How To Build a Jev Decision Workflow With No Engineering Help?

Every Monday, someone on the team opens the shared inbox and sorts hundreds of messages by hand. Refund request, bug report, sales lead, spam. It is the same decision, made again and again, and it eats hours.

Automating it sounds like an engineering project, so the idea waits in a queue. Quick chatbot prompts do not fill the gap, because they give a different answer every time and nobody trusts them.

A Jev decision workflow solves this without code. It is a decision you already make, a few narrow questions and plain rules for what happens next. Jev AI is TypeSafe AI's decision model, and it returns typed answers with confidence scores, which is what makes those rules possible. This guide walks through seven steps to build one without engineering help.

• A Jev workflow is one repeated decision, a few narrow questions and rules for acting or asking a person.
• Jev handles judgment, while spreadsheets and built-in filters handle math, dates and counts.
• Confidence bands and a human review path make the workflow safe to trust.
• Test on examples you have already labeled before you turn anything on.

A Jev decision workflow is made of five parts that run in the same order every time.

The five parts are:

1. Trigger: A new ticket, form answer, row or message arrives.
2. State: The text and fields Jev reads.
3. Questions: A few narrow Choice, Score and Noul questions.
4. Rules: Thresholds on probability and confidence decide the next action.
5. Action or review: The item is routed automatically or sent to a person.

Before you start, gather four things:

• A decision you repeat often.
• A few dozen past examples where you know the right answer.
• Written criteria for each outcome.
• Someone who can review the uncertain items.

Follow these seven steps in order. Each one produces something the next one uses. The examples use a support inbox, but the steps work for any repeated decision.

Step: 1. Pick One Repeated Decision and Split Out the Rules

Start by choosing the decision and deciding what Jev should and should not handle.

Choose with these guidelines:

• Pick a decision that repeats: It should have a fixed set of outcomes and lead to a clear next action, such as routing a message to Billing, Bug, Sales or Other.
• Start with low risk: Tagging and routing are good first picks. Leave payments and approvals for later.
• Keep rules out of Jev: Amounts, dates, counts and status checks belong in formulas or filters. TypeSafe's own notes say Jev is not a calculator, and it struggles with counting and with reading dates as ordered values.

Write the decision as one sentence, such as "Route each support message to Billing, Bug, Sales or Other." Anything that needs a calculation, such as "refund over $500," becomes a filter in front of Jev. If you want a deeper look at what Jev handles well, see how Jev differs from general LLMs.

Step: 2. Define the State Jev Will Read

The state is the material Jev judges. Keep it small and clear.

Build it with these rules:

• Include only what the questions need: For the inbox example, that is the message text and the customer plan.
• Name fields clearly: Use labels like customer_message and plan so each field is easy to point to.
• Leave out unrelated detail: TypeSafe notes that Jev can be distracted by irrelevant context, so skip old tickets and extra history.
• Convert other formats to text: Jev accepts a string, a JSON object or an array of text. Images, audio and video are not supported, so transcribe or extract text first.

English works best. TypeSafe says other languages are accepted but currently have lower accuracy, so test any non-English inputs before relying on them. By the end of this step you have a short, labeled record for every item.

Step: 3. Write Narrow Questions With Clear Options

Narrow questions give clearer answers than one broad question. TypeSafe's guide to building with System One calls this the most important idea in the whole workflow.

Write your questions this way:

• Split vague judgments: Instead of one "is this spam?" question, ask separate yes/no questions, such as whether the message asks for credentials, offers an unexpected reward or comes from a mismatched sender.
• Match the type to the need: Choice fits categories, Score fits ordered ratings such as frustration level and Noul fits yes/no checks.
• Define each option: For every Choice option, write what it covers, what belongs elsewhere and two or three example messages.
• Add an escape option: Include "Other" or "Not enough information" so Jev is not forced into a bad match.
• Ask together: Questions over the same state are evaluated independently and in parallel, so you can send them all in one request.

When you finish, you have a short list of questions, each with plain, literal wording and options that say the same thing as your instructions.

Step: 4. Set Confidence Bands for What Happens Next

Confidence bands turn Jev's answers into actions. A band is a range of confidence that maps to one action.

Start with these three bands, using TypeSafe's confidence guide as a reference:

Treat these numbers as a starting point to test, since TypeSafe says the right values depend on your domain and your data.

Adjust the bands with these rules:

• Raise the bar for risky actions: The harder an action is to undo, the higher the confidence it should need.
• Handle yes/no differently: Noul questions return a probability and no separate confidence score. Treat a middle range, such as 0.4 to 0.6, as uncertain and send those items to review.

You now have a written rule for every possible answer.

Step: 5. Build the Review Path

The review path is where uncertain items go and how you capture what a person decides.

Set it up with these points:

• Pick a landing spot: Use a sheet tab, a channel or a thread for "check first" and "send to a person" items.
• Show the full picture: The reviewer should see the record, Jev's answer and its confidence.
• Log every correction: Use a fixed format with three columns, the record, Jev's answer and the right answer.

Once this is in place, no uncertain item falls through a gap, and you have a growing record of where Jev and your team disagree.

