Today we're launching Decisions in HyperTrail, a new building block powered by JEV from TypeSafe AI. Decisions let enterprises run high-quality automated decisions on every customer profile, at a speed and cost that until now made them impractical at scale.

The problem: great decisions, too slow and too expensive at airline scale

HyperTrail builds a living profile for every customer, with thousands of insights stitched together in real time from every system you run: bookings, stays, purchases, support chats, app sessions, loyalty activity. Agents read that profile and take the right action at the right moment.

That worked well, but it had a ceiling. Every decision went through a large language model. LLMs are remarkable at open-ended reasoning, but most business decisions aren't open-ended. Is this passenger at risk of missing their connection? Is this guest eligible for an upgrade? Which of these four offers fits best? Using a full LLM to answer questions like these meant two things:

Latency. An action for a customer could take several seconds from event to outcome. That's fine for an email, but too slow for a gate change or a transaction at the point of sale.
Cost. At the volume of a large airline, with millions of passengers and tens of millions of events a day, per-token pricing added up quickly. Running a decision on every event for every customer was expensive enough that teams had to pick which moments to automate.

What JEV changes

JEV is a new kind of model. Instead of generating text, it evaluates a state and returns typed, structured answers: a yes/no, a score, a choice from a list you define. Because the possible answers are fixed in advance, JEV can't return a value outside your schema. There's no parsing and no hallucinated output.

In HyperTrail, a Decision works like this:

The profile is the state. The customer's profile and facts are rendered into the Decision's input, the same context an agent would see.
JEV answers typed questions. For example, churn_risk (score), upgrade_eligible (yes/no), or next_best_offer (spa, dining, lounge, miles).
Deterministic rules pick the action. Rules are evaluated top to bottom. The first match calls an action with its parameters, which can be static values, fields from the profile, or JEV's answers.
Everything is logged. Each run records the question answers, the rule that matched, and the resulting action, so every decision can be audited and replayed.

A JEV Decision in the HyperTrail demo: a customer profile on the right, the Decision's typed answers and matched rule, and the in-app action it triggered

A Decision in our live demo. A Pulse Ring sync lands on Maya's profile. JEV answers typed questions (recovery=0.4/poor, rest_first=0.10, routine=gentle_activation), the gentle_default rule matches, and an "Easy start" routine is pushed to her watch. The trace shows the model, token usage, and cost of every decision: here, $0.00018.

The results:

Actions in milliseconds, not seconds. JEV decisions benchmark at p90 400 ms, and profiles are served from HyperTrail's purpose-built storage in single-digit milliseconds, so the full path from event to action stays sub-second.
Up to 400x lower cost than running the same decision through an LLM.
Consistent, reviewable output. The same profile gets the same kind of answer, in the same shape, every time.

Together, these create a new operating point: every event, for every customer, can get a decision. Workloads that used to be sampled or batched can now run on the full stream.

Decisions and agents, side by side

Decisions don't replace agents; they complement them. HyperTrail continues to support agents built on Anthropic Claude and OpenAI GPT models for the moments that need open-ended reasoning: writing a personal apology after a bad stay, negotiating a rebooking, or chaining several tools together.

The pattern we recommend is simple: use a Decision whenever the outcome can be expressed as a score, a classification, or a choice, and use an agent when the moment truly needs language or multi-step reasoning. In practice, most high-volume flows are Decisions, and a small share of moments escalate to an agent. That combination makes automated decision flows fast and cost-effective at any scale.

	Decision (JEV)	Agent (Claude / GPT)
Output	Typed answers + rule match	Free-form reasoning and text
Best for	Score, classify, route, pick an offer	Compose, negotiate, multi-step tool use
Speed	Sub-second (p90 400 ms)	Seconds
Cost	Up to 400x cheaper	Per-token
Audit	Deterministic, replayable	Full trace of every step

What this looks like for an airline

A flight is delayed by 50 minutes. Within the same second:

Every affected passenger's profile is evaluated. JEV answers misconnect_risk, tier, and rebook_preference for each one.
Rules rebook passengers at high risk of misconnecting onto the best alternative, push a notification to their app, and issue lounge access to eligible passengers.
The handful of complex cases, like a family with a tight international connection and special assistance needs, escalate to a Claude or GPT agent that composes a tailored plan, or to a human agent for approval.

Hundreds of passengers get the right action before they reach the gate, at a cost that makes sense to run on every disruption, every day.

Getting started

Decisions are available today in every HyperTrail project. Create a Decision, define its questions and rules, and wire it into a trail with an Entity-to-Decision (E2D) connector. You can test any Decision against a real customer profile before you deploy it.

HyperTrail is available on the AWS Marketplace, so your usage is billed as part of your existing AWS bill and can count toward your AWS commitments. No new vendor onboarding required.