Back to playbooks

Human Command for Adverse Decisions

Use this as the baseline rule set for AI-assisted denial, closure, disenrollment, or other adverse-decision systems.

Use it when a team wants speed or consistency without making adverse action cheap and correction hard.

What problem this solves

Many harmful systems do not fail because they produce no explanation at all. They fail because the explanation is too generic, the record is too hard to reach, and the review arrives too late to matter.

Where a denial can cut off care, income, coverage, credit, or service, that is a human-command problem, not just an administrative one.

Failure pattern to prevent

The most common bad pattern is:

flag or score -> generic denial notice -> claimant cannot tell what actually happened -> appeal burden shifts downward -> wrong decision hardens

That is fake human command.

Non-negotiables

1. Do not let AI contribute to a materially consequential adverse decision without notice, reason, appeal, records, and human override.
2. Do not issue generic notices that hide the actual factors behind the decision.
3. Do not let the claimant bear the burden of reconstructing the vendor or model chain.
4. Do not let one automated signal trigger a high-stakes adverse action by itself where corroboration is required.
5. Do not allow decision ownership to dissolve across payer, agency, vendor, and model provider.

Minimum floor

If AI contributes to an adverse decision, the minimum floor is:

• notice
• reason
• appeal
• records
• human override

What has to be true in practice

• define which decisions are serious enough to count as materially consequential
• assign one named institution to own the outcome
• require specific reasons in notices and letters
• maintain the actual output-to-decision record
• provide a time-bounded appeal path that can still reverse the harm

Metrics and tripwires

Track:

• denial volume
• generic-notice rate
• records access response time
• appeal volume and overturn rate
• share of cases with named owner accountability

Tripwires:

• notices are too generic to challenge
• claimants cannot access the decision record promptly
• appeals arrive after the practical harm is already locked in
• staff cannot identify who owns the outcome

Who has to own this

This only works if ownership is clear:

• program or operational leadership for decision thresholds
• legal / compliance for rights floor and notice quality
• frontline review teams for timely reversal and override
• procurement / IT for traceability, logging, and audit access

Bridge language

“No efficiency gain counts if it comes from making denial cheap and correction hard.”

“If the institution acted on the output, the institution owns the outcome.”

Back to playbooks