Texas TRAIGA Checklist for AI Hiring and Firing

The law asks for proof, not a fairness score

Before the tool scores another candidate or termination packet, build the file that will support the deployment. Texas TRAIGA took effect on January 1, 2026, and it governs entities that deploy AI in Texas.

Only the Texas attorney general can enforce the law, and the statute gives no private right of action.

Under TRAIGA, a hiring or firing AI tool is unlawful only when the developer or deployer intends to unlawfully discriminate against protected classes, and skewed results by themselves do not prove that intent.

The statute supplies the legal floor. The pre-deployment review is good practice that makes the process defensible.

Build the file before the tool goes live

The checklist is a go/no-go test. If any answer is no, pause deployment and close the gap before the tool touches a real person. Each item maps to a record you can produce, a decision you can defend, or a gap you can close.

• Confirm the vendor contract names who owns model changes, data use, and post-deployment safeguards.
• Collect the model and data documentation.
• Identify protected-class risk and set a human review step before any adverse action.
• Build adverse-action notices that state the reason, the data considered, and the review path.
• Set monitoring and records retention for post-deployment performance and safeguards.
• Note that private employers have no workforce-facing AI disclosure duty under TRAIGA; the disclosure rule is directed at government agencies that interact with consumers.

Section 552.103 gives the Texas attorney general authority to request records describing what an AI system does, what data trained it, what it receives and produces, how it performs, its known limits, and the safeguards used after deployment.

Contract language matters because a model can change after you sign. If the vendor swaps a feature, adds a data source, or retunes a score, you need a named owner and a notice path.

Risk review keeps the team from assuming neutrality. A human review step gives the file a documented pause before an adverse action.

Notices protect the decision when the person asks what happened. Monitoring catches changes before they become a problem. If the tool changes, update the file before the next decision. You can still explain the process voluntarily.

The records request is the practical deadline

The file should exist before a request arrives. A records request is not a negotiation.

TRAIGA sets civil penalties of $10,000 to $12,000 per curable violation, $80,000 to $200,000 per uncurable violation, and $2,000 to $40,000 per day for ongoing violations.

The amounts are the starting point for a compliance budget. An ongoing issue can outpace a one-time fine if the tool keeps running while the gap stays open.

Keep the vendor file, the model file, the decision file, and the monitoring file in one place. When an AI termination decision is made, the file needs to show who reviewed the output, what the human considered, and why the final decision was made.

When an adverse action is taken, the notice should match the file. The version used for each decision should be in the file.

The statute also bars government bodies from deploying AI that scores or categorizes people from social behavior or personal traits when the purpose is to create a social score. Build the file for a lawyer, not just a vendor demo.

Make the first decision the test

Run one real decision through the checklist before the tool is used broadly. Pick a low-stakes screen, a single job family, or a small termination cohort.

Have a manager, an HR lead, and a vendor contact sign off on the file. If an answer is no, fix the gap before the next candidate or employee is affected.

Keep the signed file with the job requisition or termination packet, not in a shared chat thread. That test creates a date, a version, and a record.