Um Claude Skill que pega sua lista de vagas, seu inventário de ferramentas e seu pacote de evidências e devolve uma matriz de controles mostrando quais obrigações de AI em emprego incidem sobre cada vaga, quais controles você consegue de fato comprovar e quais lacunas acumulam multa por dia. Ele avalia os artefatos que você fornece — uma URL de auditoria de viés publicada, um export de timestamps de consentimento, um runbook de exclusão — e não as afirmações do vendor, e nunca emite um veredito de conforme/não conforme. O output é o que você entrega ao jurídico para que a primeira hora deles comece nas decisões de julgamento em vez de em “quais ferramentas vocês usam, e onde?”
Quando usar
Você vai ligar um entrevistador com AI, uma triagem de vídeo com AI ou ranqueamento de currículos, e alguém perguntou qual é a exposição antes de subir para produção.
Abriu uma vaga em uma jurisdição que seu stack ainda não atendia. Vagas remotas são o gatilho habitual: uma única publicação “US-wide” pode acionar obrigações de New York City, Illinois e Califórnia de uma vez.
A renovação anual da auditoria de viés de NYC está chegando e você precisa saber quais ferramentas estão no escopo antes de contratar o auditor — os contratos de auditoria são cotados por ferramenta.
Um candidato perguntou como a decisão foi tomada, ou pediu a exclusão da entrevista dele, e ninguém sabe qual é o runbook.
Quando NÃO usar
Para produzir a auditoria de viés em si. A NYC Local Law 144 exige um auditor independente sem relação de emprego ou financeira com o empregador que comprometa a independência. Um skill que você mesmo roda não pode ser esse auditor. Este skill verifica se existe uma auditoria independente, se ela está dentro do prazo e se está publicada na forma exigida.
Como aprovação formal. Não há score agregado no output, de propósito. Um número convida alguém a tirar print e colar num board deck como atestado de saúde.
Depois que chega uma ação ou uma notificação extrajudicial. Nesse ponto o conjunto de artefatos é material de discovery e o jurídico conduz. Gerar uma avaliação interna paralela sobre os mesmos fatos cria um documento que você não precisava.
Em stacks sem scoring automatizado. Se uma pessoa lê cada candidatura e nada ranqueia ou filtra, a maior parte da matriz não incide.
Fora dos EUA. A matriz incluída cobre regras federais, estaduais e municipais dos EUA apenas. As obrigações de emprego do Anexo III do EU AI Act são outro escopo e precisariam de um arquivo de referência próprio.
Setup
Coloque o bundle em apps/web/public/artifacts/ai-interview-compliance-audit-skill/SKILL.md no seu diretório de skills do Claude Code, com a pasta references/ ao lado.
Peça ao jurídico para revisar a matriz uma vez.references/1-jurisdiction-matrix.md guarda cada parâmetro legal contra o qual o skill avalia — a janela de aviso de 10 dias úteis de NYC, a janela de destruição de 30 dias de Illinois, o piso de retenção de quatro anos da Califórnia. Uma passada de revisão, e depois mantenha a data checked: em dia. O skill se recusa a avaliar se essa data tiver mais de 90 dias.
Preencha o inventário.references/2-tool-and-stage-ledger.md pede vagas com locais de trabalho e uma linha por ferramenta que toca um candidato entre a candidatura e a oferta. Inclua as ferramentas que você não comprou para triagem — os match scores do ATS e os fit scores das ferramentas de sourcing são os que os times esquecem, e eles ranqueiam.
Registre a capacidade de revisão. A Parte C do inventário pergunta quantas candidaturas uma pessoa de fato revisa por vaga. “400 candidatos, o recruiter trabalha o top 40” é o dado que transforma uma ferramenta de ranqueamento em um filtro funcional, não importa como a tela de configuração a chame.
Monte o pacote de evidências conforme references/3-evidence-pack-index.md. Quinze linhas de artefatos; escreva MISSING onde estiver faltando. Essa primeira passada costuma ser onde o achado real aparece.
O que o skill faz de fato
Seis passos. O escopo é resolvido antes de qualquer controle ser avaliado, porque a mesma configuração é lícita em um estado e uma violação por dia em outro — e porque o erro mais comum na prática é delimitar a auditoria por onde fica a sede da empresa em vez de por onde os candidatos se candidatam.
Resolver o escopo. Monta uma matriz vaga × jurisdição × etapa. Se uma vaga não tem lista de locais de trabalho, o skill para e pergunta em vez de inferir pelo endereço da empresa. Vagas abertas a estados para os quais o pacote de evidências não tem artefatos voltam como gap, não como not-applicable.
Classificar cada ferramenta pela configuração. Quatro perguntas: o output libera uma transição de etapa em algum limiar?; ele ordena uma lista que o recruiter trabalha de cima para baixo?; uma pessoa vê o score antes de decidir?; ele analisa vídeo ou áudio buscando características usadas para avaliar aptidão? O rótulo do próprio vendor vai numa coluna separada e nunca é a resposta. O teste da Local Law 144 depende de o output auxiliar substancialmente ou substituir a decisão discricionária, o que é um fato sobre a sua configuração, não sobre o produto deles — e o vendor tem incentivo para ler isso de forma estreita.
