It is 9:40 on a Monday morning and your shop's WhatsApp holds 63 unread messages from the weekend. Some are bookings. A few are complaints. One is a supplier invoice, eleven are spam, and two are customers who were charged twice. Before anyone replies, someone has to sort the pile.
That sorting step is exactly the job a new kind of AI was built for in the past six weeks. It does not write replies. It reads each message and picks an answer from a list you give it, with a number showing how sure it is. This guide explains what these "decision models" are, how they work, what they cost, and where they fit in a small Hong Kong business.
What is an AI decision model?
An AI decision model is an AI that answers a fixed question by choosing from options you define, such as "booking, complaint, invoice or spam", and returns a probability for each option instead of writing text. It sorts, routes and checks things quickly and cheaply, then hands the result to a person or another system to act on.
Think of it as the difference between a receptionist who writes you a memo about every visitor and one who simply points each visitor to the right door. The second one is faster, cheaper and easier to check, because the only thing you need to review is which door they pointed to.
The category is very new. According to a running comparison by the support software company eesel, TypeSafe's Jev model kicked it off on 15 September 2026. By the third week of October, Cloudflare, Amazon's Strands Labs, Microsoft, Liquid AI, Perplexity and several smaller labs had all released their own decision models.
How does a decision model work?
A decision model takes two inputs: the thing to judge, such as a message or an order record, and a short list of questions with allowed answers. It reads everything once and scores every allowed answer at the same time, returning numbers like "billing 0.98" rather than sentences. Nothing is generated word by word, which is why it is fast.
A normal chatbot such as ChatGPT writes its answer one small piece of text at a time. That is useful for an email, but slow and wasteful when all you need is a label. A decision model skips the writing completely.
Cloudflare's documentation shows a typical example. The message "I was charged twice. Please refund the extra payment." went in with three questions attached. The model returned "billing" as the team with 0.981 probability, "refund needed: yes" at 0.996, and an urgency score. Your system then routes the message on those numbers.
The three question types you will usually see
--- Yes or no: "Is this message a complaint?" The model returns the probability that the answer is yes.
--- Pick one: "Which team should handle this: sales, accounts, operations or other?" The model returns a probability for each option.
--- Score: "How urgent is this, from 1 to 4?" The model returns a weighted score across your levels.
You describe the options in plain words inside each request. There is no lengthy training project with thousands of labelled examples before you can start, which is a big change from older classification tools.
Why did so many decision models launch in October 2026?
Businesses building AI assistants discovered that most of the steps inside them are small decisions, not writing. After TypeSafe launched Jev in mid-September, large providers raced to offer faster and cheaper versions. Cloudflare, AWS Strands Labs, Microsoft, Liquid AI and Perplexity all released decision models between 1 and 16 October 2026.
The release tracker at No Code MBA lists the main launches. Cloudflare released Clef and the smaller Clef-flash on 1 October as open models anyone can download. Amazon's Strands Labs released Strands Decider 2B the same day, small enough to run on an ordinary laptop. Liquid AI followed on 7 October, Microsoft-Decision-1 arrived on 9 October, and Perplexity updated its pplx-decider models on 16 October.
OpenAI also announced a Decisions API at its DevDay event on 29 September, but as of late October it had not published documentation or a price. That gap matters if you were hoping to use a familiar brand. For now, the most usable options come from cloud and infrastructure companies.
The speed claims explain the rush. In Cloudflare's own benchmarks reported by InfoQ, Clef-flash answered in a median of 38.8 milliseconds and the larger Clef in 209.3 milliseconds, against 524.1 milliseconds for Jev. A person blinks in roughly 100 to 400 milliseconds, so these answers come back faster than a blink.
How is a decision model different from ChatGPT?
ChatGPT is a generalist that writes, explains and reasons in free text. A decision model is a specialist that only picks from options you set. It is far faster and cheaper per task, and its answers are easier to check and count, but it cannot draft replies, explain itself in detail or solve complex problems that need step-by-step reasoning.
A useful comparison is a sorting machine at a post office versus a letter writer. You would never ask the sorting machine to compose a letter. You would also never pay a letter writer to read every envelope and drop it into the right bag.
Where each one wins
--- Decision model: sorting enquiries, flagging urgent cases, spotting spam, checking yes or no rules, scoring leads, and deciding whether a case needs a human.
--- Chatbot: writing replies, summarising long documents, answering open questions, and working through problems with several steps.
The benchmarks confirm the split. On GPQA Diamond, a hard reasoning test, Cloudflare reported Clef scoring 48.0 while Jev scored 78.3. Decision models are built for fast labels. When a job needs real thinking, a full chatbot model is still the right tool.
In practice the two work together. The decision model reads every incoming message in a fraction of a second and decides where it goes. Only the messages that need a written answer are passed to a chatbot or a person, which keeps the expensive step for the work that deserves it.
What can a decision model do for a small business?
For a small business, a decision model is most useful wherever staff repeatedly sort, tag or check incoming items. Typical jobs include routing WhatsApp and email enquiries, separating complaints from bookings, flagging urgent cases for a manager, scoring sales leads, and checking photos of damaged goods before a refund. The model labels; a person or system acts.
Five realistic Hong Kong examples
--- Restaurant group: every WhatsApp message is labelled as booking, catering enquiry, complaint, job application or spam. Complaints go straight to the manager on duty instead of waiting in the general queue.
