← Back

Guide · Published November 14, 2026 · Reviewed November 14, 2026 · Ingenious AI team · Not legal advice

← All guides

How to Tell If a Tool Is ADMT: The Role Test in Colorado, California, and Connecticut

The short answer: Whether duties attach depends less on what a tool is called than on what it does in the decision. The kit's checker asks the role question directly, with three answers: the AI makes the decision on its own, the AI is a substantial factor, or the AI is a minor input and a person genuinely decides. Colorado's test is material influence on a consequential decision; California's is technology that replaces or substantially replaces human decision-making, with a defined human-decision carve-out; Connecticut's employment notices trigger on direct interaction or the tool being a substantial factor. Incidental use fails every test — and token human review passes none of them.

Check your duties and dates — free Build your first notice — free

What is ADMT? Automated decision-making technology: technology that makes, or materially shapes, decisions about people. The free duties checker operationalises the definition per jurisdiction — answer its three questions (jurisdiction, decision type, the tool's role) and it returns the duties or explains why none attach. This guide explains the reasoning behind the role question so you can answer it accurately.

The minor-input end: what genuinely incidental use looks like

At the minor end, the checker is consistent across Colorado and California: if a person genuinely decides and the tool is one minor input, the ADMT duties do not arise. Colorado's phrasing: the tool is covered only when it materially influences a consequential decision — more than an incidental input. California's phrasing is a positive test for human decision-making, covered below.

Incidental use has recognisable features: the output is one document among several the decider reads; the decider regularly reaches outcomes the tool did not suggest; the decision would proceed, substantively unchanged, if the tool were unavailable for a week. If removing the tool would halt or transform the process, the use is not incidental, whatever the workflow diagram says. Record the reasoning either way — a documented minor-input finding, with these features checked, is a defensible register entry.

The substantial middle: where most real tools live

Most operational tools sit in the middle: a score, ranking, or recommendation that a person formally approves. The tests catch this zone deliberately. Colorado: a tool that is ranking people, narrowing their options, or setting terms has crossed into material influence. Connecticut: the written notice is owed before the tool is a substantial factor. California: technology that substantially replaces human decision-making is ADMT, and the human carve-out fails where the review is token.

The practical probe for the middle zone is the override rate and its meaning. If reviewers approve the tool's recommendation in nearly every case, ask why: genuine agreement with a good tool, or a workflow where disagreeing costs time, justification, and career safety? The first can still be substantial-factor use (the duty attaches to the tool's role, not to reviewer laziness); the second is token review wearing a rota. Either way, the middle zone is covered in all three states — plan notices accordingly.

Worked example: Pioneer Valley Credit Services

Example — fictional business, real duty structure. Pioneer Valley is a fictional lender and employer that applied the role protocol to the five systems in its register. The findings, per tool:

California's human-decision carve-out, stated exactly

California's checker states the carve-out as a three-part test: the decision-making is human, not automated, if a human reviewer knows how to interpret the output, reviews the output and other relevant information, and has the authority to make or change the decision. All three must hold. A reviewer who can change the decision but cannot interpret the score fails part one; an expert reviewer without authority fails part three.

Token review is named explicitly as failing: a person rubber-stamping the tool's output does not meet the test. The carve-out is therefore an operating standard for the review step, not a label for the org chart — it describes what the reviewer does with each file. If you rely on it, evidence it per decision: the record should show review happened, by someone qualified, with authority. The full discussion is in Token Human Review and Meaningful Human Review.

Connecticut's two triggers: interaction and substantial factor

Connecticut employment adds a trigger the other states do not use: direct interaction. The interaction notice is owed when a person will directly interact with the tool, regardless of how the eventual decision weighs its output — an applicant completing an automated assessment meets the tool, and the notice follows the meeting. The second trigger, the tool as a substantial factor, carries the fuller written notice.

For Connecticut non-employment decisions, the profiling frame applies instead: profiling in furtherance of an automated decision with legal or similarly significant effects — and since July 1, 2026, without the word 'solely', so a nominal human in the loop does not end the analysis. Connecticut therefore has no comfortable 'a person signs it' exit in either track.

Applying the test: a five-minute protocol per tool

For each registered tool, run the protocol: (1) What does the output do — decide, rank, narrow, set terms, or merely inform? (2) What happens when the tool is unavailable for a week? (3) What is the measured override rate, and what does disagreeing cost the reviewer? (4) Does the reviewer meet California's three parts — interpretation, full review, authority? (5) In Connecticut employment, does a person directly interact with it? Five answers, recorded per tool per state, turn the scope question from an argument into a file.

Then verify against the checker: run each combination through the free duties checker and file the result with the register row. Where the answer is covered, the same row should name the notice workflow that follows — the inventory method is in How to Inventory Automated Decision Tools, and the boundary cases get a full guide in When Does an ADMT Duty Not Apply?

  1. For each tool, answer the five protocol questions in writing — output's effect, outage test, override rate, reviewer quality, interaction trigger
  2. Apply each state's test in its own words: material influence (CO), replaces or substantially replaces (CA), substantial factor or direct interaction (CT)
  3. Test the human-decision claim against California's three parts — interpret, review fully, authority to change — and treat token review as failing
  4. Record negative findings with reasoning and dates, not just the covered tools
  5. Verify every combination in the free checker and file the result with the register row

Sources, as carried in the kit's own data: the duty summaries in app/rules.py and the checker data in app/checker_data.py, verified against primary or authoritative sources on October 28, 2026. Run the free duties checker for your jurisdiction, decision type, and the tool's role, and compare the kit's plans for the record, builder, and review queue. More from Ingenious AI is on the hub.

Build your first notice — free

Related guides

See all ADMT guides for the full set.

Frequently asked questions

What is the simplest way to tell if a tool is ADMT?

Ask what its output does to the decision. If it decides, ranks, narrows options, or sets terms, it is materially influencing the decision (Colorado) or a substantial factor (Connecticut). If it replaces or substantially replaces human decision-making, it is ADMT in California.

A person approves every decision — does that take us out of scope?

Not by itself. California's carve-out requires the reviewer to know how to interpret the output, review it with other relevant information, and have authority to change the decision — and token review expressly fails. Connecticut's profiling rights, since July 1, 2026, do not turn on the word 'solely' either.

What if the tool is only a minor input?

Then the duties may not attach: Colorado covers only material influence, and the checker returns 'no duties found' for genuinely incidental use in Colorado and California. Record the reasoning — removal test, override behaviour, decision continuity — with the finding.

Does the EU ask the same role question?

No. The kit's checker notes Article 50 transparency duties attach to interaction and content regardless of the AI's role in a decision, so the EU result is the same for every role answer.

How do I verify my assessment?

Run the combination — jurisdiction, decision type, role — through the free duties checker and file its result with the tool's register row.

Build your first notice — free