Most of the AI news this year has been about agents that work for hours. This week’s most useful development for New Zealand businesses is smaller and less exciting: a cheap, fast way to make one routine decision at a time, with a number attached that says how sure the AI is.

That number matters more than it sounds. It turns “should we let AI do this?” from a gut call into a setting a manager can choose, measure and adjust.

What changed

On 22 October 2026, OpenAI announced that its Decisions API is generally available, after a public beta that opened on 6 October. It’s a developer service, not a chat app, so it sits behind your own systems rather than in front of your staff.

In plain English, you send it some content (an email, a form, a photo, a block of order data) and up to 400 questions about it. According to OpenAI’s developer guide, each question is one of three types:

  • Yes or no: “Is this email a complaint?” The answer comes back as a probability that it’s true.
  • Pick one: “Which team should handle this: accounts, dispatch, service or other?” The answer is one of the options you supplied, with a probability for each.
  • Score: “How urgent is this, on a scale of 1 to 5?” The answer is a rating with a confidence figure.

The launch price is US$0.10 per million input tokens, and OpenAI says there are no output-token charges. As rough arithmetic, an email of around 1,000 tokens would cost a hundredth of a US cent to assess, before regional processing premiums and other multipliers the guide mentions. The GA release added prompt caching to cut the cost of reusing the same instructions, the ability to pass structured data directly, and images in individual questions, so a photo can be compared against a reference image.

Why it matters for how teams work

A lot of office work is triage. Someone reads each incoming item, decides what it is and where it goes, and passes it on. It’s repetitive, it’s interrupted by everything else, and it’s often done by your most experienced people because they know the rules.

General-purpose AI could already do this, but it tended to return prose that someone had to interpret, and it didn’t say how sure it was. A service that returns a fixed answer and a probability fits into a workflow much more cleanly. The workflow can act automatically when the AI is confident, and send everything else to a person.

That changes the job rather than removing it. The person who used to sort the whole queue now handles the uncertain cases, checks a sample of the automatic ones, and improves the rules. That’s a better use of their knowledge, but only if the change is planned.

Concrete ways NZ businesses could use it

  • Wholesale and distribution: classify incoming customer emails as order, change, query or complaint, flag anything that mentions a delivery problem, and route each to the right inbox.
  • Professional services: check incoming documents against a short list of questions (“Is this signed?”, “Is this for the current year?”) before anyone opens them.
  • Construction: score site photos against a reference image for an obvious issue, such as missing edge protection, and send low-confidence or flagged photos to the site manager.
  • Manufacturing: compare a product photo with a reference for labelling or packaging faults, and hold anything uncertain for a person to inspect.
  • Agriculture: sort supplier and contractor paperwork by type and urgency so the office deals with time-sensitive items first.
  • Finance teams: suggest the expense category for a supplier invoice from a fixed list, with anything below your chosen confidence going to a person.

In every case, the AI makes a narrow call from options you control. It doesn’t write to customers, move money or change records on its own.

Risks, costs and governance

Confidence isn’t correctness. A probability is the model’s estimate, not a guarantee. OpenAI’s own guide says to set thresholds using labelled examples from your application, and to base them on the cost of false positives and false negatives. A wrongly routed marketing email costs almost nothing. A wrongly approved invoice does not.

Privacy. Under the Privacy Act 2020, you’re responsible for personal information you send to a provider. The guide says regional processing is available in the United States and Europe, so NZ customer data will be processed offshore. Check the provider’s data controls and terms before you send personal information, and send only what the decision needs. Our article on the Privacy Act and AI tools covers what to check.

Security. Incoming emails and documents can contain hidden instructions. Keep decisions narrow, keep the options fixed, and never let a classification on its own trigger a payment or a data change. See our explainer on prompt injection.

Lock-in and change. At launch the service supports one model. Build the integration so the decision step can be swapped for another provider, or for simple rules, without rebuilding the workflow.

Human sign-off. Decide in writing which decisions are always reviewed by a person, whatever the confidence. Anything affecting a customer’s money, an employee or a legal obligation belongs on that list.

What managers should do differently

The confidence threshold is a business decision, not a technical one. Setting the auto-approve line at 95% rather than 80% is a trade-off between staff time and error risk, and it should be made by the person who owns the process, not left as a default in the code.

Managers should also own the measures. For each automated decision, track three numbers: the share handled automatically, the error rate on a regular sample, and how long uncertain items wait for a person. If errors rise, raise the threshold. If the review queue backs up, look at whether the categories are clear enough.

Finally, recognise the work of the people who review. Their corrections are what make the thresholds trustworthy over time.

Try this in the next 30 days

  1. Pick one queue where someone spends time sorting or routing: a shared inbox, a document intake, or a photo check.
  2. Write the decision as fixed options, including an “other” or “not sure” option.
  3. Collect 100 to 200 real past examples and label the correct answer for each.
  4. Run the examples through the decision step in a test environment, with personal information removed or minimised.
  5. Compare the AI’s answers with your labels at different confidence levels, and agree the threshold with the process owner.
  6. Run it in shadow mode for two weeks, where the AI suggests and people still decide, before letting anything act automatically.
  7. Write down the decisions that always need a person, and check the workflow enforces them.

Where to start

Start with one queue and one decision, not a platform. If you’d like help choosing the right process, setting thresholds and building the review step into your existing systems, see our AI and agents and integration and automation work, or get in touch.