The AI Your Customers Talk to Is Getting Smarter - Is Yours?

If you tried an "AI chatbot" in 2023 and walked away rolling your eyes, you weren't wrong. The thing misunderstood half of what customers said, made up answers to the other half, and couldn't hold a conversation longer than two turns without falling apart. A lot of business owners tried it, got burned, and mentally filed AI under "not ready yet."

Fair. But a lot has changed in eighteen months - and the pace picked up again this summer. On June 30, OpenAI received U.S. regulatory clearance for a broad rollout of GPT-5.6, its newest frontier model. A week earlier, it also shipped a lower-latency Realtime voice model - the piece that powers AI phone assistants. Neither headline made the evening news, but together they quietly moved the goalposts on what an "AI assistant" for a small business can actually do.

What actually changed under the hood

The easiest way to think about it: the AI that customers talk to today is a different animal than the one from a couple of years ago. Not marginally better - categorically better at the things that used to make it feel dumb.

• It understands context. Older models would forget what you said three sentences ago. Current-gen models can hold a full conversation, remember that the caller already mentioned their address, and connect that to the fact they're asking about a follow-up appointment.
• It makes fewer confident mistakes. "Hallucinations" - where the AI made up an answer that sounded right - used to be common. Current-generation models are noticeably better at sticking to what they actually know and saying so when they don't, though it still isn't a solved problem.
• It sounds human on the phone. The new Realtime voice models cut response latency to roughly the pause a real person would take, and they handle background noise - a running truck, a busy office, a shop fan - without falling apart. That was the single biggest thing making AI phone assistants sound obviously robotic a year ago.
• It follows instructions more reliably. Give the current generation a set of business rules - how to greet callers, when to escalate, what never to promise - and it actually sticks to them. Older models would drift.

Why this matters for small business owners specifically

The gap between "toy" and "employee" is the gap between an AI that guesses and an AI that knows. Frontier models closed a lot of the guessing gap. But raw intelligence isn't enough on its own - a smart assistant with no information about your specific business is still going to say something wrong. The thing that turns a general-purpose model into something useful is grounding it in your business's actual rules, prices, policies, and history. (That's what a "second brain" is for, and it's why the same underlying AI can sound generic in one deployment and sharp in another.)

The demand-side pressure is also real. A growing share of routine customer service interactions are being handled by AI, and customers are increasingly comfortable with it - as long as it doesn't feel stupid. The same pattern shows up elsewhere: the businesses that quietly deployed current-gen AI in the last twelve months are pulling ahead not because the technology is magic, but because their competitors are still using tools built on last generation's assumptions.

One more thing worth watching: the rules are getting clearer

One reason small business owners have hesitated is the patchwork of state-level AI rules - disclosure requirements, consent laws, restrictions on automated calls that vary depending on where your customers live. That's been a real headache. The draft "Great American AI Act" currently moving through Congress proposes federal preemption of most state AI rules, which - if it passes - would give small businesses a single standard to follow instead of fifty. It's still a draft and could change substantially, so nobody should base a decision on it yet. But the direction of travel is toward less compliance ambiguity, not more.

If you dismissed AI a year or two ago, it's worth another look

Not because AI is suddenly perfect. It isn't. It still needs to be set up carefully, connected to your actual business information, and monitored so it doesn't drift. But the reasons you dismissed it in 2023 - sounds robotic, misunderstands people, makes stuff up, can't follow rules - are largely fixed problems now. The version your competitors are quietly rolling out this year is not the version you tried.

If you want more plain-English pieces like this, the blog has a Friday digest that sums up each week's posts.