How Decision Models and Causal AI Work Together
Where Jev and OpenAI Decisions API fit in enterprise decision making
An engineer submits a report about recurring vibration in a production asset. Before anyone can decide what to do, the business has several questions to answer. Which equipment does the report concern? Which previous incidents are relevant? Does it need specialist attention? Would changing the operating regime reduce the risk of failure, or should the component be replaced?
Those questions belong to the same workflow, but answering them well requires different capabilities.
New decision services such as TypeSafe’s Jev and OpenAI’s Decisions API could make many of the judgments within that workflow faster and more economical. Causal AI contributes to the analysis needed to understand the likely effects of changing what the business does. At Geminos, we see an opportunity to bring these capabilities together within ArchWay, assigning each a clear role.
What Jev and similar services contribute
Jev evaluates supplied context against narrowly defined questions. It returns a choice, a score or a yes/no probability that software can use directly. Its choice and score outputs also include confidence information. TypeSafe describes Jev as a System One model, emphasizing fast, focused judgments.[1]
OpenAI’s recently announced Decisions API addresses a similar need. Developers provide text or image context and questions with predefined answers, then use the results to classify content, route requests or select an agent’s next action. The announcement establishes a similar application role; it does not establish identical performance or confidence behavior.[2]
The attraction is practical. Enterprise applications repeatedly need to interpret information and choose among known options. A useful document can be ranked for relevance, a request assigned to a workflow, or a proposed response flagged for further review. Each judgment may be small, but together they can consume considerable time and model usage.
These services are candidates for making that work more efficient. Whether they improve a particular application still needs to be measured.
What causal analysis adds
Return to the maintenance example. Suppose the historical records show that equipment receiving frequent maintenance also experiences more failures. It would be a serious mistake to conclude immediately that maintenance causes failures. The equipment may receive more maintenance precisely because it is older, more heavily used or already showing signs of trouble.
That is an illustrative example of why an observed pattern can be an unreliable guide to action.
Causal analysis makes the proposed intervention explicit and examines what is needed to estimate its effect. It requires assumptions about how the relevant factors relate, evidence appropriate to the question, and scrutiny of alternative explanations. A causal conclusion is conditional on that foundation.[3]
The resulting questions are operationally important: Would bringing an inspection forward reduce failures? How much would a different operating regime change risk? Which interventions are likely to improve the outcome, and under what conditions? Causal methods can also support investigation of past events and counterfactual questions about what might have happened under different actions.[4]
A decision model may select “replace the component” from a list. The selection alone does not establish the expected benefit of replacement. That requires an analytical basis connecting the intervention to the outcome.
Equally, a causal model does not automatically deliver the right answer. Missing evidence, inappropriate assumptions or a poorly specified intervention can undermine the result. Sometimes the useful conclusion is that a further inspection or experiment is needed before the business can make a defensible comparison.
Match the capability to the question
The distinction becomes clearer when we look at the work inside an application. These are illustrative assignments, rather than claims about a particular model’s tested performance.
| Business question | Appropriate primary capability |
| Which reports are relevant to this investigation? | Retrieval, with a decision model potentially ranking or classifying the candidates. |
| Which approved investigation workflow fits this request? | A bounded judgment, operating within application constraints. |
| How likely is this asset to fail next month? | A validated predictive or Bayesian model; a forecast does not necessarily require causal analysis.< |
| Would changing the maintenance interval reduce failures? | Causal analysis of the proposed intervention and its alternatives. |
| Which maintenance schedule best meets our objectives and resource limits? | Optimization, informed by credible estimates of outcomes, costs and constraints. |
| Is this action authorized, and has the required approval been obtained? | Explicit permissions, deterministic checks and the approval workflow. |
Commercial decisions have the same structure. Classifying an enquiry and routing it to the right salesperson may suit a decision model. Estimating whether a discount would generate additional profitable sales requires evidence about the effect of the discount. Producing a credible recommendation may then require both causal analysis and an economic calculation.
The word “decision” covers all of this work. It does not make the underlying tasks equivalent.
A bounded answer can still have serious consequences
A short list of possible answers makes an interface easier to control. It does not make the consequences of choosing the wrong answer small.
“Shut down the turbine” and “continue operating” form a tightly bounded choice. So do “approve” and “reject” for a substantial payment. The business risk depends on what happens next.
