California Just Made Human Review the Law for AI-Driven Discipline. Here's What Your Performance Records Need to Show.

Short answer: California's SB 947, the No Robo Bosses Act, was signed on September 30, 2026 and becomes operative on July 1, 2027. It bars employers from relying solely on an automated decision system to discipline or fire an employee. When a system is the primary basis for that decision, a human reviewer has to corroborate the output using other information, such as supervisor evaluations, peer feedback, or personnel records. The employee then gets written notice explaining that a system was used, what data it drew on, and who they can ask about the decision. For HR, the practical question is what that reviewer checks against. On most teams, the answer is a year of performance history that nobody wrote down.

California now requires a human to check AI before anyone is disciplined or fired

Governor Newsom signed SB 947 as part of a package of workplace AI bills on September 30. A near-identical predecessor, SB 7, was vetoed in October 2025, when the governor called its restrictions overly broad. The 2026 version narrowed the notice provisions and became the first state law of its kind.

The law works on two levels. First, a California employer cannot rely solely on an automated decision system when it disciplines or terminates an existing employee. Second, if the employer primarily relies on a system's output for that decision, it has to direct a human reviewer to corroborate the decision, and it has to give the employee written notice. Fisher Phillips reports that the notice must say the employer primarily relied on an automated system, that a human corroborated the output, and that retaliation is prohibited. It also has to describe the employee data the system used and give the employee a human contact who can explain the decision.

One detail deserves more attention than it is getting. Under the new Labor Code section 1526.1(c), once a worker shows an automated system was used in a disciplinary decision, the burden shifts to the employer to prove either that it did not primarily rely on the system or that it met the review and notice requirements in full. In plain terms, the employer has to show its work.

The law's author, State Senator Jerry McNerney, put the intent in one line: "AI must remain a tool controlled by humans, not the other way around."

The law names what the human checks against

SB 947 does not leave "human review" undefined. HR Dive describes the corroboration requirement as drawing on a nonexhaustive category of supporting information, including evaluations, personnel files, and other documents. Fisher Phillips describes it as the data that produced the output or other relevant information.

That matters because a reviewer who looks at a system's flag and clicks approve is not corroborating anything. Corroboration means checking the output against something the system did not produce, such as what a manager observed or what the personnel file shows happened and when.

So the question every HR team now has to answer before July 2027 is simple to ask and harder to answer. If a system flagged one of your employees tomorrow, what would the human reviewer open?

Most performance records can't carry that weight

Picture the request arriving in August 2027. A scheduling or productivity tool has flagged an employee in your California office, and the flag is heading toward a written warning. A manager is asked to corroborate it.

Here is what the manager typically has. Last December's annual review, written in an afternoon from memory and weighted toward the six weeks before it was due. A handful of 1:1 notes, if the manager keeps them. Some Slack threads. Whatever the manager remembers about March.

Under SB 947, that thin file has a hard consequence. If the reviewer cannot corroborate the system's output, or decides it is inaccurate, incomplete, or misleading, the employer cannot use it to discipline or terminate the employee. A flag with nothing on file to check it against is a flag the employer cannot act on.

None of that is the annual review's fault. The annual review is the right moment to step back and assess a year of work. The problem is upstream: the year itself was never captured. When the inputs are thin, the review is thin, and a human reviewer asked to corroborate a system's output ends up either rubber-stamping it or rebuilding a year from recollection. Neither is what the law describes.

Year-round context is what makes human review real

The records SB 947 names are the records performance management is supposed to produce: evaluations, peer feedback, and a personnel history that shows what happened over time. The difference between a human review that holds up and one that does not is whether those records were written when things happened or reconstructed later.

Employment lawyers are giving the same advice. Kilpatrick recommends that employers using automated systems keep contemporaneous records of what the system did, what the human reviewer considered, and when notice went out for each decision. The operative word is contemporaneous.