Avaliar cada controle sobre evidência, com citação. Quatro status: evidenced (artefato citado literalmente, com o trecho que faz o trabalho), unevidenced (provavelmente ok na prática, nada a mostrar), gap e counsel-review. Não existe compliant. Um controle sem citação não pode ser evidenced — essa regra é o que impede um relatório fluente sobre documentos que ninguém tem.
Rodar as checagens determinísticas de data. O relógio da auditoria de viés, sinalizado aos 10 meses em vez de aos 12 para haver folga para contratar o auditor. O prazo de aviso contado em dias úteis contra a primeira execução da ferramenta. O consentimento de Illinois capturado antes da entrevista, não embutido num aceite posterior — a ordem é o controle inteiro. O tempo de resposta do runbook de exclusão contra a janela de 30 dias, e se ele nomeia os destinatários downstream e os backups ou apenas o ATS principal.
Ordenar a remediação por exposição. New York City conta cada dia em que uma ferramenta no escopo opera fora de conformidade como uma violação separada, e cada aviso omitido a um candidato como violação própria, em até $500 pela primeira violação e de $500 a $1.500 pelas seguintes. Uma lacuna de aviso numa vaga de 400 candidatos por mês acumula assim contra um custo fixo de remediação; a mesma lacuna numa vaga pausada, não. A lista ordena por essa diferença.
Emitir o relatório. Matriz, resultados determinísticos, remediação, e depois os itens de counsel-review na íntegra com as duas leituras expostas.
O que mudou recentemente, e por que a matriz é um arquivo
O Colorado é a razão de os parâmetros legais morarem num arquivo de referência datado em vez de na cabeça do modelo. A SB 24-205 — o Colorado AI Act de 2024, com seus programas de gestão de risco e avaliações de impacto anuais — nunca entrou em vigor. A data de início foi de fevereiro de 2026 para junho de 2026, e então a SB 26-189, sancionada em 14 de maio de 2026, a substituiu por um arcabouço mais estreito de aviso e transparência, vigente a partir de 1º de janeiro de 2027: aviso prévio ao uso, uma descrição em linguagem simples dentro de 30 dias de um resultado adverso, direito de pedir revisão humana e um piso de três anos de registros. Qualquer checklist que ainda avalie avaliações de impacto está auditando uma lei revogada. O explicador do Colorado AI Act percorre a substituta por inteiro.
Illinois também se mexeu. O AI Video Interview Act vigora desde 2020, mas a emenda ao Human Rights Act que proíbe efeitos discriminatórios da AI — incluindo CEP como proxy de classe protegida — entrou em vigor em 1º de janeiro de 2026, e as regras de aviso do Department of Human Rights seguem em processo regulatório após um primeiro rascunho retirado. As regulações de sistemas de decisão automatizada da FEHA da Califórnia valem desde 1º de outubro de 2025, com um piso de registros de quatro anos que aponta na direção oposta ao pedido de exclusão de um candidato. O skill expõe essa tensão em vez de resolvê-la.
A ordem executiva de dezembro de 2025 que instrui uma task force do DOJ a contestar leis estaduais de AI está na matriz como contexto, não como controle. Nenhuma regra estadual acima foi afastada. Trate-a como razão para manter os controles documentados e portáveis, não como razão para aposentar uma linha.
Realidade de custo
Por execução de auditoria — cerca de 40-80k tokens de input (matriz, inventário, textos de aviso, docs do vendor) e 6-10k de output. Às tarifas de lista do Claude Sonnet isso dá algo em torno de $0,30-0,60 por execução. Estimativa, a partir do formato de tokens de um stack de três vagas e cinco ferramentas.
Setup — 90 minutos, e esse número só é honesto se o pacote de evidências já existir em algum lugar. Times que o montam pela primeira vez gastam de 4 a 8 horas, boa parte descobrindo que um controle que todo mundo achava resolvido não tem artefato por trás.
Tempo de jurídico economizado — uma primeira passada de advogado trabalhista externo sobre um stack multijurisdicional leva de 8 a 20 horas a $350-700 por hora, boa parte em levantamento: quais ferramentas, quais vagas, quais estados, o que o score faz. Chegar com o inventário preenchido e uma matriz avaliada traz esse levantamento para dentro. As decisões de julgamento continuam sendo faturadas.
O que não economiza — a auditoria de viés independente. Esse é um contrato à parte, cotado por ferramenta, e este skill não o substitui.
Métrica de sucesso
Contagem de unevidenced tendendo a zero. A divisão entre evidenced e unevidenced na primeira execução é a linha de base real. Controles migrando de unevidenced para evidenced sem nenhuma mudança operacional é o resultado pretendido — significa que o artefato agora existe.
Relógio da auditoria de viés nunca além de 10 meses. Uma checagem determinística que nunca deveria disparar duas vezes na mesma ferramenta.
