Think about the small calls your office makes all day. Is this email a new job or an existing customer? Has this invoice been paid? Is this person allowed to get a text from us? Is this request urgent, or can it wait until morning?

Most of them are yes-or-no questions, or a pick from a short list. They are also where automation is safest and cheapest, as long as it is built for that kind of question and knows what to do when it is not sure.

A new kind of AI tool, built for yes and no

On October 6, OpenAI released a beta of a tool it calls the Decisions API. Instead of writing a reply, it answers a question with a likelihood that something is true, a pick from a fixed list of options, or a score, according to OpenAI’s own documentation. OpenAI says it is about ten times faster than asking its regular model the same question. That is the company’s own claim, and the tool is still in beta.

Other companies, such as Liquid AI, have released similar “decision models” too. You do not need to use any of them. The point for an owner is simpler: the AI industry is admitting that a lot of business automation is not about writing. It is about sorting.

Why “how sure” matters more than “yes or no”

A good sorting step does not just say yes or no. It says how sure it is. That one number lets you set up three lanes:

  • Sure yes. The automation acts on its own. File it, tag it, send the standard reply.
  • Sure no. The automation files it away, or ignores it.
  • Not sure. It goes to a person.

Where you draw those lines depends on what a mistake costs. Sending a “we got your request” email by mistake costs almost nothing. Approving a refund by mistake costs real money. So the refund question gets a much higher bar before the automation acts alone, and more of them land with a person.

A made-up example

Here is a made-up example. A pest control company gets about forty emails a day: new job requests, questions from existing customers, vendor invoices and junk. The owner sets up an automation that asks one question of each email: which of those four is it?

When the automation is sure, it sorts the email into the right folder and, for a new job request, drafts a reply for the office to send. When it is not sure, for example an email from an existing customer asking about a new property, it goes into a “check this” folder.

In the first week, the office checks every sorted email anyway and notes each mistake. By the end of the week, they know how often it is right in each lane, and they move the lines until the “check this” folder holds only the emails that genuinely need a person.

Four questions to ask whoever builds your automations

  1. Which steps are just sorting, and how are they built? A yes-or-no step does not need the same tool as writing a customer email. Ask what each step uses and why.
  2. What happens when it is not sure? There should be a clear lane that goes to a named person, not a guess.
  3. Can you see how sure it was? A record that shows the answer and how confident it was makes mistakes easy to trace.
  4. Has it been tested on your real messages? Ask for a test on a few hundred of your own emails or forms before it runs live, with the mistakes counted.

Where sorting stops

A sorting step can tell you a contract looks risky. It cannot write the note explaining why. Drafting replies, summarizing a long thread and working through a problem with several steps still need a different kind of tool, and usually a person checking the result. Even OpenAI’s own guide says that when one decision depends on another, they should be asked separately.

The pattern that works is plain: fast, simple sorting wherever the answer is a checkbox, a person wherever it is not sure, and the bigger tools saved for the writing.

Your one step this week: list five decisions your office makes over and over that could be answered with a checkbox. Next to each, write what a wrong answer would cost. The cheap-mistake ones are your first candidates.

If you want help working out which of yours to automate first, here is what we build.

Sources

Kush AI Automation builds in accounts you own, and our monitoring alerts when a run fails and when it never starts, so an automation that quietly stops still reaches a person. Bring the decision your office makes most and book a free call.