Two phrases will decide what California's AI workplace laws cost employers

Crypto Briefing's report on the package spends one line on cost: companies using automated tools for performance management or termination may need to re-evaluate their AI strategies, and adding human reviewers could raise costs ^[17]. It does not put a figure on that. The figure depends on a threshold. SB 947's review duty applies to decisions made "primarily" by an automated decision system ^[1]. The same report names that word, along with SB 951's "in substantial part", as the terms that will set how wide the net is ^[6].

The cost is a count times an hourly rate. Take the discipline and termination cases that cross the threshold and multiply by the reviewer time each one takes. The count can come out more than one way. Read narrowly, "primarily" covers only cases where software acted alone. Those are already barred, because SB 947 separately forbids relying solely on an automated system to discipline or fire ^[2]. On that reading the review rule adds close to nothing. Read broadly, it covers any case where a score or a flag carried most of the weight, and the reviewer count climbs with every performance cycle. A third outcome is a redesign in which a manager opens each case and the software only recommends, built so the decision falls below the threshold.

I'd expect that third outcome at employers running large HR systems, since it keeps the software and puts the human at the start of the process. The counter-case is that the redesign costs about what a reviewer would have, because a manager reading every file is the same expense under a different job title. The view is wrong if the term is read to reach a ranking that a manager merely signs off on. Then the redesign route closes and the cost scales with case volume.

Complying with SB 951 means putting a statement in a filing ^[10]. In my view the exposure is consistency. Suppose an employer tells shareholders that automation let it run with fewer people. In California it has then described a disclosure trigger ^[10], and a notice that leaves AI out of the same cut contradicts the company's own public account. Whether the two accounts can differ comes down to how "in substantial part" is read ^[6].

The package does not ban AI at work ^[8], so employers are not being asked to retire their systems or write them off. The money goes on a human step added to software they already run. Employers have 274 days from the September 30 signing ^[3] to the July 1, 2027 start of the ADS rules ^{[9] [15]}. In that window they have to decide which cases cross the line, or rather, find out whether anyone will define the line for them first. Crypto Briefing says employers will be looking for clarity on those terms well before the effective dates ^[7].