Distribuição do prazo de aviso, não a média. Acompanhe o percentil cinco de dias úteis entre o aviso e a primeira execução da ferramenta por vaga. Médias escondem as vagas rápidas, que são exatamente onde o prazo falha.
Tempo entre vaga em jurisdição nova e matriz avaliada. Deve ficar abaixo de um dia depois que o inventário existe.
vs alternativas
vs um checklist em planilha. O status quo, e ele falha em duas coisas concretas: delimita por localização da empresa em vez de por vaga, e não tem relógio, então uma auditoria de viés envelhece em silêncio além dos 12 meses enquanto a ferramenta segue rodando. Os passos 1 e 4 do skill existem por causa dessas duas falhas.
vs atestados de conformidade do vendor.HireVue, Sapia.ai e outros publicam resumos de auditoria de viés dos modelos deles. Servem, e satisfazem as linhas do lado do vendor. Não satisfazem as suas: o dever da Local Law 144 recai sobre o empregador ou a agência de emprego, e a auditoria que importa cobre a sua configuração e o seu pool de candidatos. O índice de evidências mantém artefatos de vendor e de empregador em colunas separadas para que isso não se misture.
vs um vendor de auditoria dedicado. As firmas que realizam auditorias de viés AEDT independentes fazem o que o skill estruturalmente não pode. Use o skill como a passada de preparação antes de contratar uma — chegar com um inventário de ferramentas já classificado corta o vaivém de definição de escopo — e como a rechecagem entre auditorias quando abre uma vaga em um estado novo.
vs pedir ao jurídico para tocar tudo. Correto para as decisões de julgamento, caro para o inventário. A divisão que este workflow propõe: você é dono do inventário e do pacote de evidências, o jurídico é dono da revisão da matriz e da fila de counsel-review.
Watch-outs
O modelo afirma uma conclusão jurídica.Guarda: o vocabulário de status não tem valor conforme/não conforme e não há score agregado. Decisões de julgamento vão para counsel-review com as duas leituras impressas.
Parâmetros legais desatualizam.Guarda: os limiares moram em references/1-jurisdiction-matrix.md com uma data checked:, e o skill se recusa a avaliar depois de 90 dias. Esse cenário mudou três vezes entre agosto de 2025 e maio de 2026.
Um atestado do vendor conta como a sua conformidade.Guarda: a linha 14 do índice de evidências é marcada como do lado do vendor e não pode satisfazer as linhas de auditoria ou publicação do empregador.
Delimitar pela sede.Guarda: o passo 1 não avança sem locais de trabalho por vaga, e vagas remotas se expandem para cada estado da lista aceita.
Lavagem de classificação.Guarda: a classificação vem de fatos de configuração registrados no inventário, com os dados de capacidade de revisão anexados. Uma ferramenta que ranqueia 400 candidatos onde uma pessoa revisa 40 está fazendo trabalho de decisão, diga o que disser o rótulo.
Ler unevidenced como aprovado.Guarda: essas linhas entram na remediação junto com as lacunas, com o artefato faltante nomeado.
Retenção e exclusão puxando para lados opostos.Guarda: o piso de quatro anos da Califórnia e um pedido de exclusão de Illinois colidem sobre o mesmo registro. O passo 4 sinaliza o conflito; o jurídico decide uma vez e o relatório cita esse memorando.
Stack
O bundle fica em apps/web/public/artifacts/ai-interview-compliance-audit-skill/ e contém:
SKILL.md — a definição do skill
references/1-jurisdiction-matrix.md — parâmetros legais com uma data checked:
references/2-tool-and-stage-ledger.md — inventário preenchível de vagas, ferramentas e capacidade de revisão
references/3-evidence-pack-index.md — mapeamento de controle para artefato mais o andaime de aviso e consentimento
Assume Claude para a execução. O stack auditado normalmente inclui um ATS como o Greenhouse e uma ou mais ferramentas de triagem com AI — HireVue, Sapia.ai ou similares.
---
name: ai-interview-compliance-audit
description: Audit a configured AI interviewing and screening stack against the US employment-AI rules that are live today — NYC Local Law 144, the Illinois AI Video Interview Act, the Illinois Human Rights Act AI amendment, California's FEHA automated-decision-system regulations, and Colorado SB 26-189 — and emit a control matrix plus a remediation list ordered by penalty exposure. Grades controls on evidence you supply, not on vendor marketing claims. Produces a readiness report, never a legal conclusion.
---
# AI interview compliance audit
## When to invoke
Use this skill when someone owns a hiring stack that includes AI screening, AI-scored video or async interviews, AI interviewers, or resume-ranking, and needs to know which controls are evidenced, which are missing, and what to hand counsel. Typical triggers: a new AI interviewing vendor going live, a req opening in a jurisdiction the stack has not served before, an annual bias-audit renewal, or a candidate complaint.
Inputs are configuration and artifacts. The skill reads what the stack actually does — which score gates which stage — and what documents exist to prove each control.