That is the design principle behind Upward365. Managers capture context during the year, in the moment, instead of all at once in December:

• Notes hold what a manager observed and when, including historical reviews and surveys imported at setup.
• Uplifts record recognition as it happens, so strong work is on file before the review, not remembered during it.
• Flight Checks give managers a regular read on how each person is doing between cycles.
• Briefings give 1:1s a shared agenda, so the conversation leaves a trail.

Reviews and peer feedback live in the same place as everything above.

When the annual review opens, the History Hub overlay puts that year in front of the manager. The review arrives with the year already in it, which is what turns the annual review into the payoff of a year of good management rather than a memory test.

None of this makes an employer compliant with SB 947 on its own. Compliance depends on how an employer uses automated systems, who reviews them, and what notices it sends, and that is a conversation for counsel. What year-round context does is give the human reviewer something real to check against.

Upward365 keeps AI on the file-cabinet work and people on the judgment

The new California law draws a line between what software can do and what a person has to do. Upward365 was built on the same line.

We use AI for the transactional work that buries managers: collating a year of notes, compiling what is on file, spotting trends worth a closer look. In Upward365, that AI-enabled help sits inside the History Hub overlay. It does not write reviews. It does not write comments about people. The judgment about a person stays with a person.

Some buyers want software that writes performance reviews for them. Upward365 is not built for that, and we would rather say so up front. AI Summaries, AI Recommendations, and the AI Ad-Hoc Report Generator are coming soon, and they will follow the same rule.

And when something needs a human on our side, a real person who knows your account answers the same day.

What to do before July 1, 2027

Nine months is enough time to get this right and not much more. A practical sequence:

1. Inventory every tool that scores, flags, or ranks employees. Attendance systems, productivity monitoring, quality scoring. If its output could feed a discipline or termination decision, it is on the list.
2. Map where those outputs touch discipline. For each tool, note who sees the output and what decisions it informs today.
3. Name the human reviewer. Decide who corroborates a system-driven decision and what they are expected to check.
4. Test your records. Pick five employees at random. Could a reviewer corroborate a hypothetical flag on each of them from what is on file right now? If the answer is no for most of them, the gap is your performance records, not your AI policy.
5. Build the notice into the termination packet. Include the data description and a named human contact. Fisher Phillips recommends delivering the notice at the moment of the decision rather than days later.
6. Talk to counsel about the rest of California's rules. Since October 1, 2025, California's FEHA regulations on automated decision systems have required employers to keep employment records, including automated decision data, for four years. The California Privacy Protection Agency's rules on automated decisionmaking technology add pre-use notice, opt-out, and risk assessment duties starting January 1, 2027. These are separate obligations from SB 947.

The common thread across all six steps is that the law asks employers to show a human was in charge. The cheapest time to build that record is while the year is happening.

Frequently asked questions

Does SB 947 ban AI in HR decisions? No. It bars employers from relying solely on an automated decision system to discipline or fire an employee, and it requires human corroboration and written notice when a system is the primary basis. Employers can still use automated tools; a person has to stay in charge of the decision.

When does SB 947 take effect? SB 947 was signed on September 30, 2026 and becomes operative on July 1, 2027.

Does a manager approving the system's output count as human review? Not on its own. The law asks the reviewer to corroborate the decision using other information, such as evaluations, peer feedback, or personnel records. Approving the output without checking it against anything else does not meet that description.

What happens if the reviewer can't corroborate the output? The employer cannot use it. If the output cannot be corroborated, or the reviewer finds it inaccurate, incomplete, or misleading, it cannot be the basis for discipline or termination.

What records does a human reviewer need under SB 947? The law describes a nonexhaustive category of supporting information, including evaluations, personnel files, and the data behind the system's output. In practice, the more of that record is written during the year rather than at review time, the more a reviewer has to work with.

This article is general information, not legal advice. Laws and their interpretation change, and how they apply depends on your situation. Talk to employment counsel before changing your policies.

See how Upward365 keeps a year of context ready for every review. Book a Demo