AI changed the job about a year ago. The job description, the exemption test, the confidentiality rules, and the paper trail did not get the memo. That gap doesn’t cause a problem while everything is running smoothly. It causes one the first time an accommodation request, a termination, or a discovery request asks the paperwork to explain what’s actually happening.
Other Duties as Reassigned
AI absorbed part of somebody’s job about a year ago. The job description, the classification, and the paper trail did not get the memo.
Somewhere in the company, an employee started using an AI tool. Probably in marketing, or accounting, or whoever ended up owning proposals. It worked. Their output went up, the turnaround on a recurring task dropped from two days to two hours, and nobody said anything because nothing appeared to be wrong.
There was no procurement process, no policy, and no approval, because the tool cost twenty dollars a month and required no installation. This is not a management failure. It is what happens when the friction to adopt something drops to zero. The same instinct that got the company from six people to thirty is the instinct that says try it and see.
The tool is not the problem. The problem is that the org chart quietly changed and every document describing it stayed the same. That gap is a paperwork problem today. It becomes a legal problem the first time someone outside the company reads the paperwork.
Three Things Break When a Task Moves
First, the job description stops matching the job. Job descriptions identify essential functions, and essential functions drive the reasonable accommodation analysis under the Americans with Disabilities Act (“ADA”). When the written description still lists tasks the software now performs, and says nothing about the review and verification work the role actually requires, the company is defending itself with a document that describes a job nobody holds. If the company denies an accommodation because the employee cannot perform a task the software now handles, that denial is hard to defend. That gap surfaces at the worst possible moment, which is after an accommodation request has been denied.
Second, restructuring can look like a cover story. Restructuring starts to read as pretext when it is undocumented. Administrative and support positions may be among the roles affected by automation, and at many companies those positions skew older or female. Eliminating or consolidating them can be lawful. Doing it without a contemporaneous written business reason, and then assembling that reason after the charge arrives, is how an ordinary operational decision can take on the shape of a discrimination claim. “We just stopped needing the role” is a conclusion, not a record. The file needs the reasoning, dated before the decision.
Third, overtime exemptions can quietly fail. Exempt classification breaks silently. When duties shift, the duties test (which looks at the work an employee actually does, not the title) does not shift along with them as a courtesy. Move an exempt employee’s substantive work to a tool, backfill the role with oversight and cleanup, and the exemption the company has relied on for four years may no longer hold. Misclassification can bring back overtime wages, liquidated damages, and a lookback period that may extend to three years for willful violations. And when multiple employees perform substantially similar work, the exposure may extend beyond the person who first raises the issue.
The Wage and Hour Problem Nobody Puts on the Agenda
Output expectations recalibrate to the fastest adopter on the team. Once one person is producing at three times the old pace, that becomes the informal standard for everybody, including nonexempt (overtime eligible) staff who were never told which tools they were allowed to use.
The nonexempt employee keeping up with that standard may be doing it on a personal account, on a personal phone, after hours, unrecorded. If the employer knows or has reason to believe the employee is performing work, that time may be compensable. It is also completely invisible until a former employee’s counsel asks for the app history, at which point it becomes a very legible pattern.
The Confidentiality Exposure Sitting in the Chat Window
Employees who find a tool useful use it, and they use whatever free or personal account they already have, under terms of service nobody has read, which may permit inputs to be retained or used to train the underlying system.
For a company holding client files, patient information, salary data, or an open investigation, that is a confidentiality problem with a compliance dimension. For a company holding anything it intends to call a trade secret, it is worse. Protection under the Texas Uniform Trade Secrets Act and the federal Defend Trade Secrets Act requires reasonable measures to maintain secrecy. If a company cannot show it took reasonable steps to keep information secret, a court may decide it was never a trade secret at all. Pasting the pricing model into a public chat window can undermine that argument. No policy prohibited it, because there was no policy, which is precisely the argument opposing counsel may make. The absence of a rule doesn’t read as neutral in court. It reads as a company that never asked the question.
