Ask where AI runs inside a company and the first answers are about writing: drafts, summaries, chat. Look at the request logs and a different picture appears. A large share of calls ask a model to decide something about a piece of content — which team, which policy, how urgent, which passage, which model — and then parse the answer back into a value. That is the work TypeSafe built Jev for: typed questions in, probabilities out, no text to parse1.
This guide walks through the use cases TypeSafe documents2, with a request design for each, and the rules that decide whether a given decision is a good first candidate. If Jev is new to you, the explainer covers the API and the three question types: noul (yes/no), choice and score.
The shape every use case shares
Every design below has the same four steps.
ticket, message, listing, claim
An event arrives
filter first
Code builds the state
POST /v1/systemone
One request, several questions
your policy
Thresholds decide
Two of the steps carry most of the quality. Filtering the state matters because Jev 1.13's accuracy falls as unrelated content grows, so retrieval and trimming belong before the call3. Deciding in code matters because the answer and its confidence are two separate signals: the answer says what, the confidence says whether to act. TypeSafe's guidance is to set thresholds per decision, scaled to the cost of being wrong4.
| Illustrative designs | State | Questions | What code does with the answers |
|---|---|---|---|
| Support triage | Ticket text, plan, product area | choice team; noul urgent; noul refund; score frustration | Queue, priority, refund workflow |
| Moderation | Post plus thread context | noul per policy; score severity | Allow, warn, review or remove |
| Model routing | The user's prompt | choice tier; noul needs tools | Pick the route or model for the LLM call |
| RAG context | Query plus one candidate passage | score relevance; noul answers the question | Keep the top passages, drop the rest |
| Lead routing | Inbound form, company profile | score fit; noul buying intent; choice segment | Assign owner, sequence, priority |
| Claims triage | First notice of loss, adjuster notes | score complexity; noul missing info; noul fraud indicator | Straight-through, specialist or investigator |
| Contract review | One clause or section | noul per required term; choice clause type | Checklist with gaps for counsel |
Customer support triage
Support is the use case TypeSafe's own examples start from: classify each ticket by issue and product area, detect urgency, frustration, churn risk and refund requests, and route to the right team, queue or automated workflow2. A single request carries all of it — a Choice for the team, Nouls for urgency and refund, a Score for frustration — and one answer set drives the queue, the priority and whether the refund workflow opens.
TypeSafe's intent-routing pattern goes a step further: one intent routes to deterministic code with no LLM at all, two go to different specialist LLMs, and a complexity score decides between an LLM and a person5. The expensive handlers only run for the requests that need them. The same answers also check outgoing replies: TypeSafe lists verifying support responses against policy and against the customer's request as a use2.
Moderation and trust and safety
Moderation is a set of policy questions about one post. TypeSafe describes applying company-specific criteria, detecting toxicity, harassment, spam, fraud, unsafe advice, personal-data exposure and opt-out requests, and combining severity and confidence to allow, warn, review or block2. The design is one Noul per policy plus a severity Score, so adding a policy is adding a question rather than retraining a classifier.
Two of Jev 1.13's documented weak spots apply directly. It reads literally, so each policy needs its boundary cases written into the criteria, and adversarial text written to argue for its own classification can move the answer, so moderation is the place to test with hostile inputs before rollout3.
LLM guardrails and output checks
The same pattern sits around LLM calls. TypeSafe documents checks on every LLM input, output and tool call: jailbreak and injection detection, policy violations, sensitive-data exposure, tool-call errors and response-quality failures2. Its guardrails cookbook screens each message with one request carrying a battery of Noul questions and a severity Score, then applies an action threshold and a lower review threshold in code6. The guardrails and routing guide covers that design in detail.
Model routing: choosing which LLM answers
Before an LLM call, a Choice over model tiers and a few Nouls (does this need tools, is this risky) decide where the prompt goes. TypeSafe lists building a custom router, classifying intent and domain, estimating difficulty and risk, and escalating requests that need a more expensive model2. The router does not need to be right every time; it needs a confidence gate that sends uncertain prompts up a tier, so a misroute costs money rather than quality.
Search, reranking and RAG context
Retrieval produces candidates; deciding which ones answer the question is a judgment. TypeSafe lists scoring query-to-candidate relevance, reranking, and selecting useful context for downstream AI workflows2. Its reranking cookbook takes 30-passage BM25 shortlists for 40 legal queries from the CLERC dataset and asks one question per query–passage pair; on jev-1.12, the correct passage reached first place for 18% of queries, up from 5% with BM25 alone, and the top ten for 62%, up from 38%7. The useful part for a RAG system is the cut: passages scored as irrelevant never reach the generating model's prompt, which shortens it.