Do NOT invoke this skill for:
- **Producing the bias audit itself.** NYC Local Law 144 requires an *independent* auditor with no employment or financial relationship with the employer that would compromise independence. A skill run by the employer is not independent and cannot satisfy that duty. This skill checks whether an independent audit exists, is within its clock, and is published in the required form.
- **A legal opinion, or a "we are compliant" sign-off.** The output is a readiness report with an evidence status per control. It never emits a compliant/non-compliant verdict.
- **Non-US stacks.** The jurisdiction matrix in `references/1-jurisdiction-matrix.md` covers US federal, state, and city rules only. EU AI Act Annex III employment obligations are a different scope and a different reference file.
- **Stacks with no automated scoring at all.** If humans read every application and no tool ranks, scores, or filters candidates, most of the matrix does not attach and the run is wasted effort. Confirm the classification question in step 2 before a full run.
- **Retroactive defense of a decision already challenged.** Once there is a charge or a demand letter, the artifact set is discovery material. Counsel drives; do not generate parallel internal assessments of the same facts.
## Inputs
- Required: `reqs` — the open or planned requisitions in scope, each with the job's work location(s) and whether remote candidates are accepted from other states. Jurisdiction attaches by where the candidate applies for or performs the job, not by where the company is headquartered. See `references/2-tool-and-stage-ledger.md`.
- Required: `tool_ledger` — every tool touching a candidate between application and offer, with the stage it runs at and what its output does (displayed to a recruiter, sorts a list, sets a threshold, advances or rejects automatically). Same file.
- Required: `evidence_pack` — paths or URLs for the artifacts that prove controls: candidate-facing notice text, consent capture record, published bias-audit summary URL, data-retention policy, vendor DPA, deletion-request runbook. See `references/3-evidence-pack-index.md`.
- Optional: `jurisdiction_matrix_path` — override the bundled matrix with your counsel's maintained copy. Recommended once you have one.
- Optional: `as_of` — the date to run clock arithmetic against. Defaults to today.
## Reference files
- `references/1-jurisdiction-matrix.md` — every statutory parameter the skill grades against, with a `checked:` date. The skill reads thresholds from this file and never from model memory.
- `references/2-tool-and-stage-ledger.md` — fillable inventory template for reqs, tools, stages, and what each score actually gates.
- `references/3-evidence-pack-index.md` — control-to-artifact mapping, plus scaffolding for the notice and consent language each rule requires.
## Method
Six steps. Scope resolves before any control is graded, because the same configuration is lawful in one jurisdiction and a per-day violation in another, and because the most common real-world error is scoping the audit to company headquarters.
### 1. Resolve scope before grading anything
Build a req × jurisdiction × stage matrix from `reqs`. A single remote-eligible req can attach NYC, Illinois, California, and Colorado at once. If any req lacks a work-location list, stop and ask — do not infer jurisdiction from the company address, and do not grade a partial matrix.
Flag reqs open to candidates in states the evidence pack has no artifacts for. That cell is `gap`, not `not-applicable`.
### 2. Classify each tool from configuration, not from its label
For each tool in `tool_ledger`, answer from the configuration:
- Does its output gate a stage transition automatically, at any threshold?
- Does it sort or rank a candidate list that a recruiter works top-down?
- Does a recruiter see the score before making the advance/reject call?
- Is it analyzing a video or audio interview for characteristics used to evaluate fitness?
Read the vendor's own classification as an input to be checked, not as the answer. A vendor has an incentive to say its product merely assists, and the LL 144 test turns on whether the output substantially assists or replaces discretionary decision-making — which is a fact about your configuration, not about their product. Record the classification, the configuration facts that drove it, and dissent from the vendor label explicitly where it exists.
The video/audio question is separate and additive: the Illinois AI Video Interview Act attaches to AI analysis of video interviews for Illinois positions regardless of whether the tool also qualifies as an automated employment decision tool.
### 3. Grade each control on evidence, with a citation
For each applicable control in the jurisdiction matrix, assign exactly one status:
- `evidenced` — an artifact in the evidence pack satisfies the control. Requires a verbatim citation: file path or URL, plus the quoted passage that does the work.
- `unevidenced` — the control is plausibly satisfied in practice but no artifact was supplied. This is not a pass. It is the state that turns into a gap the moment anyone asks for proof.
- `gap` — the artifact exists and does not satisfy the control, or the required artifact does not exist.
- `counsel-review` — the determination turns on a judgment call (whether a given tool substantially assists a decision, whether a notice's placement counts as before the interview). The skill states the facts and the competing readings, and stops.
There is no `compliant` status and no aggregate compliance score. Both invite the reader to treat the report as a conclusion. A control with no citation cannot be `evidenced`, which is the guard against a fluent report about documents nobody actually has.
### 4. Run the deterministic date and arithmetic checks
These are computed, not judged, and they surface first because they are the cheapest failures to fix:
- **Bias-audit clock.** Parse the published date of the most recent bias audit. Flag at 10 months, not at 12, so there is runway to schedule the auditor. Past 12 months the tool is out of clock while still in use — which accrues per-day.