For Geminos, the sensible initial uses are judgments whose mistakes are detectable and recoverable: organizing evidence, suggesting a route, or flagging a case for review. Even here, an apparently minor classification can matter if it causes important evidence to disappear from an investigation.
Confidence information can help determine when to gather more evidence or escalate. Jev’s documentation describes using different thresholds for different consequences and validating those thresholds on the intended workload.[5] A threshold does not replace an approval requirement or an operating limit.
A decision service can therefore contribute to a consequential workflow without receiving authority over the consequential action.
Probabilities need an explicit meaning
A probability attached to a classification and a probability of a business outcome under an intervention answer different questions. Confidence that a report belongs in the “bearing issue” category does not tell an engineer how much bearing replacement would reduce failure risk.
Bayesian methods have a distinct role here. They combine a specified model, prior information and observations to update estimates and represent uncertainty. They can also be used within causal analysis. A model being Bayesian does not, by itself, establish that its relationships are causal.[3][6]
The categories can overlap: a Bayesian classifier might handle the same narrow task as a decision service. The important requirement is to understand what each output means and whether it supports the action being considered.
How the capabilities could work together
Consider an illustrative application investigating repeated equipment problems. A decision service could help classify incoming reports and identify material worth reviewing. KnowledgeWay would connect the reports to the relevant asset, operating history and engineering context. Uncertain interpretations would remain visible rather than becoming unquestioned facts.
An agent could then assemble the evidence and invoke an appropriate causal model to investigate the likely effects of different interventions. The application would bring those results together with practical considerations such as downtime, staffing and cost. An engineer would review the evidence, assumptions and alternatives before deciding on the response.
This is where the complementarity matters. Faster interpretation and routing can reduce the effort needed to reach a useful analysis. Causal analysis gives the intervention assessment an explicit foundation. Human expertise helps establish whether the proposed action makes sense in the actual operating environment.
The handoffs need care. If a report is misclassified or linked to the wrong asset, sophisticated downstream analysis can still address the wrong situation. The complete workflow is what must be evaluated.
How we are thinking about this within ArchWay
ArchWay is Geminos’ reference architecture for bringing appropriate capabilities together around an operational solution. Our thinking is to make bounded decision services another reusable capability within that architecture, available where they demonstrate value.
KnowledgeWay would provide connected business context and evidence. A service such as Jev, OpenAI Decisions API or a suitable self-hosted alternative could contribute focused judgments. CauseWay would supply causal analysis where the question concerns drivers, interventions or counterfactual outcomes. Predictive models, optimization and conventional software would continue to contribute where appropriate.
Geminos agents can now invoke causal models, with the available tools and authority defined for each application. Incorporating these newer decision services is an area we are evaluating; the workflow above describes the direction we are considering, rather than an announced production integration.
We would keep the choice of provider flexible and judge candidates on the customer’s tasks: accuracy, useful uncertainty estimates, response time, total cost and deployment requirements. Application controls and accountable people would retain authority over consequential actions.
Through PathWay, successful patterns and their evaluation criteria could become reusable assets for subsequent solutions. A well-tested way to classify a report or route an investigation should make the next application easier to build and maintain.
Better decisions depend on the whole application
For customers, the opportunity is to spend less time assembling and sorting information, and more time evaluating what to do with it. Faster judgments can help people reach the relevant evidence. Causal analysis can help them assess what a proposed change is likely to achieve.
Our aim at Geminos is to make the knowledge and discipline of the best decision makers available more widely. These new services could help us do that when they are combined with shared context, appropriate analysis and clear human responsibility. The measure of success will be the quality and speed of the customer’s decisions.
Sources
- TypeSafe AI, Introduction to Jev and System One models.
- OpenAI, DevDay 2026 Recap, September 29, 2026. Decisions API was announced in limited preview.
- Judea Pearl, Causal inference in statistics An overview, Statistics Surveys, 2009.
- PyWhy, Performing causal tasks, DoWhy documentation.
- TypeSafe AI, Confidence.
- Joon-Ho Lee, Avi Feller and Sophia Rabe-Hesketh, Model-based inference for causal effects in completely randomized experiments, Stan case study.
Sources checked October 1, 2026. The business examples and proposed ArchWay uses are Geminos’ interpretation; these sources do not evaluate Geminos products.