Insight / Blog
Enterprise AI Agents: One Model for Every Decision?
Jev: AI Built to Decide, Not Write
Does every decision in an enterprise AI agent need the same level of AI?
Jev, released by TypeSafe AI on September 15, differs from the generative AI models we have grown used to. Rather than writing long answers or code, it reads a situation and returns a structured decision, such as a choice, score or boolean, along with probabilities. Software can use those outputs to decide what happens next: which team should receive an inquiry, for example, or whether a process should move to the next step.[1]
The response was swift. Jev joined Vercel AI Gateway the following day, and 13% of Vercel's paid teams used it within the first 24 hours. Vercel described this as its fastest initial model adoption to date. The figures were recorded during a free trial: they indicate early interest, not paid purchases or sustained use.[2][3]
If new AI models keep getting better at generating answers, why experiment with a model focused on decisions? The answer starts not with a model benchmark, but with the work enterprises ask their agents to do.
A complex workflow is a sequence of different decisions
Even a simple order inquiry involves understanding the customer's request, verifying their identity, retrieving the order and explaining the result. It looks like one task from the outside, but each step calls for a different kind of work.
Salesforce's official customer-support example illustrates the distinction. A large language model (LLM) handles the conversation and selects the next function to call. A defined verification procedure, however, determines whether the authentication code matches; its result controls access to order information. Natural-language judgment and rule-based verification coexist within the same agent.[4]
Microsoft's AI for its own cash-collection operation also goes beyond forwarding emails. It summarizes customer conversations, predicts the likelihood of overdue payments or disputes, and helps staff prioritize their work. Choosing whom to route a case to and deciding how to respond in light of a customer's history require different information.[5]
|
Even invoice processing separates routine work from exceptions
SAP's invoice-matching example makes the distinction more concrete. An agent compares invoices with purchase orders and goods receipts, handles differences within the permitted tolerance, then records the payment schedule and processing result. SAP describes a workflow in which the agent handles the 80% of invoices that follow standard patterns, while the finance team handles the 20% requiring separate judgment.[6]
That ratio is not an automation target for every enterprise. What matters is how the broad task of processing invoices is broken down by the conditions involved. Separating record matching, tolerance checks and exception review changes both the scope of automation and the role of staff.
Before setting the automation boundary, define which conditions the system can handle and which changes should trigger a person's review. Without that boundary, handing work to an agent may still leave people reviewing every result.
Jev aims to handle some of those decisions separately
Rather than replacing an entire agent, Jev can separate decisions made during a workflow from text generation. Vercel lists request classification, selection of the next tool or subagent, priority scoring, and decisions to continue, retry or stop as potential uses. Instead of receiving a long explanation and interpreting it again, software receives a value it can use directly.[2][3]
A short answer does not necessarily mean a simple decision. A question such as 'Should this be approved?' might involve a single amount threshold, or it might require contracts, transaction history and reasons for an exception. Having only two possible answers does not make those decisions interchangeable.
If a rule can reliably establish whether an authentication code matches, AI does not need to estimate the answer again. A task that involves reading an inquiry or business situation and choosing from defined options may, by contrast, be worth evaluating with a decision model. Combining multiple sources into a response strategy, or writing an explanation for a customer, is another role again. The purpose of breaking down the work is not to add more models, but to remove unnecessary processing at each step.
Repeated decisions add up to operating costs
These distinctions also affect cost as usage grows. In Vercel's April 2026 AI Gateway production data, requests ending in tool calls accounted for 22.2% of requests but 58.9% of tokens. Work that represented a minority of requests accounted for more than half the data the models processed.[8]
An agent may read search results, query another system, review the output and make another decision. The price of generating one final answer therefore does not reveal the cost of the entire workflow. To see where costs accumulate, examine what the agent reads and decides, and how often.[8]
Those figures alone cannot establish the savings a decision model would deliver. Whether separating repeated decisions is worthwhile must be tested in the actual workflow. If incorrect classifications lead to more retries and human reviews, a cheaper individual call may not reduce the total processing cost.
The impact of an error matters, too. A case sent to the wrong department and then reassigned is different from an incorrect approval that exposes customer information or triggers a payment. Consider both the information a decision requires and the consequences of getting it wrong when defining what can run automatically.
Break down the workflow before choosing a model
This shift matters most to business leaders setting the scope of AI adoption, AI transformation teams, AI project managers, and development and IT organizations. For them, the more useful question is not 'Should we use Jev too?' but 'Which parts of our work require which kinds of decisions?'
The exercise need not be complicated. Choose one workflow, write out its actual processing sequence, and add the information required, the decision rule and the owner of any exception beside each step. Customer support might involve understanding the inquiry, verifying identity, retrieving an order, deciding whether the issue can be resolved, responding and handing it to a staff member. This makes it easier to distinguish rule-based processing, contextual judgment and decisions people must make.
Business teams can then explain approval rules and exceptions, while developers identify the data and conditions needed to execute the workflow. Agreeing on 'what can run automatically and when it should be handed over' is more useful for defining development scope and test requirements than a broad request to 'automate customer support.'
Microsoft likewise explains that redesigning the work experience and preparing the underlying systems mattered as much as developing its cash-collection agent. Attaching a capable model was not the same as making the workflow function in practice.[5]
Agent adoption starts with breaking down the work, not choosing a model. That breakdown reveals the scope for automation, the level of reasoning each step needs and the points requiring human review. Model comparison can then follow, guided by the work each step actually demands.
Jev is not the answer for every enterprise agent. But its arrival changes the order of a familiar conversation. Before asking 'Which AI is best?', ask 'What does AI actually need to decide in our workflow?' With AI that decides rather than writes now emerging, enterprises have reason to look beyond a new model's name and examine the work they plan to delegate.
References
- [1] TypeSafe AI — Introducing System One Models & Jev
- [2] Vercel — Jev is the fastest-adopted model in AI Gateway history
- [3] Vercel — TypeSafe AI's Jev now available on AI Gateway
- [4] Salesforce — Multi-Surface - Customer Support Agent
- [5] Microsoft — Streamlining finance cash collection at Microsoft with AI
- [6] SAP Learning — Identifying Enterprise Value and Use Cases for Agents
- [7] SAP — Sourcing Assistant | SAP Ariba Sourcing
- [8] Vercel — AI Gateway production index
SURROMIND in the News
Explore recent media coverage of SURROMIND.