- **Notice lead time.** Compare the candidate-notice timestamp to the tool's first run against that candidate. The NYC requirement is at least 10 business days, counted in business days.
- **Consent ordering.** For Illinois video interviews, confirm consent is captured before the interview, not bundled into a post-interview acknowledgment. Ordering is the whole control.
- **Deletion SLA.** Check the deletion runbook's stated turnaround against the 30-day statutory window, and confirm it names downstream recipients and backups rather than only the primary system.
- **Retention floor.** California's FEHA regulations require automated-decision-system records — selection criteria, outputs, audit findings — kept four years. Flag any retention policy that deletes earlier, including a well-intentioned privacy-minimization policy. Cross-check the deletion SLA against the retention floor and surface the conflict rather than resolving it; that tension is a counsel call.
### 5. Order remediation by exposure, not by effort
Rank gaps by how the penalty accrues. NYC counts each day an in-scope tool runs out of compliance as a separate violation, and each missed candidate notice as its own separate violation — so a missing notice on a high-volume req compounds daily against a fixed remediation cost. A gap on a paused req does not. Give each gap: the accrual shape, the artifact that would close it, and who owns it.
### 6. Emit the report
Control matrix first, then the deterministic-check results, then remediation ordered by exposure, then the `counsel-review` list verbatim.
## Output format
```markdown
# AI interview compliance readiness — as of 2026-07-28
Matrix version: references/1-jurisdiction-matrix.md (checked: 2026-07-28)
## Scope
| Req | Work locations | Attaches |
|---|---|---|
| ENG-411 | NYC + remote US | NYC LL 144; IL AIVIA; IL HRA; CA FEHA ADS |
## Tool classification
| Tool | Stage | Output gates | AEDT (configuration) | Vendor label | Dissent |
|---|---|---|---|---|---|
| HireVue | async video screen | recruiter sees score before advance/reject | yes | "decision support" | yes — score precedes the call |
## Controls
| Jurisdiction | Control | Status | Citation |
|---|---|---|---|
| NYC LL 144 | Independent bias audit within 12 months | evidenced | careers.example.com/aedt — "date of most recent bias audit: 2026-03-14" |
| NYC LL 144 | 10 business days candidate notice | gap | notice fires at invite, 2 business days before |
| IL AIVIA | Consent captured before interview | counsel-review | consent is on the invite page; candidate can start without scrolling |
## Deterministic checks
- Bias-audit clock: 4.5 months elapsed — OK (flags at 10)
- Notice lead time: 2 business days vs 10 required — FAIL
- Deletion runbook: names primary ATS only; no downstream or backup step — FAIL
## Remediation (by exposure)
1. NYC notice lead time — ENG-411 is live and high-volume; each missed notice is a separate violation and each day of use accrues. Fix: move notice to the application confirmation. Owner: TA ops.
## Counsel review
1. IL AIVIA consent placement on ENG-411. Facts: [...]. Competing readings: [...].
```
## Watch-outs
- **The model asserts a legal conclusion.** *Guard:* the status vocabulary has no compliant/non-compliant value and the report has no aggregate score. Judgment calls route to `counsel-review` with both readings stated.
- **Statutory parameters drift.** *Guard:* thresholds live in `references/1-jurisdiction-matrix.md` with a `checked:` date; the skill refuses to grade and warns if that date is more than 90 days old. This landscape moved three times between August 2025 and May 2026.
- **Vendor attestation counted as employer compliance.** *Guard:* the evidence index separates vendor artifacts from employer artifacts, and a vendor bias-audit summary can only satisfy vendor-side rows. The LL 144 duty sits with the employer or employment agency.
- **Scoping to headquarters.** *Guard:* step 1 will not proceed without per-req work locations, and remote-eligible reqs expand to every state in the accepted-candidate list.
- **Classification laundering.** *Guard:* classification is answered from configuration facts recorded in the ledger; the vendor's label is a separate column, and disagreement is printed rather than reconciled.
- **Unevidenced read as a pass.** *Guard:* `unevidenced` rows sort into the remediation list alongside gaps, with the missing artifact named.
- **Retention and deletion pulling opposite ways.** *Guard:* step 4 surfaces the conflict between a deletion request and the four-year records floor as a flagged tension rather than picking one.
# Jurisdiction matrix
```yaml
checked: 2026-07-28
checked_by: REPLACE_WITH_YOUR_NAME
```
The skill reads every threshold from this file and never from model memory. Update `checked:` whenever you re-verify against primary sources. The skill warns and refuses to grade if the date is more than 90 days old — this landscape moved three times between August 2025 and May 2026, and a stale matrix produces a confident report against rules that no longer apply.
This is a working parameter table maintained by the employer, not legal advice. Have counsel review it once and then keep it current.
