Colorado’s Draft AI Rules: When Meeting Notes Influence Employment Decisions
Colorado's October 2026 draft AI rules put the focus on how employers use meeting outputs, with important exclusions for administrative summaries.
A manager preparing for a promotion discussion may find a searchable meeting archive useful for remembering projects, commitments, and decisions. The harder question arrives when the manager asks the software to identify who demonstrated leadership, rank the candidates, or explain which employee is least dependable. The archive has acquired a new purpose, and a sentence generated from incomplete meeting records may now influence someone’s career.
Colorado’s current AI rulemaking makes that change of purpose especially relevant. The attorney general released an interim draft on October 6, 2026, with a hearing and written-comment deadline scheduled for October 26. The substantive requirements being implemented are due to apply from January 1, 2027. These are different milestones, and the October document is still a proposal. The official rulemaking page is the place to check subsequent revisions.
For meeting software, the central issue is the function of the output and its influence on a consequential decision. An employer should be able to explain that function before treating either a recording tool or an entire meeting archive as covered or exempt. This article reviews the position as of October 11, 2026, and offers operational analysis rather than advice on an individual employment decision.
Start with the law Colorado actually enacted
The familiar summaries of Colorado’s 2024 AI law are an unreliable starting point for this review. SB26-189, signed in May 2026, repealed and reenacted the earlier provisions, replacing their framework with one focused on automated decision-making technology, or ADMT, in consequential decisions. Updating an implementation calendar without updating the underlying legal analysis would leave a substantial gap.
The enacted framework addresses developers and deployers separately. Its coverage reaches employment decisions, while its definitions and exclusions distinguish substantive decision assistance from routine administrative processing. Employers should assess their own role, the people and opportunities involved, and the particular deployment; a vendor’s general description of its product will rarely answer all those questions.
The timing also deserves care. The legislature’s bill page lists the act’s enactment history, while the attorney general explains the January 2027 start of the new substantive provisions. Neither the May signature nor publication of an October draft means that every proposed operational detail is already binding.
A company with employees in several states will need separate analyses. Our California workplace-AI article discusses distinct California statutes and their implementation dates. Colorado’s material-influence test should not be replaced with a California test simply because both reviews concern the same meeting archive.
Administrative summaries have meaningful exclusions
The enacted definitions in section 6-1-1701 exclude tools used solely for specified administrative functions, including summarizing and organizing information for human review. A further exclusion addresses summarization or presentation that produces no score, ranking, recommendation, classification, prediction, or other inference materially influencing an outcome. Covered ADMT must materially influence a consequential decision; incidental, trivial, and clerical uses do not qualify.
Those distinctions matter in ordinary work. A recap that records the agreed launch date, lists unresolved questions, and helps someone prepare the next agenda has a different function from an assessment that calls one participant unsuitable for promotion. The software may have produced both documents from the same transcript, with the same model, on the same laptop. Reviewing only the installation would miss the change.
Consider a hypothetical performance review in which a manager retrieves meeting passages about a delayed project. If the tool merely locates and presents the passages, and the manager evaluates the work, the exclusions deserve serious consideration. If the tool selects those passages because it has inferred that the employee is unreliable, then the selection itself becomes part of the assessment. Calling the result a summary does little to explain what happened.
This is why an inventory should preserve representative prompts and outputs alongside feature names. “Search,” “recap,” and “coaching” are convenient labels, but an actual example can show whether the system retrieved a statement, condensed a discussion, or supplied a judgment about a person. That evidence helps counsel assess coverage and helps managers understand the boundaries of an approved use.
The draft looks beyond the final click
Rule 6 of the October 6 interim draft proposes a rebuttable presumption of material influence for certain individual-related outputs reviewed during a decision process, or used to screen what the decision-maker sees, when the outcome is consistent with the output. The proposal also describes evidence of independent judgment and reiterates the administrative exclusions. These provisions remain draft rules, subject to change.
The practical concern is easy to recognize in a hypothetical promotion workflow. A manager could read an AI recommendation, open a few supporting notes, and then approve the recommendation manually. The presence of that final approval says little about the recommendation’s actual influence. A better account would explain what the manager considered, which evidence supported the choice, and whether the manager had a realistic opportunity to reach a different conclusion.
