Only rich people had drivers. Now we have Uber.

Only rich people had holiday houses. Now we have Airbnb.

Only rich people had a personal assistant. Now there’s Muse and Instinct.

This week is in 4 parts. What these AI assistants are, how they decide what to buy, whether they can even get into your website, and how to put one to work for yourself.

Meta launched Muse on 8 September. It’s an AI assistant that does things for you. Sends emails, books travel, fills in forms, buys stuff.

It hit 6.6 million installs in its first month.

Instinct is the other one. You just text it. It’s invite-only, it has spent $0 on marketing, and people are selling invites on eBay. It’s handling close to US$1 billion a year in transactions, about half of it travel. Investors just valued it at US$10 billion.

Muse isn’t in Australia yet. Meta says US and Canada only, no date.

But I’ve been using Instinct here in Sydney. It has my email and calendar, keeps a rolling to-do list and texts me morning and night asking where things are up to. It’s the first AI tool I’ve used that chases me, not the other way round.

And the traffic is already here. Cloudflare says requests from AI agents grew more than 1,700% in a year. More than half of all internet traffic isn’t human anymore.

Takeaway: your next customer might not visit your website. Their assistant will.

When a person books a hotel, they click a listing or two and often walk away. A McGill University study of hotel-booking agents found agents looked at 2 to 6 listings per session. People looked at about 1.

Every single agent session ended in a booking. A third of the people who clicked didn’t book.

What did the agents pick? High review scores and low prices. That’s it.

Microsoft built a fake marketplace to test the same thing. 80 to 100% of agents took the first offer they got. Their summary: “speed beats quality.” Giving agents 100 options instead of 3 made things worse, not better.

Under the hood, a lot of these calls are made by a new kind of model that doesn’t write anything. It just answers a question with a label. Is this the cheapest? Is it available? Does it match? TypeSafe’s Jev does that for about 42 US cents per 10,000 decisions. A hiring marketplace called Jack & Jill swapped to it and, by TypeSafe’s numbers, cut its screening cost 88%.

It’s already moving money here. Google says Kmart and Officeworks’ own shopping agents lifted conversion up to 3 times.

It also goes wrong. A YouTuber put Muse on “Allow Always” to sell something on Facebook Marketplace, with a US$700 floor. It accepted US$600 and sent the buyer his home address. Meta blamed a display bug and promised clearer permission prompts.

And Amazon has started blocking Muse from shopping on its site. It says Muse doesn’t identify itself. Meta disagrees.

Takeaway: agents shortlist on reviews, price and clear information, and they move fast. If they can’t read you, you’re not on the list.

This week I knocked on 22 of Australia’s best-known websites the way an AI assistant would, labelled as one.

13 turned me away or left me hanging.

Big caveat: real ChatGPT traffic comes from OpenAI’s own servers, and some of these sites let that through. A brand-new assistant, or a small business’s own tool, gets what I got.

Here’s the weird bit. Only 2 of the 22 say anything about AI agents in their robots.txt, the file that tells bots what they can and can’t do. The blocking happens further down, in bot protection most owners never look at.

Kmart turned away the request labelled as an AI assistant, but a plain request with no label got the full page. Airbnb’s own rules say AI assistants are welcome. Its bot protection still said no.

JB Hi-Fi and The Good Guys went the other way. Both have set up 13 tools an AI agent can use to search the range, fill a cart and check out. (JB’s homepage is blank to anything that doesn’t run JavaScript. Its checkout tools aren’t.)

And CommBank wrote the bots a welcome note.

Then I checked 54 Australian accounting and bookkeeping websites.

Good news first. None of them block AI in their robots.txt, and only 3 turned my request away. The door’s wide open.

The problem is what’s inside.

35% have no structured data at all. Those are the hidden labels that tell a machine “this is an accounting firm, here’s the address, here are the hours.” Only 39% have opening hours an agent can read. 6% show any fees.

Now picture someone’s assistant asked to “find a bookkeeper near Parramatta who does Xero and is open Saturdays.” It can’t shortlist you on things it can’t read.

Check yours. The checker is live. Type your website and in about 20 seconds you’ll see whether AI agents can get in, what they can read, and whether ChatGPT, Claude and Gemini name you when someone asks for a firm like yours.

For your web person, here’s the one-line version:

If the first line says 200, an agent like this gets in. A 403, or nothing for 20 seconds, means it’s turned away.

Takeaway: for a small firm the fix isn’t technical. Put your services, area, hours and a starting price somewhere a machine can read them.

The classic jobs:

The rule I’d set on day 1: “ask first” for money, messages and personal details. Never “Allow Always”. The Muse Marketplace sale is exactly why.

Two more. Give it a card with a low limit, not your main one. And read the terms. Instinct’s let it enter “agreements, commitments, or transactions” for you, and its first version claimed a “perpetual and irrevocable” licence over your data. That was walked back within days, but it tells you where the defaults start.

What didn’t work. Too many tools gets chaotic. I’d rather one assistant doing 5 jobs than 5 assistants. And the scheduled-task features in the big chatbots are still flaky. Instinct’s check-ins are the first ones that reliably turn up.

Takeaway: start with the boring stuff. A to-do list that chases you and an inbox that sorts itself. Leave the credit card on “ask first”.

If you joined from one of our videos, this is the one we promised. The video, the setup guide and every prompt, in one place.