When AI agents talk to each other to book your appointments

Tuesday, 11:40 pm. A message lands on your business WhatsApp: "Hi, I'm looking for a check-up appointment on Thursday afternoon, ideally after 5 pm. If nothing's free, first thing Friday works too." Your agent replies with two slots, the other side picks one, asks for an email confirmation and signs off with "thanks". Forty seconds, start to finish. The odd part isn't the hour or the speed. It's that no human typed a word of that exchange: your customer asked their assistant to find them an appointment, and their assistant talked to yours. If you run any business with a diary, this is coming sooner than you'd think, and it's worth knowing what actually changes and what doesn't.

It's no longer science fiction, but it's not what you've been told either

Two programs settling something between themselves sounds like the distant future, and partly it is. In June 2025 the Linux Foundation launched the Agent2Agent (A2A) project, an open standard originally built by Google so that agents from different vendors can find each other, describe what they can do and work together on a task. Behind it are Amazon Web Services, Cisco, Google, Microsoft, Salesforce, SAP and ServiceNow, which is more or less everyone who sells software to big companies.

That's the "official" version: agents talking over a technical channel designed for the purpose. For a small business it will take a while to arrive. What's already happening is far more mundane: your customer's assistant writes to you through the same channels a person would. WhatsApp, your website form, email. No new protocol needed; it only requires the customer's assistant to be able to compose a message, and they all can.

So the useful question isn't "when will A2A reach my business?" but "what happens when the thing writing to me isn't a person, and I can't tell?"

What changes when the writer doesn't read your tone

When a customer messages you, they read your reply: the friendly tone, the "no problem at all", the emoji your salon always uses. An assistant reads none of that. It reads data: which slots exist, how long they take, what they cost, what the conditions are. All the warmth you put into your agent's welcome message counts for little; what counts is whether the answers are clear, complete and consistent with each other.

Three things show up immediately:

• Messages arrive complete, at any hour. No "hi" followed by silence. They come with date, time window, reason and a fallback in a single sentence, at 11:40 pm or 6:10 am. If your agent needs three questions to extract the same information, the thread stalls.
• Nobody resolves ambiguity any more. With a person, "mid-afternoon" gets negotiated. An assistant takes "mid-afternoon" to mean whatever its owner told it, which may not match what your agent assumes. If your agent doesn't pin down actual times, you get an agreement that isn't one.
• Cheap to book means cheap to cancel. If it costs your customer zero effort to request a slot, it also costs zero effort to cancel it or move it three times. More volume in almost always brings more churn with it.

Put differently: your rules start to matter more than your prose. And most small-business agents have plenty of prose and very few rules.

What your agent needs in writing before this happens

When the other party is a program, it won't forgive what a person would: the "we'll sort that out later". Three things need to be nailed down, and they're the same three you'd tell a new receptionist on day one.

What it can promise on its own. Slots that are actually free in the diary, the standard duration of each service, the list price. Nothing else. No discounts, no "I'll squeeze you in after closing", no "I'll book two services back to back even though they don't fit". If someone's assistant asks for 10% off for booking two people, the right answer from your agent is "a person needs to look at that", not "sure".

What it needs from a real person. At minimum, the end customer's name and phone number, not the assistant's. And if you're going to send reminders or store health details, that person's consent, not the say-so of a program claiming to represent them. This isn't red tape: it's the difference between having an appointment with someone and having an appointment with nobody.

When to stop. A persistent assistant can propose twelve alternatives in a minute, or ask the same thing again in different words. Give it a cap: after four or five back-and-forths without closing, the agent stops, saves the thread and hands it to a person with a summary. It never sits there "negotiating".

A practical tip: ask someone you trust to message your agent the way an assistant would, with date, window, fallback and a request for a discount tacked on the end. What it replies tells you more than any demo.

The risk isn't that they get it wrong: it's that they agree too fast

When you picture two agents talking, the natural fear is that they'll get tangled up. The real problem is the opposite: both are built to close, and they close. Two agents that want to "finish the task" can agree in twenty seconds on something neither you nor your customer wanted.

Examples we already see: a one-hour appointment at 8:15 pm when you close at 9 (it fits in the diary; it doesn't fit in the life of whoever has to lock up). A "yes" to the booking being "for two people" when you'd only set aside one slot. A brand-new contact being told about "Maria's" existing appointment because the assistant said it was writing on her behalf and nobody checked that the number matches the one on file.

That's why the principle we keep coming back to, you approve, they execute, stops being a slogan here and becomes a guardrail. The approval point sits right before anything that can't be undone: confirming a booking that pushes past closing, sharing details about an existing customer, accepting any condition that isn't on your price list. Everything before that (checking availability, proposing, asking for details) it should do alone; that's exactly the part you want to delegate.

And one specific precaution about identity: a message opening with "I'm the assistant for…" proves nothing. Your agent should treat it as a polite stranger: it can book a fresh slot with a name and number, but it cannot look up, move or cancel anything belonging to an existing customer unless that customer confirms it from the number you already have.

When you don't need to worry about any of this

If your customers are local, walk in off the street or phone you, and you get a handful of messages a week, don't build anything special. Personal assistants that book on their owner's behalf are still a minority, and in plenty of sectors they'll stay that way for a good while. Getting your business "agent-ready" before you have the problem is a way of losing hours, which is precisely what we're trying to avoid.

Don't buy anything because it carries an "A2A-compatible" label either. For a small business today, that compatibility doesn't bring you a single extra customer. What does bring customers is an agent that answers clear, structured messages well, and you need that regardless of whether the writer is a program or a person in a hurry.

The good news is that almost everything required is stuff you should have anyway: real opening hours in the diary, list prices written down, three rules about what's not accepted and a clear point where the agent hands the conversation to you. If that's in place, someone else's assistant is simply a very tidy customer.

What to do this week

1. Write down, on half a page, what your agent can promise alone and what it can't. No flourishes.
2. Check that it asks for the end customer's name and phone number before confirming anything.
3. Give it a cap on back-and-forths and decide who gets notified when it hits it.
4. Run the "assistant message" test and read the whole conversation, not just the ending.

If you take one thing away: when the other party is a program, your written rules are your tone. Our receptionist agent is built on that idea: it proposes and books within the limits you set, asks for the real person's details and passes the thread to you the moment anything strays from your price list. The rest, they can sort out between themselves.