Decision models
Some conditions need judgement: “is this ticket about a staff member starting at the school?”, “how serious is this security alert?”. Flows ask these as typed decisions: a question with a fixed set of answers, so the model can only ever pick one of the answers you wrote. Model text never becomes a record, a recipient or a target.
Two places in a flow ask a question:
- A condition in words in the “Only if” list.
- A Decide step, which branches on the answer.
Both work with AI off. Nothing in a flow depends on a model being there.
Kinds of answer
Section titled “Kinds of answer”| Kind | Example | Answers |
|---|---|---|
| Yes or no | “Is the requester asking for access for someone else?” | Yes, No |
| One of these | “Which kind of request is this?” | Two or more options you write, each with a short description that helps AI choose |
| A score | “How urgent does this sound?” | Named levels, lowest first |
A question is up to 200 characters. Write it as you would ask a colleague.
The Decide step
Section titled “The Decide step”A Decide step asks one question and has one lane per answer, plus Not sure. In its panel you set:
- What it looks at: up to eight values from the run, such as the ticket’s subject and description, its category or the device model. Sensitive values are never offered, and text the requester wrote is marked as theirs when AI reads it.
- How sure it must be: shown only when a decision provider that gives calibrated confidence is on. Below this, the run goes to Not sure. Your school chooses it; we do not ship a default.
- When it is not sure: Ask a person to choose the answer (the usual choice), Use the keyword rule (yes or no questions only) or Follow the Not sure lane.
A run goes to Not sure when the answer is below the bar, two answers tie, no answer comes, the provider is down, or the school’s AI limits are reached. When a person is asked, the run’s timeline shows “Choose the answer for this record” with the AI’s guess selected; their choice continues the run down that lane.
A run can hold at most 10 Decide steps.
The question library
Section titled “The question library”Every condition in words and every Decide question is saved to the school’s
question library, on the Questions tab of Flows (you need
flow.build). A question used by several flows is asked once per record and
learns once. When you add a question, you can pick one your school already
asks.
Each question shows its kind, the words that decide when AI cannot, which flows use it, and its record per answering model (“Laya: 18 of 50 decisions so far; people agreed with 17”). Changing a question’s wording, kind or answers starts its record again, and the editor warns you before you save.
Each question has its own mode, under What runs do with AI’s answer. It saves the moment you choose it.
| Mode | What runs do |
|---|---|
| Off | AI is not asked. The keyword rule decides, or a person |
| Watch only | AI answers and the answer is recorded, but the keyword rule or a person decides |
| Suggest | A person sees the AI’s answer already chosen and confirms it |
| Automatic | Follows the answer on its own once it has earned it; until then it behaves as Suggest |
A question’s mode is capped by the school’s Flow decisions setting
(flow.decide) in AI settings: set to Shadow there, no
question goes beyond Watch only; set to Suggest, none goes beyond Suggest.
Until a school sets Flow decisions, questions run as Watch only at most.
With AI off for the school, every question is off and the keyword rules
decide.
For a condition in words in Suggest mode, a run the AI says yes to but the words do not starts as Proposed, for a person to confirm or dismiss.
How a question learns
Section titled “How a question learns”Every answer is recorded in the school’s AI suggestion ledger, visible in AI activity. Only explicit human answers teach it:
- choosing the answer when a run was not sure;
- confirming or dismissing a Proposed run;
- Should have run on the Checked list;
- Was this right? on a Decide card in a run’s timeline.
Automatic outcomes never count, and confirming an automatic answer teaches nothing.
The bar for acting alone
Section titled “The bar for acting alone”A question in Automatic mode follows an answer without a person only when all of these hold:
- that question’s wording has at least 50 answers decided by people in the last 60 days, with people agreeing at least 90% of the time;
- that record belongs to the model that answered this time;
- the answer clears the confidence bar with a single clear winner;
- the school’s Flow decisions setting allows Automatic.
The record is kept per wording and per answering model, so a question can act alone for one model and still ask a person for another. A run follows the mode earned by the model that actually answered it. A question that sees 30 answers a year will stay in Suggest, and the editor says so plainly.
Corrections
Section titled “Corrections”Correcting an automatic answer with Was this right? reverts it and is audited. Three corrections in 30 days send the question back to Suggest, with the reason written to the audit log, so a person confirms each answer again.
Which provider answers
Section titled “Which provider answers”Flows never pick a provider. Every decision follows the school’s Decision provider order in Admin, AI, under Models, where the school decides where its decision text may go.
| Provider | Where it runs | Confidence |
|---|---|---|
| Laya | On the school’s own server | Calibrated, per answer |
| TypeSafe Jev | Hosted by TypeSafe in the United States, always with masked text | Calibrated, per answer |
| The school’s chat model | Wherever that AI connector runs | None |
The first enabled provider in the order answers. If it cannot, the next enabled decision provider does. The chat model answers only when no decision provider is on. Because it gives no confidence, there is no confidence bar to set: a person confirms each answer until that model has earned its own record.
Decision text is always masked (names and IDs replaced with tokens) and wrapped as untrusted, whichever provider answers. No images or attachments are sent.
The editor says who answers in plain words: “AI is off for your school”, “Decisions are answered by your chat model, which gives no confidence”, or the providers in order. When decisions fail or the school’s AI limits are reached, runs use the keyword rules or ask a person.
What it costs
Section titled “What it costs”- Questions about the same record that are due together go in one decision call, and a question already answered for a record in its current state is not asked again.
- Decide steps are asked only when a run reaches them.
- Decisions are background work. They do not count towards the AI allowance, but they count towards the school’s daily limit on background AI, and a ticket gets at most 80 background decisions a month. When a limit is reached, runs fall back to the keyword rules or a person.
- The publish review estimates the calls a flow will make, from last month’s volume.
Related
Section titled “Related”- Flows
- Building a flow
- System flows, which use the AI features’ own decisions
- AI settings