Colorado's SB24-205 made the state the first in the US to set cross-sector rules for high-risk artificial intelligence. After a 2025 special session the effective date moved to 30 June 2026. The law may be amended again, so treat the current statute as the authority, not this summary.
The Act centres on a single idea: the consequential decision. That is a decision with a material effect on someone's access to, or the cost and terms of, things like employment, housing, lending, insurance, healthcare, education, legal services and essential government services. When an AI system is a substantial factor in such a decision, both the developer and the deployer owe reasonable care against algorithmic discrimination.
What that asks of a deployer
In practice the obligations look like an operating discipline:
- A risk management policy for the high-risk systems in use.
- An impact assessment for each system, repeated when it changes.
- Notice to people when an AI system is making, or substantially helping make, a consequential decision about them.
- For an adverse decision, an explanation of the principal reasons and a path to correct data or appeal, with human review where feasible.
Design for the appeal, not the demo
The last point is the one most systems fail. A model that produces a score cannot explain an adverse outcome unless the decision around it was designed to be explained. That means recording which inputs were used, which rule or model version applied, where a person reviewed it, and what the person was shown.
This is the seam between the model and the decision-maker, and it is where Opsian spends most of its engineering time. Agents act inside limits someone set. Every tool call is logged. Topics that must go to a person always go to a person. A system built this way can answer the question a regulator, a court or a customer will eventually ask: why did this happen to me?