---
## NYC Local Law 144 of 2021 — automated employment decision tools
**Status:** in effect; enforced by the Department of Consumer and Worker Protection since 2023-07-05.
**Attaches when:** an automated employment decision tool (AEDT) is used to substantially assist or replace discretionary decision-making for a job or promotion, for a position located in New York City. Employment agencies are covered alongside employers.
| Control | Parameter | Notes |
|---|---|---|
| Independent bias audit | Within the 12 months preceding use, renewed annually | Auditor must have no involvement in using, developing, or distributing the tool and no employment or financial relationship with the employer that compromises independence |
| Published audit summary | Publicly available on the careers or jobs section of the website | Must include the date of the most recent bias audit and the distribution date of the tool |
| Metrics published | Selection or scoring rates and impact ratios by sex, by race/ethnicity, and by intersectional sex × race/ethnicity categories | The four-fifths screen is the conventional read on the impact ratio |
| Candidate notice | At least **10 business days** before use | Business days, not calendar days |
| Notice contents | That an AEDT will be used; the job qualifications and characteristics it assesses | |
| Data disclosure on request | Within **30 days** of a written request, if not already published | Data collected, source of the data, retention policy |
| Penalty — first violation | Up to $500 | |
| Penalty — subsequent violations | $500 to $1,500 each | |
| Accrual | Each day an AEDT is used in violation is a separate violation; each failure to provide a required notice is a separate violation | This is what makes notice gaps on high-volume reqs the top remediation item |
**The load-bearing determination** is whether the tool substantially assists or replaces discretionary decision-making. It turns on your configuration, not the vendor's product description.
---
## Illinois Artificial Intelligence Video Interview Act (820 ILCS 42)
**Status:** in effect since 2020-01-01.
**Attaches when:** an employer asks applicants to record video interviews and uses AI analysis of those videos to consider applicants' fitness, for positions based in Illinois.
| Control | Parameter | Source |
|---|---|---|
| Notice before the interview | Applicant is told AI may be used to analyze the video and consider fitness | Section 5 |
| Explanation before the interview | How the AI works and the general types of characteristics it uses to evaluate applicants | Section 5 |
| Consent before the interview | Consent to be evaluated by the AI as described; no consent means no AI evaluation | Section 5 |
| Sharing limited | Videos shared only with persons whose expertise or technology is necessary to evaluate fitness | Section 10 |
| Destruction on request | Within **30 days** of the applicant's request, delete the interviews and instruct every other recipient to delete their copies, including all electronically generated backup copies | Destruction section |
| Demographic reporting | Applies only to employers that rely **solely** on AI analysis of the video to decide whether an applicant advances to an in-person interview. Report race and ethnicity of applicants afforded and not afforded in-person interviews, and of applicants hired | Reporting section |
| Reporting deadline | Annually by **December 31**, covering the 12-month period ending the preceding November 30, to the Department of Commerce and Economic Opportunity | DCEO reports to the Governor and General Assembly by July 1 on whether the data discloses racial bias |
**Ordering is the control.** Notice, explanation, and consent must all precede the interview. A post-interview acknowledgment does not cure it.
**The sole-reliance trigger** is narrow and most stacks fall outside it because a recruiter reviews before the in-person decision. Record the fact that puts you outside it; do not assume it.
---
## Illinois Human Rights Act, as amended by HB 3773
**Status:** statutory obligations in effect since 2026-01-01. Illinois Department of Human Rights rulemaking is still open — proposed amendments to Title 44, Part 2520 of the Illinois Administrative Code were published 2026-05-15, after an earlier draft was withdrawn.
| Control | Parameter |
|---|---|
| Discriminatory-effect prohibition | AI may not be used with the effect of subjecting employees or applicants to discrimination on a protected basis, in recruitment, hiring, promotion, renewal, selection for training or apprenticeship, discharge, discipline, tenure, or terms and conditions of employment. Intent is not required |
| Proxy prohibition | Using zip code as a proxy for a protected class is prohibited outright |
| Notice | Required whenever AI is used in a covered employment decision, regardless of whether the use has any discriminatory purpose or effect |
**Notice mechanics — timing, form, and content — are the subject of the open rulemaking.** Grade the notice-existence row now and put the mechanics row in `counsel-review` until the rules land.
---
## California — FEHA regulations on automated-decision systems
**Status:** in effect since 2025-10-01.
**Attaches when:** an employer uses artificial intelligence, machine learning, algorithms, statistics, or other data processing to facilitate human decision-making on recruitment, hiring, or promotion of applicants or employees in California.
| Control | Parameter |
|---|---|
| Records retention | **Four years** for automated-decision-system records, including selection criteria, relevant outputs, and audit findings |
| Third-party liability | The employer is answerable for discriminatory outcomes of a tool sourced from a vendor or run by an agent |
| Anti-bias testing | Not mandated. Evidence of testing may support a defense; the absence of it is admissible against the employer |
**The retention floor conflicts with privacy-minimization defaults and with deletion requests.** Surface the tension; do not resolve it in the report.
---
## Colorado SB 26-189 — automated decision-making technology
**Status:** signed 2026-05-14; effective **2027-01-01**. Attorney General rulemaking pending, and key terms will be defined there.
