Rethinking AI customer support after California’s right-to-a-human law

Governor Newsom signed Assembly Bill 1609, the Right to Human Customer Service Act, on 28 September 2026. The law doesn’t come into force for another year, though the news has already got support directors in larger companies pulling up their current AI setups and asking questions they hadn’t planned to ask this quarter.

Any business selling into California with annual revenue over $500 million now has to give customers a route to a human agent within 15 minutes of asking for one, which means AI can still take the first message, but a human has to be reachable inside that window or the regulator gets involved.

The law itself stops at the California border, although the thinking behind it travels further, which is why UK businesses already running AI in support have a reason to look at the handoff path now, rather than wait for a regulator or a bad review cycle to force the conversation.

Customers have stopped tolerating bad bots

Most people put up with chatbots fine until the bot stops being useful, at which point patience tends to fall off a cliff, which is roughly where the Qualtrics 2026 Consumer Experience Trends Report pitched its findings. Nearly one in five consumers now see no benefit at all from AI customer service, a failure rate four times higher than other AI applications in daily use.

That kind of figure doesn’t come from one bad conversation. It accumulates over months of similar experiences, in which the bot handles the first question fine, loses the thread when things get more complicated, and leaves the customer stranded when they finally ask for help.

Resolution is what customers came for in the first place, because a reply that lands in three seconds but doesn’t fix anything is worse than a slower reply that does. Businesses that treated chatbot rollout as a way to shrink the ticket queue are now watching the consequences land in their churn reports and their Trustpilot feeds.

The California law is what happens when that kind of frustration builds up enough political weight to turn into statute, and it almost certainly won’t be the last of its kind. Whether the UK follows with a similar right is a conversation for another time, though waiting for the ICO to send a letter before reviewing an escalation path is a slow way to lose the customers a business least wanted to lose first.

The handoff is the point of failure

Most AI customer support deployments fail in the same place, and the point of failure is rarely the first exchange. The bot handles the opening well, which is exactly why customers let it try, and the problems only really start when the AI runs out of road and has to hand the ticket over.

As eesel.ai put it in a September 2026 piece on handoff best practice: “the handoff from AI to human is the single highest-stakes moment in any AI support setup, and it’s the part most teams bolt on last.”

A bot handing an unresolved ticket to a human agent who has no context, no transcript and no idea what’s already been tried is actively worse than no bot at all, since the customer has now explained themselves twice, waited twice, and watched the company fumble the one thing meant to make the whole experience smoother. Getting that goodwill back is hard work, assuming it’s possible at all.

Take a Saturday-morning ticket at a small UK retailer running its tier-1 chat through AI. The customer’s delivery hasn’t arrived and the tracking has been stuck on ‘out for delivery’ since Thursday. It pulls up that same screen, apologises in the generic way bots do, and tells her to wait another 48 hours. She asks for a human and the bot passes the ticket on, at which point an agent opens it half an hour later, sees “customer wants human assistance” in the handover notes and nothing else, and asks her to describe the problem from the beginning.

That’s a cold handoff, and it’s the default behaviour of most bolt-on AI tools on the market, which burns customers who never log a formal complaint because the ticket closes eventually anyway. A warm handoff is the opposite picture: the full conversation travels across with the ticket, including what the AI tried, the customer’s order detail, the escalation reason and a sensible next step for the agent. The code to do all of that is the easy part. Keeping every handoff warm as the queue grows is the discipline that slips first.

AI’s other job is behind the scenes

Sam Tolley, founder of SupportBuilt, made a point on LinkedIn recently that’s worth pulling out of the feed, because most of the public debate about AI in customer support is stuck on one question: how much of the customer-facing work can the bot absorb. He argues that framing misses where AI often does its most useful work.

Most of that work happens where customers never see it. It triages tickets so a critical case doesn’t sit behind a run of password resets, summarises long conversations so an agent coming back from lunch can pick up in thirty seconds instead of ten minutes, and flags when complaints about one specific feature have tripled in a fortnight, which is the kind of pattern a human manager working from a weekly dashboard tends to spot too late. The same system can catch a customer who’s been in touch three times in two weeks about related issues, because that customer is probably on their way out and nobody has told the retention team yet.

None of this is visible to the customer. What they see is an agent who takes their ticket faster, with more context, and more consistently than would otherwise be possible. And this kind of internal AI tends to earn its keep more reliably than deflection does, because it isn’t trying to replace the parts of the experience customers actually value.

The obvious question is why so few teams have prioritised it so far, and part of the answer is commercial. Vendors selling AI support tools have good reasons to push the customer-facing side, since “we cut ticket volume by 40%” is an easier number to put on a slide than “our agents resolve 15% faster.” The second number is often worth more in practice, but it’s a harder sell in a room full of people who’ve never actually worked a support queue themselves.

A head start for UK SMEs

AI adoption in the UK is still catching up with the hype attached to it, which the ONS Business Insights figures from June 2026 make clear: 29% of UK businesses now use some form of AI, up from 9% in September 2023, with the figure sitting at 44% among businesses with 250 or more staff. The direction of travel is obvious, even if the scale still has some ground to cover before adoption and customer expectation line up.

That gap works to the advantage of most UK SMEs, since there’s still a chance to set the AI customer support stack up around the right idea first time, rather than redesign it after Trustpilot starts filling up. Three moves are worth making now, whether a deployment is live or still on the whiteboard.

The simplest is to decide before launch what the AI should never handle. Billing disputes, account security and anything involving a customer who’s clearly upset should go to a person every time, regardless of what the AI’s confidence score suggests it can do.

The second is to make the exit to a human obvious and one click away, because burying “talk to an agent” three menus deep is the fastest way to produce the kind of consumer frustration California’s law was passed to address.

The move most teams miss, though, is to treat every handoff as data, since a cluster of escalations around one topic almost always points to a missing help article, and writing that article will do more for the queue than retraining the model will.

AI in customer support works best when it makes the human part of the operation better, which is where California’s law is useful even for businesses it will never touch directly: it’s a signal about where customer expectations are heading, and there’s no prize for being the last business to work this out.

Frequently asked questions

What does California’s Right to Human Customer Service Act require?
Businesses with over $500 million in annual revenue have to give California customers a route from automated systems to a human agent within 15 minutes of a request. Enforcement sits with public prosecutors, and there is no private right of action for consumers.
Does the California law apply to UK businesses?
Only where a UK business sells into California and crosses the revenue threshold, although the regulatory signal is more immediately relevant for most UK SMEs than the letter of the law itself.
What’s the difference between a cold and a warm handoff?
A cold handoff passes the customer to a human agent with no conversation history or context, forcing them to start again, whereas a warm handoff carries the full thread across, so the agent can pick up without asking the customer to repeat themselves.
Should SMEs use AI for customer support at all?
Yes, with the caveat that the highest return often comes from AI that supports agents (triage, summarisation, pattern detection) rather than AI that tries to replace them.