Branch on an AI decision

AI is great at judgement calls: is this email urgent? which category is this? But a judgement is only useful in a flow if you can act on it — and you can’t reliably branch on a paragraph of prose. The fix is to make the AI step return a typed value: not “Well, this looks fairly urgent…”, but exactly urgent.

A typed value turns a fuzzy AI judgement into a real, predictable value you can route on. Ask AI → Switch (or Branch) gives you deterministic routing off an AI decision — classify an email and send each kind down a different path, every time.

Typed answers, not prose

The Ask AI step has a Response format setting. Leave it on Free text and you get prose. Change it, and Routario coerces the model’s reply into a real value and puts it in the step’s ai_parsed output:

Response format	You get back
Free text	The text as written
Integer	A whole number
Decimal	A number with decimals
Yes / no	A true/false value
JSON	A parsed object
JSON list	A parsed list
Choice	Exactly one value from a list you define

Read the typed result downstream as {{ <label>.ai_parsed }} — for example {{ classify.ai_parsed }}. (The raw, uncoerced text is still available at {{ classify.ai_response }} if you want it.)

Choice is the one that unlocks routing, so the rest of this guide focuses on it.

Choice: one of a fixed list

Pick Choice as the response format and you supply a list of allowed values. Routario steers the model to answer with exactly one of them, then checks the reply against your list and hands back your value. If the model goes off-script, the step fails loudly rather than passing junk downstream.

Setting the list of choices

You type the allowed values into the Enum values field. Routario is forgiving about how you write them — all of these produce the same list, [rock, paper, scissors]:

You enter	Parsed list
["rock","paper","scissors"] (chips / JSON)	[rock, paper, scissors]
rock, paper, scissors (commas)	[rock, paper, scissors]
rock paper scissors (spaces)	[rock, paper, scissors]
rock; paper; scissors (semicolons)	[rock, paper, scissors]
rock ,, paper , (stray spaces / empties)	[rock, paper] (blanks dropped)
(empty)	the step errors — Choice needs values

It also “just works” from a flow value

If your choice list comes from an earlier step or a variable, it may arrive looking like ['rock', 'paper', 'scissors'] (single quotes). Routario understands that form too — you don’t need to do anything special.

How a reply is matched

Once the model answers, Routario matches it against your list. Matching is lenient about formatting but strict about content — the result is always one of your exact values, in your casing. With the list [rock, paper, scissors]:

The model returns	Result
rock	rock
Rock or ROCK	rock — matching ignores case, returns your casing
"rock".	rock — surrounding quotes and . , ! ? ' " are stripped
rock (with stray spaces or a newline)	rock — surrounding whitespace is stripped
I'd pick rock	error — not an exact value
lizard	error: Response 'lizard' is not one of: rock, paper, scissors

The returned value uses your casing, not the model’s

Define your options as [Urgent, Normal, Spam] and a reply of urgent comes back as Urgent. So you can write your choices the way you want them to read, and branch on them with confidence.

A reply that doesn’t match raises an error and the run fails on that step — which is what you want: a misclassification should stop and surface, not quietly flow into the wrong branch. (In practice this is rare: the step instructs the model to answer with just one of the values.)

Worked example: triage incoming support email

Every inbound support email is auto-sorted into urgent, normal, or spam and routed down a different path.

1. Incoming email trigger. The flow starts when a message arrives at the flow’s own inbox address. The trigger hands the email to the steps that follow as plain values — {{ from }}, {{ subject }}, {{ body }}, and any {{ attachments }}.

2. Classify it with Ask AI (label: classify). Add an Ask AI step and set the prompt to:

   Classify this support email.
   Subject: {{ subject }}
   Body: {{ body }}

   Set Response format to Choice and Enum values to urgent, normal, spam. The decision is now available as {{ classify.ai_parsed }} — guaranteed to be exactly one of those three.

3. Route on the decision with a Switch. Switch on {{ classify.ai_parsed }}, with one case per value:

   • urgent → a Send notification step that pings the on-call channel.
   • normal → add it to the queue — a Create record in your tickets table, or a Create task.
   • spam → leave the case with no onward step; the run just ends and the message is dropped.

Why the Choice format is what makes this work

Step 3 can switch on a clean value — urgent — because step 2 is guaranteed to return one of the three options. If step 2 were Free text and replied “This looks pretty urgent to me”, the Switch would have nothing reliable to match. Typed Choice is the contract that lets a deterministic Switch sit downstream of a fuzzy AI judgement.

It ties straight back to the two tables above: the three values can be entered as chips or as urgent, normal, spam; a reply of Urgent or urgent. still resolves to urgent; and a reply of maybe urgent? errors rather than guessing.

Two outcomes? Use Branch instead

Switch is for three or more outcomes. For a yes/no decision — approve this, or not? — use a Yes / no response format and a Branch step on {{ classify.ai_parsed }}.

Where to go next

• Switch and Branch — the routing steps that act on the decision.
• Ask AI — reference — every field on the AI step, including the other response formats.
• Routing a flow — how branching, switching, and merging fit together.