A customer opens a support ticket, returns to your pricing page, and then looks up how to cancel. You may already capture all three interactions. Deciding what to do next is where things get interesting.
Should someone reach out? Does the customer need help with a product issue? Would another promotional email make the situation worse? What is the right image to show this customer?
These are the kinds of questions that make Jev, a new model from TypeSafe AI, really interesting. It takes context and returns structured, probabilistic answers that can be used for a variety of automation and experiences. For organizations moving customer data through Tealium, that creates an opportunity to explore more nuanced decisions inside the workflows they use today.
A model built for decisions
TypeSafe introduced Jev earlier this month September as its first “System One” model, designed for fast, bounded judgments. You provide the context and define the questions and possible answers. Jev returns structured values through its API, with probability signals your application can use.
There are three question types:
- Choice: Select from a defined set of options, such as routing a customer to technical support, billing, or an account manager.
- Score: Evaluate something against a rubric, such as the urgency of a service issue.
- Noul: Evaluate whether a statement is a yes or no and return a probability between zero and one, such as whether a customer’s message expresses an intention to cancel.
Multiple questions can evaluate the same context in one API call. Choice and Score also return confidence values. The answers arrive as structured JSON, resulting in defined values to work with.
That is an appealing format when the next step is updating an attribute, evaluating an audience, or routing a task. Your team defines what the answer means for the business and which actions it can trigger.
Tealium Functions supports Jev (available now)
Tealium Functions gives customers a way to call external model APIs using event or visitor data. A function can assemble the relevant payload, send it to a model, and receive a result. Returned values can then flow through Tealium, where configured enrichments write them to profile attributes for audience evaluation and downstream activation.
The useful part is how much of the surrounding work may already be in place. You have events, customer profiles, audience rules, and connected destinations. You can build on those investments to test a new kind of decision.
Recognizing when a customer needs help
Consider a subscription customer whose usage has declined over the past month. Yesterday, they submitted a support message about a recurring problem. Today, they visited the cancellation page.
A cancellation-page visit alone tells you very little about the reason. The recent support issue and change in usage give that visit more meaning.
A team could bring that sequence together in a payload and ask Jev to assess the strength of the warning signs against a defined rubric. A separate Noul question could evaluate whether the support message expresses an intention to leave. Choice could help classify the issue for the right team.
Those answers could inform a retention workflow: prioritize a service follow-up, route the issue to someone equipped to resolve it, or pause an unrelated promotion where the connected systems support that action.
The score gives the team a way to evaluate the combination of behaviors and the customer’s own words, then test whether the resulting intervention helps.
Choosing a more useful next step
Now imagine a shopper comparing two expensive products. They have returned several times, reviewed delivery information, and asked whether one model works with equipment they already own.
A discount given as an incentive for purchase wouldn’t solve for a compatibility question, and would just leak margin.
A retailer could ask Jev to choose among a small set of next steps: offer compatibility help, provide a product comparison, explain delivery options, or continue the existing experience. The payload would include the relevant browsing sequence and the customer’s question, where permitted.
The result could become an attribute that a connected experience platform uses at its next decision point. If the answer is ambiguous, the workflow could keep the existing experience or offer a simple way to ask for help.
This also gives teams something specific to measure. Did customers who received compatibility help complete their purchase more often? Did they return fewer products? The test is whether the decision improves the experience and its outcome.
Describing the customer you want to reach
Marketers often know who they want to reach before they translate that audience into attributes and rules. Jev could give them another way to evaluate the fit: describe the customer in plain language, define what a strong match looks like, and ask the model to assess the profile.
Imagine a retailer planning for the holidays. The team describes three audiences: seasonal gifters, loyal high-value customers, and casual shoppers. A customer’s purchase history shows orders concentrated around the holidays, several with gift wrap. This week, they have browsed multiple products in different categories. Together, those signals give the model more context for assessing a gifting pattern than purchase timing alone. The context from each profile informs a Score question, with criteria describing what a weak, partial, or strong match might look like.
Those scores can become profile attributes that a marketer uses to build audiences or help determine which message a connected channel serves. The gifter could see gift guides and shipping deadlines. The high-value customer could see early access. The casual shopper could see new arrivals or deals in a category they have browsed.
The marketing team can refine those descriptions as a campaign evolves and the workflow could rescore profiles as relevant new events arrive, using the same activation paths already in place.
Speed and cost expand what is worth testing
TypeSafe reports end-to-end response times of 70–500 milliseconds and launch pricing of $0.042 per million input tokens, with no output-token charge. Its published evaluations show substantial speed and cost advantages over the chat-based models it tested.
What might be more interesting though is that this model call is fast and inexpensive enough that teams could consider using it for smaller, more frequent decisions. More moments become worth evaluating.
An initial experiment could run alongside an existing workflow, recording Jev’s answers without changing customer treatment. Compare those answers with what your team would have chosen and what actually happened. That gives you evidence for deciding where to introduce automation.
Room for the next idea
Tealium’s role is to give customers the flexibility to connect their data to the models and services that suit their ideas. Functions extends that flexibility to API-accessible systems, including those without a prebuilt integration.
As new approaches like Jev emerge, you can evaluate them against a real customer problem, using the data streams and activation paths you have already built.
We continue building infrastructure for your best ideas. And that gives you room to try something that was not on anyone’s roadmap when you first connected your data.