This replaced SB 24-205, the 2024 Colorado AI Act, which never took effect. SB 24-205's delayed start moved from 2026-02-01 to 2026-06-30 (SB 25B-004, signed 2025-08-28) and was then superseded. **The risk-management program, annual impact assessments, and broad algorithmic-discrimination duties of SB 24-205 are gone.** Any checklist still grading against them is auditing a repealed statute.
| Control | Parameter |
|---|---|
| Pre-use notice | Clear notice that a covered automated decision-making technology will be applied, before use |
| Post-adverse-outcome disclosure | Plain-language description of the technology's role, within **30 days** after a consequential decision producing an adverse outcome |
| Human review | The individual may request meaningful human review and reconsideration of the decision |
| Records | Retain relevant records at least **three years** |
Grade Colorado rows as forward-looking readiness until the effective date. Do not report a Colorado gap as accruing exposure today.
---
## Federal posture — context, not a control
Executive Order "Ensuring a National Policy Framework for Artificial Intelligence" (signed 2025-12-11) directs a Department of Justice AI Litigation Task Force, stood up from 2026-01-10, to challenge state AI laws in federal court, and directed Commerce to identify state laws suitable for challenge by March 2026.
**No state law in this matrix has been displaced by it.** Preemption of a state statute requires a court to say so or Congress to act. Treat the federal posture as a reason to keep controls documented and portable — not as a reason to retire any row above. Note it in the report's assumptions section so the reader knows it was considered.
---
## Not covered by this matrix
Add rows before relying on the skill for any of these: EU AI Act Annex III employment obligations; Maryland's facial-recognition consent requirement; Texas TRAIGA; New York State requirements distinct from the City's; sector rules for federal contractors; and any collective-bargaining commitments on automated evaluation.
# Tool and stage ledger
Fill this out before the first run. It is the input the classification step reasons over, and it is the artifact that makes a later audit reproducible. Replace every `REPLACE_` value.
## Part A — Requisitions in scope
Jurisdiction attaches by where the job is located and where the candidate applies from, not by where the company is headquartered. A remote-eligible req attaches every state in its accepted-candidate list.
| Req ID | Title | Work locations | Remote-eligible states accepted | Volume (applicants/mo) | Status |
|---|---|---|---|---|---|
| REPLACE_ENG-411 | Backend Engineer | New York, NY | US-wide except CO | 400 | open |
| REPLACE_SLS-102 | Account Executive | Chicago, IL | IL only | 120 | open |
| REPLACE_OPS-220 | RevOps Analyst | Remote US | CA, NY, IL, TX | 60 | planned |
If a req's accepted-state list is "anywhere in the US," write that out and expect the matrix to attach every jurisdiction in `1-jurisdiction-matrix.md`. That is usually the finding, not a formality.
## Part B — Tool ledger
One row per tool that touches a candidate between application and offer. Include tools you did not buy for screening but that score or rank anyway — ATS match scores and sourcing-tool fit scores are the ones teams forget.
| Tool | Stage | What it outputs | Who sees it and when | Does it gate a transition? | Threshold | Video/audio analysis? | Vendor's own label |
|---|---|---|---|---|---|---|---|
| REPLACE_HireVue | async video screen | competency scores 1-5 | recruiter, before advance/reject | no auto-advance; recruiter decides | none | yes | "decision support" |
| REPLACE_Greenhouse | application review | match score, sorts the list | recruiter, list is worked top-down | no | none | no | "ranking aid" |
| REPLACE_vendor | resume screen | pass/fail | nobody; auto-rejects | yes | score under 60 auto-rejects | no | "efficiency filter" |
### The four classification questions
Answer from configuration. The vendor's label goes in its own column and is never the answer.
1. **Does the output gate a stage transition automatically, at any threshold?** An auto-reject threshold is the clearest case.
2. **Does it sort or rank a list a recruiter works top-down?** Rank order changes who gets reviewed at all when volume exceeds review capacity. Record the review-capacity number — "400 applicants, recruiter reviews the top 40" is the fact that matters.
3. **Does a human see the score before making the call?** A score presented before the decision is a different fact from a score available afterward on request.
4. **Does it analyze video or audio for characteristics used to evaluate fitness?** Independent of the other three, and it is what pulls the Illinois AI Video Interview Act in for Illinois-based positions.
### Configuration facts to capture per tool
- Where the score appears in the recruiter's interface, and whether it can be hidden.
- Whether the score is recorded on the candidate record and for how long.
- Whether the vendor retrains on your candidate data, and whether that is contractually disclaimed.
- Which sub-processors receive candidate video or audio. This is the list the Illinois destruction obligation reaches.
- Who at the vendor can access candidate video, and under what contractual limit.
## Part C — Review-capacity reality
Fill this in honestly; it decides whether a ranking tool is functionally a filter.
| Req | Applicants/mo | Applications actually reviewed by a human | Effective filter rate |
|---|---|---|---|
| REPLACE_ENG-411 | 400 | 40 | 90% never human-reviewed |
A ranking tool with a 90% effective filter rate is doing decision-making work regardless of what the configuration screen calls it. Record this and let the classification step use it.
