Decisions API, hallucination detectors and agent memory: AI decides, a person confirms

OpenAI made the Decisions API generally available: fast answers to questions with a limited set of options, yes or no, one from a list, a score on a scale. IBM packed the same kind of decisions together with chat and speech recognition into one model, and n8n published three pieces in two days on how to catch a model's fabrications and where to keep its memory. For a business owner it comes down to one question: which decisions can be trusted to a model, and which must an employee confirm.

What came out

On October 22, OpenAI announced that the Decisions API is out of beta. The developer sets a question and the allowed answers, passes text, an image or JSON, and gets one of three response types: a probability from 0 to 1, a choice from a list with a confidence score, or a rating on a scale. OpenAI says the answer arrives up to 10 times faster than from the same GPT-6 Luna model through the regular Responses API. Box classifies documents and routes them onward: on 30 insurance certificates the route was chosen about 5 times faster at the same accuracy, and low-confidence documents go to people for review. Sierra picks the next action in a live customer conversation in under 280 ms and pays 90% less per call.

On October 15, IBM showed Granite Switch 4.1 at 8B and 30B parameters on the Hugging Face blog. The idea is different. Instead of a separate service, adapters sit on top of one base model: a decisions adapter, a Guardian adapter for harm checks and a built-in Turbo CTC speech model take up about as much GPU memory as the base alone. The adapters share the KV cache, so a decision after the model's answer costs almost nothing. For anyone hosting a model in-house, this is a direct answer to the Decisions API.

On October 22, n8n broke down hallucination detection methods in agent workflows. It lists four types of errors: an invented fact, an answer that ignores the supplied context, made-up details and self-contradiction. And five ways to catch them, from checking grounding in the context to confidence thresholds with handoff to a person. The line that mattered most to me: the threshold should reflect the cost of a mistake. An internal summary will survive an inaccuracy; a change to a customer record won't.

The same day, n8n compared stateful and stateless approaches in agents: the model itself remembers nothing between requests, so state lives outside, in a database or cache, and the request handlers stay stateless. The next day, October 23, came a guest piece on agent memory by Veljko Kovac of Cognee: short-term memory within the current session, long-term memory in a separate store. RAG reads and never writes back; agent memory begins where the agent writes.

Why this matters to a business

Take a car rental company. An email arrives with a photo and the text "here's the car after the trip." The first decisions are small: is this a return, a complaint or a payment question; is there visible damage in the photo; how serious is it on a scale from scratch to dent. The Decisions API and models like it were made for exactly these questions: the answer is cheap and fast, and it can go straight into a field on the request. But the next decision, whether to charge the client's deposit, costs money and the relationship. An employee confirms that one: the model prepares the request, the person presses the button.

So the rule is: split decisions by the cost of a mistake. Classification and routing run automatically with a confidence threshold. Anything that changes money, a contract or a customer record goes through confirmation.

How I have it set up

In the document parsers for a PDR company in the US, I arrived at this in practice. They read insurance estimates, reports and photos with VINs. I started building them in 2023, and they've been in use since mid-2023: back then about half of the documents went through without manual correction. In 2024 I moved to new recognition methods, and about 75% went through without correction. In 2025 I added new data and calculations, and the share grew to roughly 95%. In 2026 I connected AI validators and shipped the release version: about 99.9% of documents go through without manual correction.

There are three validators, on Anthropic, OpenAI and Gemini services, each for its own task. How the checks work inside is something I don't disclose. The principle is the one n8n describes: several independent checks instead of one score, with the last word going to the employee who sees the result and confirms it. The parsers free up about 60 hours of manual work a week. I haven't tried the Decisions API myself; I'd test it on routing incoming documents, where you need a "what is this and where does it go" answer before the main parsing.

What I would do

I'd write out every small decision in the process: document type, urgency, who it goes to. For each one I'd define the allowed answers. If the answers can't be listed, it's not a task for a decision model.

For each decision I'd name the cost of a mistake. Cheap mistakes run automatically, expensive ones go through confirmation. I'd set the confidence threshold by that cost, not by average quality, and I'd log every decision together with its input and confidence. After a month the log shows where the model gets it wrong and which way to move the threshold.

I'd keep state outside the model: in a database, with the request number and the history of steps. Then the employee can see what the model decided and why.

I'd start with one type of decision and one queue for human review. I'd add the rest once the first queue stops surprising me.

Related reading

On verifying model output: the article AI in data processing: extraction and verification and the page AI in a process where the result can be verified. The parser case: a PDR company in the US saves about 60 hours a week. From the digests on the same topic: document data extraction in 2026, what AI remembered about a customer and a model on top of your database. Earlier I wrote about the reliability of AI agents and control over the process and about how to test an AI agent. If your process has decisions an employee makes by a template, tell me which ones. Discuss your task.