A Colorado AI act impact assessment for a high-risk AI agent sets out its purpose and context, the risks of algorithmic discrimination and how they are mitigated, the data it takes in and produces, how its performance is measured, how consumers are informed and how it is monitored. Each section must describe the agent's steps, not just its output.
What does the impact assessment require, in plain words?
Colorado has passed an AI act aimed at high-risk systems that make consequential decisions. Deployers of those systems must complete an impact assessment before use, repeat it regularly, and update it after any intentional and substantial modification. In substance, it must contain:
- a statement of the system's purpose, intended use cases, deployment context and benefits;
- an analysis of whether it poses known or reasonably foreseeable risks of algorithmic discrimination and, if so, their nature and the steps taken to mitigate them;
- the categories of data it processes as inputs and the outputs it produces;
- an overview of any data used to customise it;
- the metrics used to evaluate its performance, and its known limitations;
- the transparency measures taken, including notices to consumers;
- the post-deployment monitoring and user safeguards, including oversight;
- after a substantial modification, how far the system's use matched or departed from the developer's intended uses.
Records of the assessments must be kept for a period the act sets. One assessment can cover comparable systems, and an assessment made under another law can be reused if it is reasonably similar in scope. For context on the act, see Colorado AI act and AI agents making consequential decisions.
Who has to produce one?
Deployers of high-risk systems, subject to the act's relief for certain small deployers. A third party can carry out the assessment on the deployer's behalf, but the deployer is responsible for it. The developer's documentation is the main input on how the system works, its training data and its known risks. The wider deployer duties are set out in Colorado AI act duties for deployers.
What does an agent change in the assessment?
Most assessment templates assume a model with fixed inputs and a single output. An agent has neither. It decides what to read, calls tools, may ask the consumer questions and produces a recommendation or an action. Each section of the assessment needs to reflect that.
Purpose and context. Describe each capability that touches the decision, and its depth: reading, drafting for a person, or acting within limits. An agent that drafts a lending recommendation at depth 2 has a different risk profile from one that approves small loans at depth 3.
Data inputs. List every source the agent can reach, not only the fields a model receives. Note where each comes from: declared by the consumer, signed by a third party such as an employer, or sealed by a system of record. Discrimination can enter through which sources the agent trusts.
Discrimination risk. Consider the agent's conduct as well as its output. Does it ask some applicants more follow-up questions? Does it reject documents in formats that some groups use more often? Does its language suit consumers with limited English?
Substantial modification. Define it for agents in advance. A new model, a new source, a new capability or a higher depth should each count.
How do you build the assessment for an agent?
| Requirement | Control | Evidence it leaves |
|---|---|---|
| Purpose, use cases, context, benefits | Capability list with depth and the decision each one affects | Capability register extract |
| Discrimination risk and mitigation | Outcome and conduct tests by group on the deployer's own past cases | Test results and mitigation actions |
| Data inputs and outputs | Source list with provenance class; output types per capability | Source register and output catalogue |
| Customisation data | Description of briefs, instructions and examples used to configure the agent | Configuration record |
| Metrics and limitations | Pass floors for grounding, refusal, escalation and harm; known failure modes | Certification results and limitation notes |
| Transparency measures | Notice text and adverse decision reason format | Approved notices |
| Monitoring and safeguards | Outcome monitoring by group; human oversight design; suspension triggers | Monitoring reports and oversight records |
| Modification statement | Change log comparing actual use with intended use | Change log and updated assessment |
If you already run a fundamental rights impact assessment for EU purposes, the structure overlaps considerably. Our fundamental rights impact assessment template for AI agents can serve as a starting point, and do AI agents need a fundamental rights impact assessment explains when that EU duty applies.
What will an enforcement authority ask?
- Show us the impact assessment for this agent, and the earlier versions.
- Which groups did you test outcomes for, and on what data?
- What sources does the agent use, and could any of them disadvantage a protected group?
- What did you define as a substantial modification, and has one occurred?
- What has post-deployment monitoring found since the last assessment?
What does the impact assessment not cover?
It does not certify the agent as non-discriminatory; it records the analysis and mitigation. It does not replace the separate duties to notify consumers, explain adverse decisions and offer appeals, though it should describe them. It does not cover non-high-risk agents. And it is not a privacy assessment, although Colorado's privacy law may require a data protection assessment for some profiling, which can be combined. For the full map, see AI agent regulation mapped to controls.
This is an explanation of the rules, not legal advice; confirm how they apply to you with counsel.
Questions people ask
How often must the assessment be repeated?
Regularly, at the interval the act sets, and again after any intentional and substantial modification. For agents, changes come often, so link the assessment to your change process.
Can one assessment cover several agents?
The act allows one assessment for comparable systems. Agents with different capabilities, sources or decisions are rarely comparable enough, so expect one per agent or per capability set.
Who should sign off the assessment?
The business owner of the decision, with input from compliance, risk and data teams. The sign-off should be by someone who can stop the agent if the assessment finds unmitigated risk.