# Evidence pack index
Every control graded `evidenced` needs a citation from this pack: a file path or URL plus the passage that does the work. A control with no artifact is `unevidenced`, and `unevidenced` sorts into remediation next to real gaps — because the difference between "we do this" and "we can show we do this" only matters on the day someone asks, and on that day there is no time to build the artifact.
## Part A — Artifact inventory
Fill in the path or URL. Leave `MISSING` where it is missing; that is the point of the exercise.
| # | Artifact | Satisfies | Side | Location |
|---|---|---|---|---|
| 1 | Published AEDT bias-audit summary page | NYC LL 144 publication + metrics | employer | REPLACE_URL |
| 2 | Independent auditor's report and engagement letter | NYC LL 144 audit + independence | employer | REPLACE_PATH |
| 3 | Candidate AEDT notice text, with the send trigger documented | NYC LL 144 notice + contents | employer | REPLACE_PATH |
| 4 | Timestamp export: notice sent vs tool first run, per candidate | NYC LL 144 10-business-day lead | employer | REPLACE_PATH |
| 5 | Illinois video notice + explanation text | 820 ILCS 42 Section 5 | employer | REPLACE_PATH |
| 6 | Consent capture records with timestamps | 820 ILCS 42 Section 5 ordering | employer | REPLACE_PATH |
| 7 | Sub-processor list for video and audio | 820 ILCS 42 Section 10 sharing limit | employer | REPLACE_PATH |
| 8 | Deletion runbook naming downstream recipients and backups | 820 ILCS 42 destruction, 30 days | employer | REPLACE_PATH |
| 9 | Sole-reliance determination memo | 820 ILCS 42 reporting trigger | employer | REPLACE_PATH |
| 10 | Illinois HRA AI-use notice text | IL HRA as amended | employer | REPLACE_PATH |
| 11 | ADS records retention schedule showing four years | CA FEHA ADS | employer | REPLACE_PATH |
| 12 | Anti-bias testing results, if any were run | CA FEHA ADS defense | employer | REPLACE_PATH |
| 13 | Vendor DPA with retraining and access terms | supports 7, 8, 11 | vendor | REPLACE_PATH |
| 14 | Vendor's own bias-audit summary | vendor-side rows only | vendor | REPLACE_URL |
| 15 | Data-retention policy published on the careers site | NYC LL 144 disclosure on request | employer | REPLACE_URL |
**Row 14 cannot satisfy rows 1 or 2.** A vendor's audit of its model is not the employer's audit of the employer's use. The NYC duty sits with the employer or employment agency. Vendors sell attestation packets that read as if they close the employer's obligation; they close the vendor's.
## Part B — Notice scaffolding
Adapt; do not paste. The bracketed values are the ones that make a notice specific enough to be worth anything, and a notice that omits them is the most common `gap` finding.
### NYC AEDT notice — at least 10 business days before use
> We use an automated employment decision tool to help evaluate applications for [JOB TITLE]. The tool assesses the following job qualifications and characteristics: [LIST THEM — the actual assessed dimensions, not "fit"]. The results of our most recent bias audit and our data-retention policy are published at [URL]. You may request an alternative selection process or an accommodation by contacting [ADDRESS]. To request information about the data we collect for this tool, its source, and our retention policy, contact [ADDRESS]; we will respond within 30 days.
Send it on the application-confirmation event, not on the assessment invite. Invite-time sending is what fails the 10-business-day count on fast-moving reqs, and it is the single most common lead-time failure.
### Illinois video interview — notice, explanation, and consent, all before the interview
> This interview will be recorded and may be analyzed by artificial intelligence to consider your fitness for [JOB TITLE]. How it works: [PLAIN-LANGUAGE EXPLANATION — what the system does with the recording]. The general types of characteristics it uses to evaluate applicants are: [LIST THEM]. Your video will be shared only with people whose expertise or technology is necessary to evaluate your fitness for this position. You may request that we delete your interview at any time by contacting [ADDRESS]; we will delete it, and instruct everyone who received a copy to delete theirs including backups, within 30 days.
>
> [ ] I consent to being evaluated by artificial intelligence as described above.
The consent checkbox must be reachable and actionable before the recording starts. If a candidate can begin recording without passing the consent control, that is a `counsel-review` at best.
### Illinois HRA AI-use notice
Notice mechanics are in open rulemaking as of the `checked:` date in `1-jurisdiction-matrix.md`. Give notice that AI is used in the decision, keep the text versioned, and grade the mechanics row as `counsel-review` until the rules land.
## Part C — Evidence hygiene
- **Version the notice text and keep the diffs.** The question is never "what does the notice say," it is "what did it say on the day this candidate applied."
- **Timestamps beat policies.** A policy saying notice goes out 10 days ahead is weaker evidence than an export showing it did.
- **Keep the negative determinations.** The memo explaining why a tool is out of scope is an artifact. An undocumented determination reads later as an oversight.
- **Watch the retention conflict.** The four-year California floor and a candidate's Illinois deletion request point opposite ways on the same record. Decide it with counsel once, write the decision down, and cite that memo rather than re-deciding per request.
