In Britain, Co-op Legal Services, the legal arm of the retailer, has an AI model record and grade calls from people who are often newly bereaved and asking about wills and estates. In California, a law signed on September 30 will stop employers from using AI to read how their employees feel. The two stories look like opposites, but they describe one rule that will outlast both: AI may check what a worker did, and it may not be used to guess what the worker feels.
When the sample disappears
Take an adviser who handles 20 calls a day, about 400 a month. Under the old way, a supervisor might hear 8 of them. Under the new one, all 400 are graded on 50 or more points each, which is about 20,000 small judgments a month about one person. No manager reads 20,000 results, so what reaches the manager is the percentage. That is why the contents of the score decide so much: a few badly chosen checks can pull down a good adviser every month, and she will never be able to see why. We have already covered how scoring every call changes the way staff feel. The next question is what a score is allowed to contain.
What the California law bans, and what it leaves alone
The law is AB 1883, and it starts on January 1, 2027. It covers any tool that collects what employees say, do or show without a person watching, and call recordings are inside that. Those tools may not use AI to infer an employee's emotional state, or to collect neural data such as brain activity from a headband. Software that infers mood from call transcripts is likely covered too.
Nothing in it stops a system from checking whether an adviser read back the appointment, mentioned the fee or asked the required question. A second California law, which starts on July 1, 2027, adds that a human must back up any AI-based discipline or firing with other evidence. So the pattern is that AI may grade, a person must decide, and AI may not read minds. Nothing in the law reaches the Co-op in Britain, but any employer who scores calls will face the same question from staff.
Why the law draws the line at feelings
A missed step is a fact. Anyone can open the transcript, see that the adviser never mentioned the fee, and fix it on the next call. A score for tone is an opinion from a model, and the person marked down has nothing to point at. Kaiser Permanente nurses in California who staff advice phone lines have complained that an AI tool rating their empathy and tone of voice graded them wrongly. A calm voice on a serious call is exactly what a good nurse sounds like, and exactly what a tone model can mistake for coldness.
Here is what that looks like for a clinic manager who runs a reception line. She could score every call on five checks: patient's date of birth confirmed, earliest slot offered, appointment read back, fee mentioned, next step stated. Each failed call comes with the transcript line where the step was missed, and the coaching conversation takes two minutes. Add a "warmth" score on top and that conversation turns into an argument about whether the receptionist sounded kind.
Call-scoring products often add sentiment analysis, so I expect the first effect of the California law to be that vendors rebuild those scores around checklists, at least for staff. Employers elsewhere who copy the checklist approach now will not have to rebuild later.
If you manage people and use call or chat scoring, or are about to buy it, ask which of its scores are built from tone, mood or sentiment, and switch those off first. Keep the ones a person can check against the transcript and explain in one sentence.