The screening stage deserves equal attention. Suppose software presents only meetings that it classifies as showing poor collaboration, leaving out successful work and the employee’s explanation of a dispute. The person reviewing the resulting packet might exercise sincere judgment while seeing a systematically narrowed record. A review of the displayed paragraph alone would miss how the packet was assembled.
For an employer preparing now, a useful exercise is to walk one proposed workflow backward from the decision. Identify the output the manager sees, the selection process behind it, the source material available to the system, and the material the system leaves out. This gives a much clearer account of influence than a generic statement that a human remains responsible.
A correction process needs the underlying evidence
Under the enacted law, an adverse outcome involving covered ADMT can trigger rights concerning inaccurate personal data and a request for meaningful human review and reconsideration, to the extent commercially reasonable. The definition of that review requires an authorized, trained reviewer who considers relevant available primary evidence and does not default to the system output. The correction provision does not require correction of opinions, predictions, scores, or protected evaluations. See sections 6-1-1701 and 6-1-1705.
In a meeting archive, factual errors and evaluations can be tangled together. Imagine a summary attributing a missed deadline to an employee who had actually warned that a dependency belonged to another team. A later assessment may describe that employee as a poor planner. Correcting the attribution and reconsidering the assessment are connected tasks, although they address different parts of the record.
A workable review would let the reviewer find the relevant exchange, understand who was speaking, inspect the surrounding discussion, and consider additional evidence supplied by the employee. The reviewer also needs enough context about the tool to understand whether the disputed text came from transcription, summarization, retrieval, or a later evaluation prompt. Asking the same model to explain its previous answer cannot establish whether the original account was accurate.
Preservation requires judgment. Keeping every recording indefinitely increases exposure and can make review harder, while deleting every source immediately can leave an organization unable to investigate a serious dispute. Legal, privacy, and HR teams should settle retention rules for decision evidence and compliance records together, including access restrictions and any applicable preservation duties. A meeting organizer should not have to invent that policy after a complaint arrives.
Build a reviewable workflow before January
The legislature’s enacted-law summary identifies documentation, recordkeeping, notice, adverse-outcome disclosure, and correction and review obligations. The right preparation is to connect each relevant obligation to a person and a system, rather than putting all responsibility into a procurement questionnaire.
For a proposed use of meeting analytics in employment, the review can begin with a compact record:
- Purpose and decision: Describe the employment decision and the exact assistance the system is expected to provide.
- Evidence and limits: Identify the meetings included, known gaps, speaker-attribution risks, and the outputs a manager may receive.
- Authority and challenge: Name the person who can investigate errors and reconsider a decision, with access to the necessary material.
- Changes and ownership: Assign responsibility for reviewing new prompts, integrations, model settings, and expanded uses of the archive.
These are practical implementation suggestions, not a substitute for mapping the final requirements. Their value becomes apparent when someone changes a weekly recap into a quarterly employee comparison. Without an owner for that change, a use approved for project coordination can enter an HR process without anyone reassessing it.
The October comment period also gives organizations a reason to test their proposed process against difficult examples now. A company that cannot determine whether a particular retrieval or summarization workflow falls within an exclusion can formulate a concrete question for counsel or the rulemaking process. A sample workflow, with personal information removed, is more useful than an abstract objection to regulation of AI.
Local processing addresses a different part of the problem
Keeping meeting content on a managed device can reduce the number of external systems receiving it. That can be valuable when the archive contains personnel discussions, customer information, or unreleased work. It does not establish whether an inference is sound or whether using it in a promotion decision is appropriate.
Summit’s security architecture documents local recording and transcription, with summarization available through Local Models. Selecting a hosted model sends relevant prompt content to the configured provider, and optional sync and exports introduce their own data paths. Those choices belong in the deployment review alongside device access and retention.
An employer choosing a local workflow still needs to decide which uses of meeting records are permitted, how errors are corrected, and when a new employment use requires review. The most useful preparation for Colorado’s January deadline is a record of those decisions that matches how people actually work, backed by a fresh check of the final rules once they are adopted.