O Redator de Trilha de Auditoria de Decisões de IA: Transforme uma Decisão Assistida por IA em um Registro de Conformidade Defensável
Porque é que este prompt importa
Regulations like the EU AI Act's Article 14 human-oversight requirement and a growing number of U.S. state automated-decision laws (such as NYC Local Law 144-style bias-audit and notice rules) carry per-violation fines, several of which now run into tens of thousands of dollars per undocumented incident. More damaging than the fine itself: without a record drafted at decision time, teams reconstruct the decision from memory and scattered Slack messages weeks later, and the reconstructed account routinely contradicts itself in ways that look like a cover-up to opposing counsel — a reconstruction problem, not necessarily a wrongdoing problem, that still sinks the case.
Para que o usamos
An HR coordinator used an AI resume-screening tool to auto-reject 40 candidates for a mid-level engineering role on a Tuesday afternoon. On Friday, legal receives a demand letter from a rejected candidate's attorney alleging discriminatory screening, and compliance needs a complete audit-trail record for this specific candidate within 48 hours — including what data went into the AI tool, what it recommended, and who signed off — before legal can respond.
Prompt
Role: Act as an AI governance and compliance documentation specialist who has written audit trails that held up under regulator review and legal-discovery scrutiny. Context: My organization used an AI/LLM-based tool as part of a decision affecting a person outside the engineering team — for example a hiring screen, a credit or loan decision, a content-moderation action, a pricing decision, or a support denial. Decision type: [DECISION TYPE]. The AI tool used: [AI TOOL/MODEL NAME + VERSION]. Decision date: [DECISION DATE]. Applicable regulatory framework, if any: [REGULATORY FRAMEWORK, e.g. "EU AI Act Article 14 human oversight requirement" / "state-level automated decision disclosure law" / "internal AI governance policy" / "none identified yet"]. Here is what I know about the decision: Input data provided to the AI: [INPUT DATA SUMMARY] AI output or recommendation: [AI OUTPUT OR RECOMMENDATION] Human reviewer name and role: [HUMAN REVIEWER NAME/ROLE] What the human reviewer did with the AI output: [APPROVED AS-IS / OVERRODE / MODIFIED — AND WHY] Required retention period for this record, if known: [RETENTION PERIOD] Task: Produce a structured audit-trail entry documenting this AI-assisted decision, suitable for a compliance file or a regulator / legal-discovery request. Constraints: - Do not invent or infer any fact I have not provided — including dates, reviewer names, data fields, or reasoning the AI model might have used. If a required field is missing, label it "NOT PROVIDED — must be filled in before filing" rather than guessing a plausible value. - Do not draft legal conclusions or state whether this decision complies with any law — flag applicable-seeming requirements only as "may be relevant, confirm with legal/compliance" and stop there. - Keep the tone factual, neutral, and third-person, as if this record could be read aloud in a hearing — no defensive language, no hedging phrases like "to the best of our knowledge," and no speculation about intent. - Explicitly separate what the AI recommended from what the human reviewer actually decided — these must never be merged into one combined statement, since the accountable decision-maker is the human, not the model. - If the human reviewer overrode or modified the AI output, state the override reason in the reviewer's own words if provided, or flag it as "NOT PROVIDED" if not — never infer it. Output Format: 1. Header block: Decision Type | Date | AI Tool & Version | Applicable Framework (flagged, not asserted) 2. "Input Summary" — bullet list of what was fed to the AI, in plain factual language 3. "AI Output" — exact restatement of what the AI recommended, clearly labeled as AI-generated, not a statement of fact 4. "Human Review" — reviewer identity, action taken (approved / overridden / modified), and stated reasoning 5. "Gaps Flagged" — a bare list of any NOT PROVIDED fields that must be filled before this record is filing-ready 6. "Retention Note" — the stated retention requirement, or "NOT PROVIDED — confirm with compliance" if absent
Resultado
Decision Type: Candidate Screening — Automated Resume Rejection Date: 2026-10-06 AI Tool & Version: ScreenFit AI Resume Matcher v4.2 Applicable Framework: NYC Local Law 144 automated employment decision tool bias-audit requirement may be relevant — confirm with legal/compliance.
