October 6, 2026
California put a 15-minute clock on reaching a human. The clock is the easy part.
AB 1609 gives Californians the right to reach a person within 15 minutes. Getting a person on the line is easy. Handing them a case they can act on isn't.
On September 28, Governor Newsom signed AB 1609, the Right to Human Customer Service Act. Starting January 1, 2027, a business with more than $500 million in revenue that serves California customers has to give people a visible way to ask for a human during business hours. Once someone asks, the company has to make a good-faith effort to connect them within 15 minutes, or offer a callback or appointment within one business day. Chatbots have to say they’re chatbots. Penalties run $5,000 for a first violation and $10,000 after that (Quarles has a clear summary).
The first draft of the bill said five minutes. Lobbying got it to fifteen.
Most of the coverage is about the clock. We think the clock is the easy part.
A person on the line is not the same as help
Picture the compliant version. A customer spends ten minutes with a bot about a double charge. They type “agent.” Fourteen minutes later a person joins and writes, “Hi! How can I help you today?”
That company met the law. The customer now has to explain the double charge again, to someone who can’t see what the bot already said, what it checked, or what it half-promised. Industry group Internet.Works opposed the bill partly on the grounds that time pressure pushes agents “to prioritize speed over actually solving” the problem. That’s a fair worry, and it points at the real issue. The law measures how fast a person shows up. It can’t measure whether they show up holding anything.
A handoff that works carries three things with it:
- The thread. Every message, in order, so nobody asks the customer to start over.
- What the AI was about to do. If it drafted a reply or proposed a refund, the person should see that draft and decide on it, not discover it later.
- Why it stopped. “Customer asked for a human” is different from “this is a billing dispute the AI isn’t allowed to touch.” The second tells the person where to start.
None of that is in the statute. All of it decides whether the 15 minutes were worth anything.
The better handoff happens before anyone types “agent”
There’s a second problem with designing around the request button. It assumes the customer is the one who notices the AI is out of its depth. That’s backwards. By the time someone is typing “human” in all caps, the AI has usually already answered something it shouldn’t have.
The business knows ahead of time which questions are routine and which aren’t. Store hours, shipping times, how to reset a password: safe. Refund exceptions, a damaged custom order, anything involving a contract or a complaint: not safe for an AI to settle alone, at least not yet. That map should be drawn before launch, and the AI should hand off on its own when a conversation crosses into a topic it hasn’t been trusted with. The customer never has to ask.
This is how we built Celeste. It goes live one topic at a time. On topics the owner hasn’t graduated yet, Celeste drafts the answer and holds it, and a person sees the conversation and the draft together and approves, edits, or takes over. The AI proposes. Rules written in code decide what it’s allowed to send. The owner graduates a topic when the record shows Celeste gets it right, not because a model says it feels confident.
If you’re a small business, this still applies to you
AB 1609 only covers companies over $500 million. A 12-person shop in Fresno is exempt. It won’t stay irrelevant, though. Laws like this set what customers expect, and once a customer has been told by the big retailers that they have a right to a person, they’ll expect it from you too.
The good news is a small team can do this better than a giant call center can. You don’t need a 15-minute queue SLA. You need three things:
- A short, honest list of what your AI answers on its own, and everything else goes to you.
- Handoffs that arrive with the full conversation and the draft attached.
- A way to see, over time, which held answers you approved without changes. Those are the topics ready to graduate.
That last one matters most. A handoff policy you never revisit just turns into a bigger inbox. One that tracks what the AI got right lets you hand off less each month, for reasons you can point to.
California decided customers deserve a person within 15 minutes. Fine. The businesses that come out ahead will be the ones where that person already knows the whole story when they say hello. If you want to see how a topic-by-topic rollout works for a small team, here’s the product, or talk to us.