California's new law on AI and firing takes effect next July. Here's how a company fails it

August 2027, a composite

This is an illustrative scenario, not a real company. It's August 2027. A distribution business in California with a few hundred employees receives a letter from a former warehouse worker's lawyer. The worker was let go in July after three months of written warnings. Each warning cited the same thing: a pick-rate score from the warehouse system, below target.

The supervisor did look at the numbers. He talked to the worker twice. He believed he'd done a human review. But when the company's lawyer asks for the record, there's nothing to show what he checked. The warnings quote the score and nothing else. Nobody sent the worker a notice saying software played a major part. Nobody can say which data went into the score.

Nobody bought an AI firing tool. They bought a productivity dashboard, and the dashboard became the reason.

That's the pre-mortem. Now let's work backward through what went wrong, and what to fix while there's time.

What the law actually requires

California Governor Gavin Newsom signed SB 947, which its supporters call the "No Robo Bosses Act," on September 30, 2026. According to Ogletree Deakins' summary of October 30, it takes effect on July 1, 2027. The key points:

• No sole reliance. An employer may not rely solely on an automated decision system (ADS) to discipline or fire someone.
• Human corroboration when it's the main factor. If an employer primarily relies on an ADS, it must direct a person to corroborate the decision using the data behind the output, or other evidence such as supervisor evaluations, personnel files, work product, peer reviews or witness interviews. If the output can't be corroborated, or turns out to be inaccurate, incomplete or misleading, it can't be used.
• A written notice afterward. When software was the primary basis, the employee gets a plain-language, stand-alone notice saying so, saying a human reviewed it, naming a contact, and stating that retaliation is prohibited.
• Access to their own data. On request, the employee gets a meaningful description of their own data the system used.
• Penalties. $500 per violation, enforced by the Labor Commissioner and public prosecutors, with possible claims under California's Private Attorneys General Act. Once ADS use is shown, the employer has to prove it didn't primarily rely on it or that it complied.

This is a summary for operators, not legal advice. Your employment counsel should decide how the law applies to you.

Failure one: not knowing what counts

The definition of an ADS is broad: any computational process from machine learning, statistics, data analytics or AI that produces a simplified output, such as a score, classification or recommendation, used to assist or replace human judgment and that materially affects people. Ogletree notes that productivity scoring tools, attendance analytics, performance trackers and customer-rating systems could all potentially qualify. The law excludes things like spam filters, firewalls, calculators, databases and datasets.

In our scenario, nobody thought of the pick-rate score as "AI." It came bundled with the warehouse software. That's the most likely failure: the system that matters isn't the one with AI in its name. It's the scheduling tool that flags "attendance risk," the help-desk platform that ranks agents by customer rating, the sales dashboard that sorts reps into a bottom tenth.

Fix now: list every system that produces a score, rank or flag about a person, including features inside vendor platforms you already pay for. Ask, for each one: has this ever been cited in a warning or a termination?

Failure two: a review that left no trace

The supervisor in our scenario did review. He just didn't record it. Under SB 947, the employer has to show compliance once the system's role is established. A review that only lives in a manager's memory is very hard to show.

The law also asks for something specific: corroboration using the data behind the output, or other relevant evidence. "I looked at the score and agreed" is not corroboration. Corroboration looks like: "I checked the scans for those weeks. Two of the low days were when she was covering returns, which the system counts as idle. The remaining weeks are below target, consistent with what her lead reported."

Fix now: build the review into the workflow, not the policy manual. When a warning or termination cites a system output, the form should ask the reviewer which data they checked, what they found that the system didn't account for, and whether the output holds up. If it doesn't, the form should make it impossible to proceed on that basis. That's exactly the kind of check we argued for when Anthropic required human review in high-risk uses: a human step that is real, specific and recorded.

Failure three: the notice nobody sent, the data nobody could find

The worker's lawyer asked for a description of her data. The company had to ask the vendor what the pick-rate score included. It took six weeks.

Fix now: for each system on your list, find out what data it uses and how to pull it for one person. Draft the postuse notice template while the law is still eight months away. Name the contact person. Put the notice into the same workflow as the review form, so sending it isn't something someone has to remember.

There's a second judgment question underneath all this. "Primarily relies" isn't defined in the law, and Ogletree expects it to be tested in court. A company can try to stay on the safe side of that line by making sure the score is never the main reason. That's a legitimate choice, but it only works if the other reasons are real and written down. Otherwise it's the same decision with a different label.

Where to start this quarter

A reasonable plan for a small or mid-sized California employer:

1. Inventory every scoring and flagging system that touches people, built-in features included.
2. Mark which ones have been cited in discipline or terminations in the last year.
3. Redesign the warning and termination forms to capture what was checked and why.
4. Template the postuse notice and the data-description response, and test the data pull.
5. Train supervisors on what corroboration means, with examples from your own work.

If your company is building or buying tools that score, rank or flag employees, our AI applications and agents team designs the review steps, records and notices into the software itself, so compliance doesn't depend on someone's memory. Get in touch and tell us which score your managers lean on most.

Related work: Sutton Place.