Review criteriaProposed; access notes below.
Human review
Proposed: Access to real case data, dissemination of allegations or any filing or account action.Suggested boundaries; source approvals below.
Setup notes
Use an offline synthetic worksheet. Keep revision-specific review records.Guide notes; proposed setup untested.
Cost
Workflow cost not established. Check current tool pricing, setup and human review costs before trying it.No measured cost or time added.
Content depth
Source-linked example · suggestions untestedDepth does not establish successful execution.
Check source evidence and limits ↓

Source-backed example. Practical decision guidance is proposed and untested; this is not a full implementation guide.

The problem

Referenced transactions may not support the alert’s conclusions.

Suggested task · untested

Proposed: Assemble an offline alert-evidence review for a qualified compliance reviewer.

What you could produce

A calculation-linked findings sheet, information gaps and a versioned human-review record.

Where these suggestions come from
  • Core problem: Our suggestion; no result reported by the source
  • Agent job: Our suggestion; no result reported by the source
  • Expected output: Our suggestion; no result reported by the source
  • Relevant job roles: Our suggestion; no result reported by the source
  • Benefit type: Our suggestion; no result reported by the source
Suggestions and their limits

Describes the proposed purpose and review output; measured benefit and successful delivery are not established.

Editorial synthesis based on the linked source. This framing does not add evidence of successful delivery or results.

Before you try it · suggestions untested

Original decision guidance and first check are proposals. The source supports the attributed task; this guidance has not been executed.

Read the attributed source evidence and original caveats below ↓

Why use this

  1. References alone do not validate calculations.

  2. Changed evidence needs fresh approval.

Why an agent / simpler alternative

Useful judgment

Explain uncertainty without alleging wrongdoing.

Simpler alternative

Use fixed rules for arithmetic-only checks.

Unsuitable when

The intended output is an autonomous allegation, account restriction or regulatory filing.

Expected deliverable and review checks

Proposed artifact

A calculation-linked findings sheet, information gaps and a versioned human-review record.

Critical checks

  1. Recompute referenced amounts.

  2. Use explicit time zones.

  3. Separate signals from decisions.

Missing output evidence

Unknown: no independently inspected run artifact or live outcome.

First check — proposed, unexecuted

Status

proposed_unexecuted

Synthetic fixture

Four fictional transactions with one currency mismatch and a narrative that cites the correct IDs but overstates their total.

Acceptance checks

  1. Catch the mismatched currency.

  2. Reject the unsupported total.

  3. Retain an unresolved posture rather than allege intent.

Stop conditions

  1. Real customer records, filing instructions or requests to infer criminal intent.

Tasks: Research & discoveryWriting & contentData & analysis

Useful for: General use

Editorial relevance from the title and source-supported actions; not evidence of the operator’s occupation.
OpenAI Agents SDK; Amazon BedrockIndependent testing: No
Failure handling, costs and limits

Failure signals and proposed responses

  1. Signal

    Cited total is wrong

    Why it matters

    References can conceal errors.

    Proposed response

    Recompute; reject unsupported findings.

  2. Signal

    Approval predates revision

    Why it matters

    Evidence has changed.

    Proposed response

    Require qualified re-review.

Effort and cost — unknown

Setup work

  1. Define evidence and calculation rules.

Ongoing review

  1. Review contradictions and the current revision.

Unknowns

  1. Unknown: implementation effort, review time, usage charges and error frequency. No savings estimate established.

Limits

  1. A synthetic exercise does not validate an AML program.

Full source evidence

Official recipe · Tool workflows

Original task title: Prepare an evidence-checked AML alert review

Editorial candidate. This is source-backed reporting or a documented workflow. BotCurrent has not independently tested this task. A proposed Canadian adaptation is not evidence of a Canadian deployment or compliance certification.

Original source

https://github.com/openai/openai-cookbook/blob/main/examples/partners/AWS/evidence_grounded_aml_agent_with_bedrock.ipynb ↗

Attributed to: deepakjain108; OpenAI Cookbook AWS partner contribution

Supporting source context

Project documentation supporting the linked recipe’s scope, permissions and limits. These references do not establish an independent test or add counted jobs.

Official recipe
Original action type
tool_action
What the source reports
Reads a synthetic case and transactions, applies explicit pattern checks, validates evidence-linked findings, and optionally drafts a narrative from an approved analysis revision.
Source outcome limits
Default execution is a deterministic fixture, not an agent run. Demonstration patterns are not regulatory rules or proof of wrongdoing. No independent execution test.
Evidence attribution caveat
This official recipe describes a procedure; it does not prove successful execution or independently reproduced results.
Source approval description
Qualified-human approval is required before optional drafting; no filing tool is provided.
Named tools
OpenAI Agents SDK; Amazon Bedrock
Model or version
openai.gpt-5.6-sol, as specified in the recipe
Publication date
2026-07-23
Source last checked
2026-10-14
Demonstrated in Canada
Unknown
Canadian availability
Unknown
Proposed Canadian adaptation
Proposed: an authorized local owner must verify availability, privacy, policy and professional-review requirements. Canadian execution is unknown.
Recommended approval boundaries
Proposed: Access to real case data, dissemination of allegations or any filing or account action.