Step: 6. Test Before You Go Live

Testing shows whether your questions and bands hold up on real records, before any live item is affected.

Run these checks:

• Run a labeled sample: Use past records where you already know the right answer, and compare Jev's output with your labels.
• Check confidence against accuracy: TypeSafe suggests plotting confidence against accuracy on your own data. If answers above 0.85 are often wrong, raise the bar from Step 4.
• Try tricky wording: Include sarcastic, vague and mixed-topic messages.
• Rerun after every change: Any edit to a question or option needs a fresh test.

If the results match your labels, you are ready for live traffic. If they do not, return to Step 3 and tighten the option definitions.

Step: 7. Launch and Monitor

Launch on a slice of live volume first, then widen it as the results hold.

Roll out in this order:

• Start with a slice: Run the workflow on a small share of incoming items and keep human review on.
• Record each live answer: Save Jev's answer and confidence next to the final outcome, so you can compare them. For a tamper-evident record, see how to build an AI audit trail.
• Review the correction log weekly: Use the Step 5 log to find patterns, then tune your bands and option definitions.
• Widen only when results hold: Add more volume once corrections drop.

After a few weeks, the workflow runs mostly on its own, and your team only sees the items that need judgment.

PromptQL runs the whole workflow from a plain-English thread, in workspaces where Jev has been provisioned. You describe the decision and the rules, and PromptQL runs Jev across your records.

Follow these steps:

1. Check availability: Ask, "Is Jev available in this workspace? If not, tell me before we start."
2. Share your decision map: Paste the outcomes, the criteria and the rules you keep outside Jev, then ask PromptQL to restate them and flag anything unclear.
3. Test on a labeled sample: Ask PromptQL to run Jev on the sample and list every record where Jev disagreed with your label, with its confidence.
4. Set your bands: State the three confidence bands and ask for three outputs, acted on, check first and send to a person.
5. Run the full set: Run it on your records, then rerun it on each new batch.
6. Invite reviewers: Bring teammates into the thread to correct items, and apply their corrections to the criteria.
7. Turn results into a dashboard: Ask PromptQL to build a view of the outcomes in the same thread.

Here is a ready-to-copy prompt for the support inbox example:

Is Jev available in this workspace? If not, tell me before we start.If it is, build a triage workflow for the messages in the Inbox table. Label each message Billing, Bug, Sales or Other, and use Other when a message does not clearly fit. Handle one rule outside Jev by flagging any refund request over $500 for manual review. First, run Jev on the 40 messages in the Labeled tab and show me every message where Jev's answer differs from my label, with its confidence. Then apply these bands to the full set. Above 0.85, mark the message as done. From 0.6 to 0.85, add it to a Check First list. Below 0.6, add it to a Send to Person list. If Jev is not available, stop and tell me.

Keep your criteria in the thread so later runs follow the same rules, as explained in how to stop context decay. To share results with your team, see how to build dashboards and apps from a multiplayer AI thread.

Four things matter before real records flow through the workflow.

Keep these points in mind:

• Cost: Jev is priced at $0.042 per million input tokens, and output is free, so estimate your spend from your average record length.
• Versions: TypeSafe's models page explains that jev-latest is an alias that moves when a new release ships. Pin a specific version such as jev-1.13.0 once you have tuned your bands, and retest before you switch. The response's model field shows which version answered.
• Safety: TypeSafe notes that Jev can be vulnerable to adversarial content, so text inside a record can try to steer it. Keep hard rules and permissions outside Jev.
• Data: The records you send go to TypeSafe. According to its privacy policy, TypeSafe does not train or fine-tune on your input and does not disclose it to third parties other than service providers. Check what your own policies allow before you send sensitive content.

The cheapest test happens before you open any tool. Pick 10 records from your inbox, write down the answer you expect for each and the band it should land in.

Then compare notes with a teammate. If the two of you disagree on more than two records, the criteria are too vague for any model, and rewriting them now saves you a week of corrections later. When you both agree, you have your first labeled sample and your first draft of option definitions, ready to paste into whichever tool you use.

Can Jev Write the Reply After It Decides?

No. Jev returns typed answers and never writes text. Let Jev make the decision, then pass the result to a chat model that drafts the reply. This guide to a Jev and LLM workflow shows how the two fit together.

What Happens If Jev Is Unavailable?

Treat it the same as a low-confidence answer. Send the affected items to your review list and rerun them later. Add a fallback rule to every automation, so it pauses and alerts a person when Jev cannot be reached.

Which Decisions Should Stay With a Person?

Keep these with a person:Hard-to-undo actions: Payments, account closures and approvals.High-stakes calls: Hiring, legal and medical decisions.Cases without examples: Anything new, where you have no past answers to test against.Jev can still sort and flag these items, and a person makes the final call.

Can You Reuse One Workflow for a Different Decision?

Reuse the structure, which means the trigger, the state layout, the bands table and the review path. Write new questions and options for the new decision and test them again. TypeSafe notes that thresholds do not carry over between question types, so set new bands each time.

1. Jev Pricing — How Jev Works - Jev AI Community - www.jevai.org
2. Jev Workflows | Decision intelligence for software teams - Jev AI - thejevai.com