Connecticut AI Employment Tools: Notice Fields for Hire, Promote, Fire

The trigger is influence, not the label

If your employment process uses an AI tool, the trigger is whether the tool substantially influences the decision. Public Act No. 26-15 defines a covered tool as automated employment-related decision technology, or AEDT: software that analyzes personal data and produces outputs—forecasts, suggestions, categories, rankings, or ratings—that can materially affect an employment decision. If the tool ranks applicants and the manager follows the ranking, the influence pulls the process into the rule.

Ask whether the output changes what a reasonable decision-maker would do. When the manager would have hired the same person without the score, the tool may be a convenience. A score that narrows the field means the tool is doing work.

The Connecticut statute leaves out immaterial changes to tasks, duties, hours, or assignments, along with choices about employee safety, health, scheduling, or monitoring productivity. Those exclusions matter for routine operations. They do not cover a tool that recommends termination, blocks promotion, or filters resumes before a human sees them. Document the role of the tool in the process. A short note in the job file can say what the tool did, who reviewed the output, and what the final decision was.

The notice fields are the file

The written notice is the part that survives a later dispute. Make it specific enough that a reviewer can tell what the tool did and who can answer questions. Before a covered employment-related decision, the employer must give written notice identifying the AEDT's trade name and purpose, the personal-data categories and sources, the assessment method, and a contact for the employer. The list below is the minimum set of items a manager should be able to point to in a file.

• Write the tool's trade name and purpose.
• List the personal-data categories and their sources.
• State the assessment method.
• Name an employer contact for questions.
• Confirm the notice goes out before the covered decision.

The trade name matters because a generic label like AI screening does not identify the product. If the vendor changes the name, update the notice. Make the data categories concrete. Resume data is better than applicant information. Work history from the application is better than relevant data.

The assessment method can stay high-level. Say enough to show what the tool is doing, such as ranking applicants by predicted job performance. The contact should be someone who can answer follow-up questions. A shared inbox is better than a department name. Timing is the easiest failure. Put the notice in the workflow before the decision meeting.

The dates split the duties

The AEDT interaction-disclosure and pre-decision written-notice duties apply to use on or after October 1, 2027. From that date, Connecticut deployers must inform employees and applicants, in clear terms, that they are dealing with AEDT. The interaction disclosure is separate from the written notice: one is the statement that employees and applicants are dealing with AEDT, and the other is the file-level record.

Starting October 1, 2026, Connecticut WARN notices for plant closings or mass layoffs must answer whether the layoffs are connected to the employer's AI use or some other technological change. If the layoff is unrelated to the tool, say so. When the layoff follows a change in tooling, say that too. The answer should be consistent with the business record. A manager should not wait for a layoff to figure out whether the tool was part of the reason. The same file that shows the tool's role in hiring can show its role in a reduction in force.

Bias testing is mitigation, not immunity

Using covered AEDT cannot defeat a discrimination claim, though courts and state human-rights authorities may weigh bias-testing or comparable preventive steps as a mitigating factor. Keep the testing file separate from the decision file. It should show what the employer did before relying on the tool, not what the tool said after the fact.

A useful file includes the date of the test, the population used, the metrics reviewed, and the people who reviewed the results. Also show what changed after the test, if anything.