On September 8, Meta released Muse, a personal agent that connects to a person’s email, calendar, payments and shopping and works through tasks on its own. On September 29, OpenAI released Dots: always-on agents inside ChatGPT, each with its own cloud computer and browser and access to thousands of apps. The same day, Meta added Muse for Small Business. Google’s assistant already places calls to businesses to check prices and availability on a person’s behalf. None of this is a demo anymore. It’s a product with a price, and people are paying for it.
The thing I keep coming back to is not the technology. It’s the customer. A share of the people who would have typed “dentist near me open Saturday” into Google are now going to say it to an assistant, and then walk away from the phone. The assistant does the rest. Whatever it does next is the whole game for a local business, and almost nobody is playing it yet.
What an agent does when it shops
A person who lands on your website looks at the photos, skims the copy, forms an impression, and maybe calls. An agent does none of that. It has a goal (“table for four, Saturday, around eight, outdoors”) and a list of candidates. For each one it tries to answer a few questions: Are you open? Do you have what I need? What does it cost? Can I book it right now? Am I allowed to? If it can get a true answer it can act on, it acts. If it can’t, it moves to the next candidate. There is no “but the website was beautiful” tiebreak.
An agent never reads your homepage. It asks a question and takes the first true answer it can act on.
That changes what “findable” means. For twenty years it meant ranking: be on the first page, have the reviews, buy the ad. Ranking still matters, because agents start their search where people do. But ranking only gets you into the list. What gets you the booking is being answerable: hours the agent can confirm, a price it can quote, an availability it can check, a booking it can make, and rules it can trust. A business that ranks first and can’t be booked loses to the one that ranks fourth and can.
Four ways an agent can reach you
An agent reaches a business the same ways a person does, plus two ways only an agent can. I think of them as four doors, and it’s worth being honest about which ones are open today.
- It reads your website. Not the way a person does. It looks for structured data (the machine-readable block that says “these are our hours, this is our address, this is what a cleaning costs”), a file called
llms.txtthat describes the business in plain language, and pages it can quote without guessing. This door has been open since the first AI search products. Most local sites still have it shut, which is why assistants get hours and prices wrong so often. - It calls you. Google’s agents already do this. If the phone rings out to voicemail, the agent reports “couldn’t confirm” and the person picks somebody else. A phone that answers on the first ring, at 9pm, and can actually book, is worth more this year than it was last year, for a reason nobody planned.
- It calls your tools directly. This is the new one. There is now an open standard called MCP (Model Context Protocol) that lets an assistant call a piece of software the way it would call a calendar: “what’s open Friday afternoon?” and get back a real answer in milliseconds. A business can publish its own small server that answers those questions from its real booking system. Claude, ChatGPT and Gemini all speak it. Very few businesses have one.
- It talks to your agent. The customer’s assistant asks; your business’s assistant answers, holds the slot, applies the discount you allow, takes the deposit, and hands anything unusual to a person. This is the earliest of the four. The big assistants are still working out how they will pay and book on people’s behalf. But the side you control (your agent, your rules, your booking system) can be built now, and it’s the same work either way.
What actually changes at the front desk
Here is the practical version. When an agent asks about your business, the answer has to come from somewhere. Today it comes from wherever the agent happened to look. Most of the time that’s a Google listing an employee updated two years ago, a website with a PDF menu, and a phone that rang four times.
| The agent asks | Where the answer lives now | If it can’t get one |
|---|---|---|
| Are you open? | Google listing, website footer, a sign on the door. Often three different answers. | Assumes closed. Moves on. |
| Do you have it? | Someone’s head, a POS, a booking calendar behind a login. | Reports “unclear.” Moves on. |
| What does it cost? | A PDF, a “call for pricing” page, or nowhere. | Ranks you last. Moves on. |
| Can I book it now? | A phone, a form that emails somebody, a third-party widget. | Books the competitor with a live calendar. |
| Can I trust the answer? | Reviews, license numbers, a cancellation policy in fine print. | Prefers the business with verifiable facts. |
Look down the right-hand column. Every one of those outcomes is silent. You don’t get a missed-call notification when an agent skips you. You just get fewer Saturdays.
The other change is that rules matter more than copy. Persuasion doesn’t work on an agent, but limits do. Can it hold a table for ten minutes? Can it offer ten percent to a first visit? Does it need a deposit? What happens to a request it isn’t sure about? A business that has written those answers down can let a machine act on them safely. One that hasn’t will either say no to everything or have a bot promising things the owner never agreed to.
Five things I’d do this month
- Ask an assistant about your own business. Open ChatGPT, Claude or Gemini and ask five things a customer would: are you open Sunday, how much is X, can I book for Friday, do you do Y, where do I park. Write down what it gets wrong. That list is your homework, and it’s usually longer than people expect.
- Fix the facts once, everywhere. Hours, address, phone, the five most-asked prices. Same answer on your site, your Google listing, Apple Maps and in your structured data. Agents cross-check, and disagreement reads as unreliable.
- Make the phone answer. First ring, every hour, able to book. Whether that’s a person or an AI concierge, the voicemail door is the one you can close fastest.
- Write your rules. One page: what can be booked without a person, how long a hold lasts, the most anyone can be offered, when a deposit is required, who gets a text when something is booked. You need this whether the request comes from an agent, a new hire or a franchise partner.
- Choose tools with real APIs. If your booking system can’t be read or written by other software, no door can be opened on top of it. This is the same advice I gave in the routing piece, and agents make it urgent.
Where this goes
I don’t know how fast the fourth door opens. I do know that the businesses that answer agents well are going to get a share of customers out of proportion to their size, because for the first time a two-chair salon and a national chain are being compared on the same terms: can you answer, can you book, can I trust you. That’s a good deal for small businesses, if they show up for it.
It’s also why we packaged the four doors as AI Business. Two of them we’ve been building for clients for a while (a readable website, a phone that answers). The other two are new, and we’re building them on open standards so nothing depends on which assistant wins. The first step is the same one I suggested above: ask the assistants what they say about you today. We’ll do it for you, for free, and send you the transcript.
What do AI assistants say about your business right now?
We ask ChatGPT, Claude, Gemini and Perplexity five customer questions about your business and send you what they answered, what they got wrong, and the three fixes that matter most. Free, within 48 hours.