Automated Decision-Making: Systems, Law, and Verdicts

Automated decision-making (ADM) is any process where software makes a decision about a person or a case without a human making the final call. Loan pre-approvals, spam filters, fraud holds, resume screens, and insurance quotes are all automated decision-making systems, and so is a ticket router, just with lower stakes.

The stakes are what matter. Regulators in several jurisdictions now treat automated decisions with "legal or similarly significant effects" differently from routing a support email, and builders need to know which side of that line their pipeline sits on. This page is an orientation, not legal advice.

How automated decision-making systems work

Every ADM system has the same four parts:

1. Inputs: data about the person or case.
2. Logic: rules, a trained model, or a judgment model that turns inputs into a decision.
3. Output: approve, reject, flag, route, score.
4. Action: something happens because of the output.

The design question is where a person sits. Fully automated means the output triggers the action directly. Human-in-the-loop means a person confirms before anything happens. Human-on-the-loop means a person monitors and can override afterward.

Where decision models fit

A decision model like Jev, from TypeSafe AI, is purpose-built for the logic step: closed questions in, a probability per answer out (what a decision model is). That shape makes some ADM obligations easier to meet, not harder:

• Every decision has a question. The wording is the logic, readable by an auditor.
• Every decision has a probability, so uncertain cases can route to a person by rule.
• Every decision can be logged with its evidence and question version, the spec in verdict logging and audit trails.

What the model can't do is make a high-stakes decision legitimate on its own. That part is process design.

The legal landscape, briefly

A few regimes matter most to builders, stated at the level of general orientation:

• GDPR Article 22 (EU and UK) gives people rights around decisions based solely on automated processing that produce legal or similarly significant effects, including, in many cases, the right to human intervention.
• California's CCPA regulations now address automated decision-making technology, covered on its own page.
• Hiring-specific rules, such as New York City's law on automated employment decision tools, require bias audits and notices for certain tools. Resume screening covers the hiring posture this site recommends.
• The EU AI Act classifies some uses, including many in employment, credit, and education, as high-risk, with added obligations.

Rules differ by jurisdiction and change often. Confirm the current text with counsel before relying on any summary, this one included.

Designing ADM you can defend

Four habits hold up under most regimes:

• Keep a human decision on anything significant. The model recommends; a person decides. On this site, irreversible actions never trigger on a lone verdict.
• Log the why. Inputs used, question version, probabilities, and who confirmed.
• Offer a path to contest. People affected by a decision need a way to ask for review.
• Measure disparities. Check outcomes across groups on your own data before launch and on a schedule after.

Low-stakes routing (which queue, which folder) rarely needs any of this, and treating it as if it does wastes effort. The skill is telling the two apart, which the decision model versus LLM comparison helps frame on the technical side.

Frequently asked questions

What is automated decision-making?

A process where software makes a decision about a person or case without a human making the final call.

Is automated decision-making legal?

Generally yes, with conditions that vary by jurisdiction and stakes. Decisions with significant effects on people face the most rules. This is not legal advice.

What counts as a "significant" automated decision?

Typically decisions affecting legal rights, employment, credit, housing, education, or access to essential services. Check the specific law that applies to you.

Does a human reviewer make a system not automated?

Only if the review is meaningful. A person rubber-stamping every output may not count as real human involvement under some regimes.

How do decision models help with ADM compliance?

They make the logic readable and every decision loggable with a probability, which supports audits. They don't replace legal review or human decision-making.