Input Summary: - Candidate resume (PDF, 2 pages) submitted via careers portal on 2026-10-04 - Job requisition: Senior Backend Engineer, req #ENG-2291 - Screening criteria configured in tool: minimum 5 years backend experience, required Python/Go proficiency, degree-equivalent experience accepted
AI Output: ScreenFit AI Resume Matcher scored the candidate 41/100 against the configured criteria and recommended "DO NOT ADVANCE," citing "insufficient backend-specific keyword match" and "employment gap exceeding 18 months unexplained in resume text." This is the AI tool's generated output; it is not a statement of fact about the candidate's actual qualifications.
Human Review: Reviewer: J. Alvarez, Talent Acquisition Coordinator. Action taken: Approved the AI recommendation as-is; no override. Stated reasoning: NOT PROVIDED — must be filled in before filing. The recruiter did not record a rationale beyond accepting the tool's score; the recruiter should be asked directly whether the resume was independently reviewed before rejection.
Gaps Flagged: - Human reviewer's independent reasoning for accepting the AI score: NOT PROVIDED - Confirmation of whether a bias audit has been run on this tool configuration for this job requisition: NOT PROVIDED - Candidate's protected-class information, if any claim is later made: NOT PROVIDED, do not infer
Retention Note: NOT PROVIDED — confirm the required retention period with compliance before filing; most applicable employment-decision frameworks require a minimum of 2+ years.
Toda equipe que adota uma ferramenta de IA para uma decisão com consequências reais acaba recebendo o mesmo pedido do jurídico ou da conformidade: mostre o registro. Não um resumo de memória semanas depois, mas o que foi dado à IA, o que ela recomendou, quem revisou e o que essa pessoa realmente decidiu — escrito próximo ao momento em que aconteceu. A maioria das equipes não tem isso, porque ninguém projeta uma trilha de auditoria até o dia em que precisa dela desesperadamente.
Este prompt trata essa lacuna como o ponto central. A seção Role atribui à IA o papel de especialista em documentação de conformidade em vez de um assistente genérico, porque o estilo de escrita que parece bom em um post de blog — confiante, fluido, preenchendo pequenas lacunas com inferências que parecem razoáveis — é exatamente o estilo que prejudica uma organização sob escrutínio judicial. Um registro com detalhes inventados é pior do que nenhum registro.
A seção Constraints é onde a engenharia real acontece. Proibir fatos inventados e conclusões jurídicas é a parte fácil; a instrução mais difícil é forçar o modelo a separar "o que a IA recomendou" de "o que o humano decidiu" em duas declarações distintas, nunca mescladas. Regulamentações como as disposições de supervisão humana do EU AI Act existem precisamente porque a responsabilidade deve recair sobre uma pessoa identificada, não sobre um modelo — então um registro que mistura essas duas linhas anula silenciosamente seu próprio propósito.
A seção "Gaps Flagged" no formato de saída é a decisão de design que torna isso genuinamente útil, e não apenas um modelo de boa aparência. A maioria da documentação de conformidade falha não porque alguém mentiu, mas porque ninguém percebeu que um campo estava faltando até que um regulador ou o advogado adversário o solicitasse. Fazer a IA listar explicitamente "NOT PROVIDED — must be filled in before filing" transforma um registro incompleto em uma lista de verificação acionável, em vez de uma falsa sensação de completude.
Execute este prompt imediatamente após qualquer decisão assistida por IA que afete uma pessoa fora da equipe — contratação, crédito, moderação, exceções de preço — enquanto os detalhes ainda estão frescos o suficiente para preencher as lacunas que ele sinaliza.