--- Property agency: each new lead is scored from 1 to 4 on how ready the buyer seems, based on their message. Agents call the 4s first.
--- Online shop: when a customer sends a photo of a parcel, the model answers "visible damage: yes or no". Cloudflare's Clef accepts up to four images per request, so photo checks are possible.
--- Clinic or beauty centre: any message that mentions pain, bleeding or a reaction is flagged "urgent, needs a person", so it never sits unanswered behind routine appointment questions.
--- Trading company: incoming emails are tagged invoice, purchase order, shipping update or other, then saved into the right folder.
None of these require the AI to speak to a customer. That is the appeal for cautious owners: the model makes a small, checkable choice behind the scenes, and every customer-facing word can still come from your own team.
If your business already uses several AI models, a decision model can also act as a traffic controller, sending simple jobs to a cheap model and hard ones to a stronger one. Our guide to LLM model routing explains that idea in more detail.
How much does an AI decision model cost?
Decision models are among the cheapest AI tools available. Microsoft-Decision-1 charges about US$0.042 per million input tokens with free output, Cloudflare's Clef-flash costs US$0.09 and Clef US$0.24 per million, and AWS's Strands Decider 2B is free to run yourself. Sorting 10,000 short messages typically costs well under HK$5.
A "token" is a small piece of text, roughly three quarters of an English word. A typical customer message plus your list of options is around 500 tokens, so 10,000 messages add up to about five million tokens.
What 10,000 sorted messages cost (input only, HK$7.8 to US$1)
--- Microsoft-Decision-1: about US$0.21, or roughly HK$1.60.
--- Cloudflare Clef-flash: about US$0.45, or roughly HK$3.50.
--- Cloudflare Clef: about US$1.20, or roughly HK$9.40.
--- Strands Decider 2B: no usage fee, but you supply and maintain the computer it runs on.
Prices come from Microsoft's published rate as reported by eesel and from Cloudflare's Workers AI pricing. Cloudflare also gives every account 10,000 free "neurons" a day, which eesel estimates covers roughly 900 Clef or 2,400 Clef-flash decisions daily. The real cost for a small business is therefore setup time, not usage. Someone still has to connect the model to WhatsApp, email or your order system.
What do people get wrong about decision models?
The most common mistakes are assuming a decision model is smarter than a chatbot, treating its confidence number as the chance it is correct, and letting it make serious decisions about people without review. It is a fast sorter, not a judge. Every option list also needs an "other" choice, because the model cannot say "I do not know" on its own.
Misconception 1: "It is a better ChatGPT."
It is not. It is narrower. It wins on speed and cost for labelling, and loses badly on reasoning, as the GPQA Diamond scores above show.
Misconception 2: "0.9 means it is 90% likely to be right."
Not always. Cloudflare's documentation states that the confidence value shows how concentrated the probabilities are, and "isn't the chance that the answer is right." You only learn real accuracy by testing on your own past messages.
Misconception 3: "Bigger is always better."
Cloudflare's results show the small Clef-flash matching Clef on a customer service test, 77.0 against 76.3. However, on a test with 150 categories plus messages that fit none of them, Clef-flash dropped to 66.8 while Clef held 97.4. Short lists suit small models; long or messy lists need the larger one.
Misconception 4: "It can decide anything."
Microsoft warns that Decision-1 should not be the sole basis for high-impact decisions about people, such as employment or credit. Screening job applicants or approving payment terms should always keep a person in the loop. Our explainer on human-in-the-loop AI shows how to design that check.
One Hacker News commenter summed up why honest confidence matters: "A 96% model with overconfident outputs is operationally worse than a 94% model with honest ones." In plain terms, a sorter that knows when it is unsure is more valuable than one that is always certain.
How do you start using a decision model safely?
Start with one repetitive sorting job, write four or five plain-language options including "other", and test the model on 200 to 500 of your past messages before going live. Let it act alone only above a high confidence level, such as 0.9, send everything else to a person, and keep a log so you can measure mistakes every month.
A five-step starter plan
--- Pick one job: choose the sorting task your team repeats most often, such as tagging the morning WhatsApp backlog.
--- Write clear options: describe each label in one sentence, for example "complaint: the customer is unhappy with food, service or a charge". Always add "other".
--- Test on history: run a few hundred old messages through the model and compare its labels with what your staff actually did.
--- Set a threshold: let the model route automatically only when it is very sure. Below that line, a person decides.
--- Review monthly: check a sample of decisions, especially the ones customers complained about, and adjust your option descriptions.
Most of these steps need a developer or an IT partner the first time, because decision models are sold as building blocks rather than ready-made apps. The good news is that the building blocks are cheap, so a small pilot can be tested without a large budget.
The bottom line on AI decision models
A decision model is the AI version of a fast, tireless sorter. It will not write your emails or replace your judgement, but it can read every incoming message in milliseconds, decide where it belongs, and tell you how sure it is, for a few Hong Kong dollars per ten thousand messages.
For a small business, the smartest first step is not a big AI project. It is choosing one sorting job your team repeats every day, testing a decision model against your own history, and keeping a person in charge of anything that really matters. We understand AI. UD stands with you.
Reviewed by the UD AI team.
Wondering which sorting jobs in your business could be handed to a decision model, and which should stay with your team? Take the free AI Ready Check and we will walk you through it step by step, from spotting your most repetitive decisions to planning a safe first pilot.