AI decision models explained: choices and checks
Follow a support request through rules, model classification and application checks. Learn what decision models such as Jev do, and what they do not prove.
An AI decision model returns a bounded choice or assessment that software can use. The application still needs to check whether the choice is justified, request missing context and confirm permission before acting.
In an illustrative support inbox, a message says that downloading a report stopped working. A model suggests the technical-support queue instead of writing a paragraph. The application decides whether to route it, ask a question or request review.
Selecting the technical-support queue has not repaired the download, explained its cause or given anyone access to the report. Those are separate jobs. Keeping them separate makes it easier to decide which part needs a model and which part belongs in ordinary code.
Separate choosing from writing and doing#
Here, decision model means a model used to return a bounded choice or assessment for a program. It does not mean that classification is new, or that other language models cannot make such choices. The useful distinction is the job the application asks the model to perform.
TypeSafe documents Jev as accepting supplied context and typed questions rather than generating free-form text. Its interfaces include choosing among options, scoring against levels and assessing a statement. Those are Jev-specific interfaces, not a standard every classifier implements.
For our support example, choosing a queue is one decision. Drafting a reply is a writing task. Moving the request into a queue is an application action. A useful design may combine all three, but the output of the first is not evidence that the other two happened.
Say what each option means#
In this illustrative inbox, define technical_support as help with a feature that fails, account_admin as a request to change account details, and other as a request outside those categories. Add needs_context for messages that cannot be assigned from the available information. These are proposed application categories, not Jev defaults.
“The report download fails after I press Export” provides more routing evidence than “It stopped working.” The latter does not even identify the feature. Making the model choose between only the first two queues hides that missing information; it does not resolve it.
Jev’s Choice interface associates options with written criteria. Treat those criteria as part of the application contract. A returned name is meaningful only when the program and model were given the same definitions. Do not silently change what other means between versions.
One message may need several teams. A single primary-queue choice is appropriate only when the workflow has an agreed rule for that situation. Otherwise, ask separate questions or let a person resolve the assignment. Designing testable decisions develops this question-and-context contract.
Compare a rule, a classifier and a writer#
Use the same support workflow to compare the alternatives. The choice below follows what information is already known, not how fashionable a model is.
| Situation | Take | Because |
|---|---|---|
| A maintained ownership table maps the submitted component ID to one team. | Use the lookup | The application already has the exact rule and input. Interpretation adds no missing information. |
| Only a free-text symptom is supplied, and several queues are plausible. | Assess the message | A classifier can propose a category. Missing context or competing interpretations still need a defined response. |
| The team needs a helpful explanation tailored to the customer. | Write a reply | Selecting a queue does not produce the explanation. A separate writing step needs the relevant facts. |
- Situation
- A maintained ownership table maps the submitted component ID to one team.
- Take
- Use the lookup
- Because
- The application already has the exact rule and input. Interpretation adds no missing information.
- Situation
- Only a free-text symptom is supplied, and several queues are plausible.
- Take
- Assess the message
- Because
- A classifier can propose a category. Missing context or competing interpretations still need a defined response.
- Situation
- The team needs a helpful explanation tailored to the customer.
- Take
- Write a reply
- Because
- Selecting a queue does not produce the explanation. A separate writing step needs the relevant facts.
A generative model can also return structured data. OpenAI’s structured-output documentation provides one example of schema-constrained responses and warns that structured outputs can still contain mistakes. The comparison is not typed decisions against necessarily unparseable prose.
Start with the distinction between explicit rules and interpretation. Consider a specialised decision model when its interface fits the remaining task, then compare it with a suitable baseline on your own cases.
Compare the clear report-download symptom with the message “It stopped working.” Predict whether the latter supplies enough information to choose a queue. In the activity, choose Unclear, then switch back to Download.
Try it: change the kind of support request
The message contains a known billing code that is already mapped. Adding a model would not make this lookup more informative.
Invented inbox cases; no model is called or accuracy measured.
JavaScript is needed to change the case. The billing example remains visible above. A failed report download supports a technical-support proposal; a written explanation needs a writer. For “It stopped working”, the illustrative result is needs_context: ask which feature or action failed. No queue is selected and no action is taken.
needs_context preserves the missing information as an explicit outcome. The application’s next step is to ask which feature failed and what the person was trying to do, then assess the fuller message before choosing a queue.
Make the application handle the result#
For a proposed queue change, preserve the request identifier, input revision, option definitions and model result. Check that the queue still exists and that this workflow may move this request. Those checks are our proposed application design; a model choosing the right string does not perform them.
Keep a service timeout separate from other. The first means no usable classification arrived. The second is a classification result. Turning both into the same value makes an outage look like a change in what customers are asking about.
Likewise, a reported confidence value is not permission. Confidence and calibration explains how to evaluate such signals. Type-safe output versus correctness separates a valid value from a correct and authorised action.
Test the decision, not only the format#
Save clear requests, ambiguous messages, unrelated requests and messages with more than one intent. Write down acceptable outcomes before comparing models. Check queue assignment, unnecessary referrals and missed requests separately. A system that sends everything to a person avoids some automatic mistakes but may not accomplish the intended routing work.
Use repeatable evaluation cases to compare the entire routing step, including failures and follow-up work. Measure time and cost rather than inheriting a vendor’s performance claim. A decision model earns its place when the complete workflow benefits under unchanged requirements.
Questions people also ask
2 QUESTIONSCan a generative language model return a structured choice too?
Yes. Supported structured-output modes can constrain a generative model's response to a schema. Compare actual task quality, latency and complete cost rather than assuming that one model category is always preferable.
Does a decision model summarise a book?
Choosing a category or assessing a supplied passage is different from writing a summary. A pipeline may use decisions to select material for a separate writing stage, but that stage cannot recover important passages which were wrongly excluded and never supplied.