Sales: lead qualification and routing
Inbound leads arrive as forms, emails and company profiles. TypeSafe describes matching them to an ideal customer profile, scoring industry fit and company maturity, detecting buyer relevance, pain points and purchase intent, and prioritising and routing leads2. A fit Score, an intent Noul and a segment Choice give the CRM three fields it can sort and assign on, with low-confidence leads left for a person to read.
Insurance claims and financial crime
Claims and alerts share a structure: a large stream, most of it routine, a small part needing a specialist. For insurance, TypeSafe lists classifying first-notice-of-loss reports and adjuster notes, detecting complexity, missing information and potential fraud indicators, prioritising straight-through processing or specialist review, and escalating uncertain or high-risk cases to a human adjuster2. For financial crime it lists evaluating transaction narratives and KYC documents, matching entities across inconsistent records, prioritising alerts and routing ambiguous cases to investigators2.
In both, Jev ranks and routes; it does not decide a claim or file a report. Its documented weakness with numbers and dates also matters here: amounts, thresholds and time windows are computed in code from fields the model extracts or the system already holds3.
Legal and compliance checklists
A contract review is a checklist: is each required term present, is any prohibited clause there, what kind of clause is this. TypeSafe lists classifying contracts, policies, filings and marketing claims, detecting missing clauses, prohibited claims and policy violations, and escalating uncertain findings to counsel2. Asking one Noul per required term over one clause at a time keeps each state small and each answer checkable, and the output is a gap list for a lawyer rather than a verdict.
TypeSafe's SEC filings cookbook shows why confidence matters for this kind of classification. Across 60 annual reports classified by industry, a confidence cutoff of 0.9 split them in half: the confident half was right 90% of the time, the other half 40%, which rose to 70% when reported one level up the industry hierarchy8. The design that follows is to act on the confident half and handle the rest differently, not to accept one average accuracy figure.
Catalog, listings and feature extraction
Retail platforms with many sellers normalise listings, extract attributes from titles and descriptions, and detect prohibited listings and counterfeit signals2. Because Jev does not generate text, attribute extraction works best as a Choice over known values, or as a check over candidates a parser has already proposed3.
The last use is less obvious and often the most valuable: turning text into features for a classical model. TypeSafe describes extracting probabilistic signals — purchase intent, churn risk, urgency — from natural-language data and combining them with structured data in models trained on real outcomes2. A churn model that already uses billing and usage data gains a column for how a customer's support messages read.
Recruiting and other decisions about people
TypeSafe lists recruiting: evaluating applications against explicit, job-related criteria, scoring evidence for required competencies, routing candidates and escalating uncertain cases for human review2. Decisions about people deserve a stricter design than the others in this guide. Write the criteria as job-related evidence, keep a person deciding, log every answer with the model version that produced it, and check outcomes across groups. The European Commission lists AI tools for employment and the management of workers, such as CV-sorting software, among the AI Act's high-risk uses, with the rules for that area applying from 2 December 20279. Treat hiring, lending and insurance decisions as regulated work, not as a classification exercise.
Choosing the first use case
1 · fit
High volume
Thousands of the same judgment a day, where per-call cost and latency add up.
2 · fit
Closed answer set
The outcome is a label, a yes/no or a level your code branches on.
3 · fit
Ground truth exists
Past decisions or a reviewed sample you can score the model against.
4 · care
Reversible or reviewed
Where a wrong answer is costly or affects a person, confidence gates send it to someone.
The best first candidate is a decision you already make thousands of times a day, whose answer is a label your code branches on, and for which you have past outcomes to test against. Support triage and moderation usually qualify. Run the questions over a few hundred labelled examples with a pinned model version — TypeSafe's jev-latest alias moves with each release10 — find the confidence threshold above which the answers are right often enough, and start by acting only above it.
Running these decisions through INFRO
INFRO implements TypeSafe's endpoint at the same path, POST /v1/systemone, with the same request body, so each of these designs runs through the same gateway as the application's LLM calls, on the customer's own TypeSafe connection and TypeSafe bill. TypeSafe connections are available in production: Jev through INFRO shows the setup, and the System One reference the request, the response and the errors.
With a connection in place, a metadata value such as {"workflow": "claims-triage"} on each request groups Jev decisions by use case in usage and traces, beside the LLM calls the same workflow makes. Spend limits apply to Jev requests as they do to any other, checked before the call. Each response carries usage.cost, the estimated TypeSafe charge, so the cost of a decision and the cost of the escalations it triggered can be read per workflow.