Something happened this fall that most dentists missed, because it didn't happen in dentistry, but still might change how you practice forever.

In the space of a few months, Google, Meta, OpenAI and Anthropic each launched a personal AI agent. Not a chatbot you ask questions to. An assistant that works in the background, on its own computer in the cloud, connected to your email, your calendar and your apps, and that keeps going after you close your phone.

I want to be clear upfront: I'm not telling you to plug one into your practice tomorrow. For anything touching patient data, they're not ready yet.

But I'm convinced this is the biggest change coming to how practices run. And the dentists who prepare now will be the first to benefit when it's ready.

Here's what came out, what each one is good at, and a plan for the next 6 and 18 months.

Quick one before we dive in.

Picture this: your patient leaves the consultation with one personal link. On it, they see their own teeth in 3D, every appointment in order, and one clear price they can accept and pay online, upfront or in installments.

And it keeps working once they say yes. Throughout treatment, the same link shows their next appointment (they confirm in one tap), your notes for each visit, what they've paid and what's left. Everything in one place.

That's Aria: the bridge between you and your patient, from the first quote to the last visit. It's in free beta in the US.

What your patients will see with Aria

See exactly what your patient receives →

The five personal AI assistants launched in 2026

Gemini Spark (Google) Announced May 19 at Google I/O, it runs 24/7 on Google's cloud and works across Gmail, Docs and Slides. Those connections are off by default, you turn them on one by one, and it asks for your approval before important actions. Best if your life already runs on Google.

Muse (Meta) Launched September 8. You connect your apps one at a time (email, calendar, payments and more), and when a service has no direct connection, Muse uses a web browser to get it done. There's a free tier, then $20 or $100 a month. One thing to know: Meta uses your queries to train its models by default. You can turn that off. The most accessible of the five.

Dots (OpenAI) Launched September 29 inside ChatGPT. Each dot has its own computer, browser and memory, and can reach more than 4,000 apps. It comes with built-in rules about when to act alone and when to ask you first, plus a read-only mode. For now it's reserved for the higher-priced plans. OpenAI already has a beta for healthcare workspaces, which tells you where this is heading. The most powerful, and the most expensive.

Claude Duet (Anthropic) Launched October 21, the newest of the five. It runs on its own virtual machine in the cloud, with its own browser, so it keeps working after you close the app. Every morning it sends you a briefing drawn from your connected apps (Gmail, Slack, Google Calendar), then watches them during the day for anything that needs action. It asks for explicit confirmation before sending messages, making payments, signing documents or deleting files. For now it's a waitlisted research preview in the US, starting with Max subscribers. The newest, and still in preview.

OpenClaw The open-source one that went viral this year (you may have seen it called Clawdbot). It runs on your own machine and you talk to it through WhatsApp or Telegram. Extremely powerful, and also the clearest example of the risks: researchers found a vulnerability allowing remote code execution with a single click, and hundreds of malicious add-ons were found in its skills marketplace. Fascinating to watch. Never on a computer that touches your practice.

Think about your week. How much of it is actually dentistry?

Confirming appointments. Chasing unscheduled treatment. Insurance verification. Answering the same five questions about pricing and parking. Following up with the patient who said "I'll think about it." Managing schedules, suppliers, payroll questions, reviews.

The ADA's own research says it plainly: practices are fighting staffing shortages, and dentists are spending a growing share of their week on non-clinical work.

This is exactly the kind of work these agents are being built for. Rules, follow-ups, coordination, memory.

Here's the part I find interesting.

Most professions are worried AI will take away the core of their job. Dentistry is one of the rare cases where it's the opposite. The core of your job is physical, clinical and human: your hands, your judgment, the patient in the chair who's scared and needs to trust you. No agent does that.

What agents can take is everything around it. Which means a dentist in 2028 may spend more of the day being a dentist than a dentist does today.

What agents can take, and what stays yours

(Getting new patients is another story, and a topic for another email.)

These agents are built for consumers, and none of them was designed with HIPAA in mind.

Under HIPAA, any tool that handles patient data needs a signed Business Associate Agreement. And an agent that reads your email and fills forms on its own is a new kind of risk: it can be tricked by a malicious email, click something it shouldn't, or send something to the wrong person.

That will get solved. Healthcare versions with BAAs, audit trails and encryption are clearly coming. Until then:

No patient data. No practice email. No PMS access.

What is safe today, and what has to wait

Your preparation timeline: now, 6 months, 18 months

Pick one assistant and give it your personal admin: your calendar, travel, family logistics, the bills you always forget.

The goal isn't productivity. It's learning how to delegate to an AI: how precise you need to be, where it makes mistakes, what you want it to ask you before acting. That skill transfers.

This is the most valuable thing you can do, and it costs nothing. An agent can only follow a process that exists somewhere other than in your head, or your office manager's head.

Start with the tasks that eat the most time:

And for each one, the most important line: what must always go to a human. An anxious patient, a complaint, anything clinical.

One process, written so an assistant could follow it

List the software you use (PMS, phones, imaging, scheduling, payments). For each one, ask the vendor two questions:

The practices with clean data and connected tools will be able to switch on agents first. The ones with everything locked in old software will wait.

When a vendor offers a healthcare-grade version, don't hand over the whole practice. Pick one workflow, like reactivating patients who haven't been seen in 18 months.

Start in draft mode, where the agent prepares and a human sends. Expand only once you trust it.

Open ChatGPT, Claude or Gemini and paste this (no patient names, no identifiers):

Fifteen minutes per process. Do one a week and in six months you'll have something most practices won't have: a practice that can be explained.

That's it for today.

I'd love to know which task eats the most of your week. Hit reply and tell me. I read and answer every reply myself, and the answers will shape the next emails.

Salim

P.S. Aria, the treatment sequencing tool we're building, is still in free beta worldwide. Try it here.