Colorado ADMT: Build Decision Evidence Before January 2027

An employment-screening tool uses the wrong employment end date from a reference letter. When an applicant challenges the result, can the team retrieve the exact document, input values, decision, and notice, then explain what changed after correction? That is a document-evidence problem as well as an AI-governance problem.

Colorado's revised Automated Decision-Making Technology Act takes effect January 1, 2027. It addresses covered technology that materially influences consequential decisions, with requirements for developers and deployers and consumer rights concerning data and review. Confirm the workflow's scope, the organisation's role, and applicable exemptions before mapping obligations to a system.

As of November 11, the Attorney General's rulemaking page lists an October 6 interim draft and an October 26 hearing. Treat draft procedures separately from enacted requirements and check adopted rules before implementation. The architecture below is a practical evidence design for an in-scope workflow.

Follow the evidence through three stages

1. 1 · Capture

   Preserve the inputs

   Link the case, source document versions, input values, and system version.

2. 2 · Decide

   Record the outcome

   Retain the output, applicable rules, review work, final result, and notice.

3. 3 · Correct

   Connect later changes

   Link a challenge, validated corrections, reconsideration, and any revised outcome.

Connect each decision to its evidence

A system inventory identifies owners, versions, intended uses, and risk classifications. It does not explain an individual decision. Create a decision record that links the person or case, decision purpose, applicable system and version, input snapshot, output, business rules, human reviewer, final result, and the notice provided.

In the screening example, retain the reference letter's version and the employment dates supplied to the decision tool. A link that always opens the newest letter cannot establish what an earlier decision used. Store durable version references rather than uncontrolled copies when another system remains authoritative.

Make corrections traceable

The law includes rights to request personal data and correct factually incorrect personal data used by covered technology. Following an adverse consequential decision, consumers can also request meaningful human review and reconsideration. A correction workflow should identify the challenged field, its source, the submitted evidence, validation work, the accepted value, and whether reconsideration or a new decision is needed.

For the disputed employment date, preserve the original value, the corrected letter, and the reviewer's reason for accepting or rejecting the correction. Keep the earlier decision explainable while recording which later decisions use the validated value.

A corrected decision should link back to the earlier one. Record what changed, who authorised the correction, which system version processed it, and what notice was issued. If the challenge is rejected, preserve the reason and the evidence considered, subject to applicable retention and privacy rules.

Separate system evidence from case evidence

These sets have different access patterns. Engineers may need aggregated monitoring data without viewing every consumer document. A case reviewer may need the decision package without access to unrelated model-development material. Use document-level controls and record every access that the process relies on.

Use integration events to preserve context

Have the decision application send the case ID, decision ID, input values, system version, and source document references when work completes. Send later events for a correction, review, or changed outcome. Deduplicate retries and reconcile missing events so a temporary integration failure does not leave a decision without its evidence.

Avoid treating an audit log as the entire evidence package. Logs can show that an API call occurred, while the linked record explains what was decided and which version of each document was considered.

A readiness drill for December

• Select one consequential-decision workflow and reconstruct a completed case.
• Correct one material input without destroying the original decision context.
• Verify that notices and reviewer decisions reference the exact evidence used.
• Verify that consumer disclosures contain the appropriate case evidence without exposing internal or unrelated records.
• Test access for the case reviewer, compliance team, and system administrator.
• Confirm retention and deletion rules for both case evidence and system documentation.

Make each decision and correction explainable

A useful evidence record connects the source information to the decision, notice, review, and any later correction. Test that chain on one real workflow before expanding it, so the organisation can explain both the original outcome and what happened when the facts changed.

Regulatory update reviewed: November 11, 2026. Confirm final Colorado rules and sector-specific requirements before implementation.