What Texas Actually Requires Now
The Texas Responsible Artificial Intelligence Governance Act (“TRAIGA”), HB 149, took effect January 1, 2026. It is considerably narrower than employers were bracing for. It prohibits developing or deploying an AI system with the intent to unlawfully discriminate against a protected class, and disparate impact alone is not sufficient to demonstrate discriminatory intent under the statute. It also dropped the impact assessment requirements that appeared in an earlier version of the bill. Enforcement runs primarily through the Attorney General. Employees cannot sue under it, and the Attorney General must give notice and a 60 day cure before seeking penalties.
The provision that matters for a company this size is not just the prohibition. It is the structure around enforcement. TRAIGA expressly recognizes substantial compliance with the NIST Generative AI Profile or another nationally or internationally recognized AI risk management framework in connection with its liability provisions. If the Attorney General investigates, the company may have to describe the system and produce all relevant documentation as identified by the Attorney General. A company with documented AI governance has something to produce. A company without it has a shrug and a Slack thread.
For anyone with employees in a second state, the math gets worse before it improves. For example, Colorado Senate Bill 26-189 takes effect January 1, 2027, and its obligations differ meaningfully from the Texas approach. Colorado’s law covers AI tools that materially influence hiring and other employment decisions about Colorado applicants and employees, and it requires advance notice, disclosures after the decision, and record retention. One policy drafted with the stricter requirements in mind is achievable. One policy drafted for Texas and quietly applied to the Denver hires is a discovery exhibit waiting for a caption.
What Unverified Output Costs the People Standing Next to It
A useful illustration comes out of Gauthier v. Goodyear Tire & Rubber Co., a wrongful termination case in the Eastern District of Texas. The former employee’s lawyer filed a brief citing two cases that did not exist, along with quotations that appeared nowhere in the authority cited, all generated by an AI tool and none of it adequately checked. The court sanctioned the lawyer two thousand dollars.
The employer defending the case reported spending $7,521.26 on a reply brief that required tracking down cases that did not exists. Even after Goodyear flagged the fake citations, the lawyer did not go back and check his work until the court ordered him to explain himself.
The company that did not irresponsibly use AI still had to pay to clean it up (i.e. investigate the cases that did not exist for purposes of a reply brief).That is the shape of this risk generally. Unverified output is rarely most expensive for the person who generated it. It is most expensive for whoever has to deal with the consequences, and inside a company, that is usually the employer. Nobody at the company had to be a bad actor for the company to pay for it.
The court’s response is the real lesson. It did not ban AI. It updated its paperwork. Within months, a judge in the same district required every filing to state whether AI was used and certify that someone checked the output. Swap the lawyer for an employee drafting a client proposal and the employer is the one paying for the cleanup. Verification has to be a written part of the job, not an assumption.
Five Things Worth Doing Before Year-End
- Write the policy. Approved tools, prohibited categories of information, a verification requirement, and a disclosure obligation (such as requiring employees to tell their manager when AI helped produce a work product). Two pages beats zero pages, and a dated two-page policy creates a record that the company actually addressed the issue.
- Update the job descriptions. Update the job descriptions for every role whose work has actually changed. Make review and verification an explicit responsibility instead of an assumption.
- Document the business reason. Document the business reason for any role that shrinks, merges, or disappears, and do it before the decision rather than after the charge.
- Recheck exempt classifications. Make a habit of rechecking exempt classifications wherever duties have moved.
- Adopt a recognized risk management framework. Adopt and regularly audit a risk management framework, being sure to keep the record of adoption. Under TRAIGA, substantial compliance with a recognized framework can matter when a company discovers and addresses a potential violation through its internal review process.
None of the steps above require slowing down adoption. None of it requires anyone to admit that the last eighteen months were handled badly. They were handled the way every growing company handles a cheap tool that works. The paperwork simply has to catch up with the org chart before somebody else notices the gap first.
AI does not need permission to change how work gets done. But once the work changes, the documents describing it need to change too. Treaty Oak’s HR process audit looks for those gaps across job descriptions, classifications, hiring and separation practices, and other HR processes—before an outdated piece of paper is asked to explain a job that no longer exists.