Anti-Bribery and Corruption Risk Detection Prompt

This prompt is an ingress filter for global compliance teams and AI platform engineers who need to automatically triage unstructured business communications—emails, chat logs, call transcripts, and CRM notes—for potential violations of the FCPA and UK Bribery Act. Its job is not to provide legal advice or make final determinations, but to act as a high-recall, structured classification layer that prevents risky content from entering downstream systems like CRM logging, generative AI summarization, or data warehouses without review. The ideal user is an engineer integrating this prompt into a compliance middleware layer, or a compliance officer configuring the risk thresholds that determine routing behavior.

Deploy this prompt when you need a consistent, auditable first pass over high-volume communications where manual review is impossible. It is specifically designed to detect bribery indicators (quid pro quo language, win-at-all-costs phrasing), facilitation payment discussions (grease payments, customs expediting), improper gift and entertainment language (luxury travel, excessive hospitality), and third-party intermediary risk signals (undisclosed agents, vague consulting fees). The prompt expects an [INPUT] text and an optional [CONTEXT] with metadata like sender, recipient, and jurisdiction. It outputs a structured [OUTPUT_SCHEMA] with a risk level, detected signals, and a confidence score, enabling deterministic routing to a human review queue for high-risk items or automated processing for low-risk ones.

Do not use this prompt as a standalone compliance program. It is a classification and triage tool, not a replacement for compliance officers, legal counsel, or a comprehensive anti-corruption framework. It will not understand nuanced local law exceptions, and it cannot investigate or verify the truth of a statement—it only flags language patterns. Avoid using it in isolation for final employment decisions, contract rejection, or regulatory filings. The output must always be treated as a signal for human review in high-risk cases, and you should implement a feedback loop where compliance officers can mark false positives and false negatives to continuously refine your eval datasets and prompt thresholds.

The prompt template below is the core instruction set for an AI system performing anti-bribery and corruption (ABC) risk detection. It is designed to be strict, evidence-bound, and auditable. The model is instructed not to invent facts, to flag ambiguity for human review, and to map its findings to a predefined risk taxonomy. This is not a conversational prompt; it is a classification contract. Use it as the system prompt or the primary instruction block in a structured API call where the input text is injected into the [INPUT] placeholder.

codeCopy
You are an anti-bribery and corruption risk classifier. Your task is to analyze the provided business communication for indicators of bribery, corruption, facilitation payments, improper gifts, or third-party intermediary risk.

# RISK TAXONOMY
Use the following categories to classify the text. A single communication may have multiple flags.
- BRIBERY_OFFER: An explicit or implied offer of something of value to influence a decision.
- BRIBERY_REQUEST: A solicitation for a bribe or kickback.
- FACILITATION_PAYMENT: A small payment to expedite a routine governmental action.
- IMPROPER_GIFT: A gift, hospitality, or entertainment that is excessive, secretive, or tied to a pending decision.
- THIRD_PARTY_RISK: Language suggesting a third party (agent, consultant, partner) may be used to make improper payments.
- AMBIGUOUS: Language that is suspicious but lacks sufficient context for a definitive classification.
- NO_RISK: No indicators of bribery or corruption are present.

# INPUT
[INPUT]

# CONSTRAINTS
- Do not invent facts not present in the text.
- If the text is ambiguous, classify it as AMBIGUOUS and flag it for human review. Do not assume innocence.
- Cite the specific phrases or sentences that support each classification.
- If no risk is found, classify as NO_RISK and provide a brief, factual summary of the text's benign content.

# OUTPUT SCHEMA
Respond with a valid JSON object conforming to this schema:
{
  "risk_flags": ["BRIBERY_OFFER", "IMPROPER_GIFT"], // Array of strings from the RISK TAXONOMY
  "citations": [
    {
      "flag": "BRIBERY_OFFER",
      "quote": "We can make it worth your while.",
      "rationale": "This phrase implies an offer of value in exchange for a favorable decision."
    }
  ],
  "requires_human_review": true, // Boolean. Must be true if AMBIGUOUS or any risk flag is present.
  "summary": "The email contains a veiled offer of value linked to a contract award decision."
}

To adapt this template, replace the [INPUT] placeholder with the text you want to classify. The [RISK_TAXONOMY] can be expanded to include more granular categories specific to your organization's policy, such as CHARITABLE_CONTRIBUTION or POLITICAL_DONATION. The [OUTPUT_SCHEMA] is a critical contract for your application; ensure your downstream parser expects this exact JSON structure. For high-stakes deployments, consider adding a [CONFIDENCE_SCORE] field to the output schema to allow for more nuanced routing, and always set requires_human_review to true for any non-NO_RISK finding to create a mandatory human-in-the-loop checkpoint.

Before integrating this prompt into a production pipeline, you must test it against a golden dataset of known-clean and known-risky communications. Common failure modes include the model hallucinating a risk where none exists (a false positive) or, more dangerously, classifying a veiled bribe as NO_RISK (a false negative). Your evaluation suite should specifically test for obfuscated language, such as code words or euphemisms, and confirm that the model correctly defaults to AMBIGUOUS and requests human review rather than making a risky assumption.

This prompt is designed as a classification gateway, not a final decision engine. It should sit at the ingress layer of any workflow that processes business communications, expense reports, third-party due diligence documents, or contract negotiations. The prompt's job is to triage content into risk tiers and extract structured indicators that downstream systems and human reviewers can act on. Because the regulatory stakes under the FCPA and UK Bribery Act are high, the implementation harness must enforce deterministic routing, audit logging, and human-in-the-loop review for anything above a low-risk threshold.

Wire the prompt into an application by first constructing the [INPUT] from the raw text source, ensuring you strip any prior system instructions or metadata that could cause prompt confusion. The [JURISDICTION] placeholder should be populated from a configuration map based on the entity's operating regions, defaulting to 'FCPA and UK Bribery Act' if unknown. The [RISK_THRESHOLD] should be set to a low sensitivity level (e.g., 'LOW') to avoid false negatives, as missing a facilitation payment signal is costlier than a false positive. After model invocation, parse the JSON output and implement a strict routing switch: route risk_level: HIGH and risk_level: CRITICAL directly to a compliance review queue with an attached audit record. Route risk_level: MEDIUM to a secondary review bucket for batch analysis. Only risk_level: LOW with a confidence score above 0.95 should be allowed to pass without human review. Log every classification, the raw model response, and the routing decision to an immutable audit store for regulatory examination.

For model choice, use a model with strong instruction-following and low refusal rates on sensitive topics, as over-refusal on legitimate compliance screening creates a dangerous blind spot. Implement a retry layer that catches JSON parse failures and malformed outputs, re-invoking the prompt with the error message appended as a [REPAIR_INSTRUCTION]. If the model fails to return valid JSON after two retries, escalate the entire input to human review and log the failure for prompt debugging. Never allow a parse failure to silently drop a message. Finally, integrate an eval harness that runs a golden dataset of known bribery indicators, clean business language, and edge cases (e.g., legitimate gift policies) against every prompt version before deployment to measure precision, recall, and over-flagging rates.