
Breakdown / AI agents · Business · AI explained
When AI chooses which business gets your money.
An assistant can save you the search. It also gets a say in which restaurant, shop or service wins your business.
Imagine asking your assistant to find dinner for four under $80. It checks what's nearby and comes back with a cart. You review the order and approve it.
You've saved yourself a search. But why did it pick that restaurant?
DoorDash is already moving in this direction. On September 30, it announced a US beta where people can order by text. It says the service searches local menus, suggests a cart and handles checkout after you confirm in the thread. There's a waitlist to try it. DoorDash's announcement.
You still approve the purchase. The assistant gets a say in which business reaches you in the first place.
What happens when the recommendations stop?
Lukas Campos runs Audioscrape, a service for searching audio and turning speech into text. In a September 30 essay, he says ChatGPT began suggesting his service in conversations and signups surged. Then the suggestions stopped. People could still find Audioscrape in the directory, but ChatGPT had stopped bringing it up. He calls this the “router economy”: the assistant sends your request to a business, and that business gets a chance to win a customer. Campos's account.
We haven't checked his signup figures independently. Other developers describe a similar problem in an OpenAI forum thread: their tools were still listed, but ChatGPT wasn't finding them in conversations. They disagree about the cause, and we haven't found a public explanation from OpenAI that settles it.
For a small business, that's an uncomfortable position. Your service can be working perfectly while the assistant that brought you customers stops mentioning it. You need to know what changed before you can do anything about it.
How does an assistant know what to use?
Suppose you run a document translation service. An assistant needs to know which languages and files you handle before it can tell whether your service fits someone's request.
OpenAI's documentation says ChatGPT and Codex use a tool's name, description and the information it accepts to decide when to use it. It recommends checking whether the assistant finds your tool when someone names it, when they describe the task without naming it, and whether it leaves your tool alone when something else would be a better fit. OpenAI's guidance.
That helps explain how a tool gets chosen. It doesn't tell us how competing businesses are ranked. And a clear description can't make a bad translation good. Getting picked brings someone to your service; the result determines whether they'll want to use it again.
People still want to browse
You might happily hand over a repeat grocery order. A holiday is different. You may want to look through photos, check the map and send a few places to friends before deciding.
Airbnb's Brian Chesky wants the company to work better with agents while keeping the browsing and shared planning people enjoy. TechCrunch interview.
Why this one?
Back to that dinner order. Did the assistant choose the restaurant because it was cheapest, arrived fastest or had the best options for four people? Those are different reasons to spend $80.
We'd want it to explain the choice, show alternatives and give us the full price before we pay. If a recommendation is sponsored, say so where we can see it.
The next time an assistant picks a restaurant, hotel or service for you, ask what else it considered. Saving you a search shouldn't mean leaving you guessing why that business got your money.
Sources
- Text DoorDashDoorDash
- The Router EconomyLukas Campos
- Plugin discovery reportsOpenAI Developer Community
- Optimize MetadataOpenAI
- Brian Chesky interviewTechCrunch