Skip to main content

NYC Local Law 144: Automated Employment Decision Tool Governance

New York City's Local Law 144 is one of the clearest examples of why AI governance has to be route-specific. The law applies to automated employment decision tools used to substantially assist or replace discretionary decision-making in hiring or promotion. It requires an annual bias audit before use, public disclosure of the audit summary, and notice to candidates or employees. That means the governance question is not simply whether the employer uses AI. It is whether a specific workflow materially influences employment decisions and whether the organization can prove the right controls around it.

Keeptrusts helps on the runtime side of that problem. It can segregate the hiring lane from other HR and productivity routes, escalate certain HR-oriented outputs for human review, and preserve evidence around how the route behaved. It does not perform the legally required bias audit by itself, and it does not replace candidate notice or published summaries. That limitation is important. Honest governance is better than overclaiming tooling.

Use this page when​

  • You use AI in hiring or promotion workflows that may fall within Local Law 144.
  • You need stronger runtime controls around HR-related outputs and reviewer intervention.
  • You want to separate employment-decision tooling from general HR or recruiting assistance.

Primary audience​

  • Primary: HR technology leaders, employment counsel, compliance and platform teams
  • Secondary: recruiters, talent-operations leaders, internal audit

The problem​

Employment AI is risky because a workflow can move from helpful to decision-influencing very quickly. A recruiting assistant that starts by summarizing resumes may soon rank candidates, recommend interview questions, or generate promotion narratives. Once the route materially assists or replaces discretionary judgment, Local Law 144 concerns become much more concrete.

The biggest implementation mistake is to bundle all HR AI together. Teams call everything a recruiting assistant, but the actual use cases are different. Interview scheduling is not the same as candidate scoring. Writing a job description is not the same as a promotion recommendation. If those workflows share one route and one approval posture, the organization has no defensible boundary for where the law becomes relevant.

The second mistake is to confuse internal monitoring with the required legal bias audit. Even if the platform can flag risky output patterns, the law still requires an annual bias audit and related notice obligations. Runtime controls are essential, but they are not the whole compliance story.

The solution​

Keeptrusts supports a more defensible Local Law 144 operating model by treating employment-decision routes as their own lane. Put candidate or promotion recommendation flows behind a stricter chain. Use bias-monitor to escalate narrow HR-style patterns that combine employment context with protected-characteristic language. Use human-oversight so the route does not deliver decision-sensitive output directly. Preserve evidence with audit-logger and export workflows so reviewers can inspect how the tool behaved over time.

This is valuable precisely because the current bias-monitor is narrow. It does not pretend to solve every fairness question. It creates a runtime tripwire for one class of HR-sensitive content, which is a realistic way to support reviewer judgment without claiming the gateway performed the legal audit for you.

Implementation​

For an employment-decision lane that should always stop for review, use a focused HR control stack.

pack:
  name: nyc-ll144-hiring-lane
  version: "1.0.0"
  enabled: true

providers:
  targets:
    - id: hr-reviewed-provider
      provider: openai
      model: gpt-5.4-mini-mini
      secret_key_ref:
        env: OPENAI_API_KEY
      data_policy:
        zero_data_retention: true
        training_opt_out: true
        retention_days: 0
        allow_internet_egress: false

policies:
  chain:
    - pii-detector
    - bias-monitor
    - human-oversight
    - audit-logger

policy:
  pii-detector:
    action: redact
    redaction:
      marker_format: label
      include_metadata: true

  bias-monitor:
    threshold: 0.85

  human-oversight:
    action: escalate

  audit-logger:
    retention_days: 365

This route is designed for candidate evaluation summaries, promotion support, and employment-decision narratives where direct model output should never become final action. It is not necessary for job-description drafting or scheduling workflows, and that separation matters. Over-scoping creates friction. Under-scoping creates legal exposure.

The most useful docs for this rollout are Bias Monitor, Human Oversight, Reviewing Alerts and Evidence, Export Evidence for a Review, Managing Policy Changes, and Pass Compliance Audits.

Results and impact​

Organizations that separate the hiring lane usually gain clarity fast. Recruiters and HR operations teams know which workflows are ordinary productivity tools and which ones trigger a more formal review process. Legal and compliance teams gain evidence that the route does not silently deliver unreviewed content. Internal audit gets a cleaner way to compare runtime behavior with the bias-audit and notice program.

That still leaves work outside the gateway. The annual bias audit, the required notices, and the public audit summary remain organizational duties. But the runtime control layer makes those obligations easier to support and easier to defend.

Key takeaways​

  • Local Law 144 is route-specific because employment-decision use is route-specific.
  • Keeptrusts does not replace the required annual bias audit or notice obligations.
  • bias-monitor is useful as a narrow HR escalation tripwire, not a full fairness certification engine.
  • human-oversight is appropriate for candidate and promotion outputs that should never flow straight through.
  • The most important governance move is separating hiring decision routes from general recruiting productivity tools.

Next steps​