X02 — A support inbox sorted in two languages
A shop’s inbox mixes everything: a refund request next to a delivery question, a compliment next to an angry customer charged three times, in English and in Italian. This workflow sorts every message with one Decide with a Model step — the team, the urgency, whether something arrived damaged — and pays for a written reply only where the answer says one is needed and the model is sure.
A reference run: 8 messages, 8 sorted, 2 reply drafts — 0.052 credits. The run’s outputs.
The case
Section titled “The case”Eight messages from a week of a home-and-gifts shop: broken headphones with a burnt smell, a parcel stuck in tracking (in Italian), a chair with a cracked leg, a question about a pushchair (in Italian), a triple charge (in Italian), a compliment, a customer hesitating over a pair of trainers, and flowers for a funeral tomorrow morning that have not left (in Italian). The messages are the example’s own, written for it.
What it shows
Section titled “What it shows”- One step, three answers. Team (a choice of five, at 0.7), urgency (a score on three levels) and damage (a yes/no at 0.8) in one call per message: 0.004 credits each.
- Two languages, the same questions. The questions are in English; the Italian messages are sorted as well as the English ones — the same was true on the 30 messages of our measurement set.
- Pay only when it is needed and sure. The reply draft — the one step that writes, and the only one that costs more
than a decision — runs only when the team is refunds and
all_accepted: a Combine Conditions, then a Filter. The other six messages cost nothing more. - A queue, not a pile. Every message comes out in the Work queue with its team and probability, the urgency level
and score to sort by, the damage flag, whether the model was sure, and the draft when there is one (
nullotherwise).
The workflow
Section titled “The workflow”
| Node | Type | Why |
|---|---|---|
| Inbox | input/json_value |
The messages: [{"id": "…", "text": "…"}] |
| One message at a time | enumerate/json |
One iteration per message |
| Triage the message | ai/decide |
team (refunds, replacements, deliveries, product_advice, other), urgency (low, medium, high), damaged |
| A refund request? | utility/predicate equals refunds |
|
| A refund, and the model is sure | utility/boolean_combine all with all_accepted |
A doubtful refund gets a person, not an automatic draft |
| Only sure refund requests | utility/filter |
The draft’s branch runs only for those |
| Draft the refund reply | ai/text_generation |
Three sentences, the customer’s language, no date or amount promised |
| Work queue | aggregate/json, the draft with whenAbsent: null |
Every message with its answers |
Reading a run
Section titled “Reading a run”| Message | Team | Probability | Urgency (score) | Damaged | Draft |
|---|---|---|---|---|---|
| M-101 headphones burnt, “I want my money back” | refunds | 1.00 | medium (1.07) | yes | written |
| M-102 “il tracciamento è fermo da quattro giorni” | deliveries | 0.99 | medium (1.05) | no | |
| M-103 chair with a cracked leg, “send me a new chair” | replacements | 1.00 | medium (0.85) | yes | |
| M-104 “il passeggino chiuso entra nel bagagliaio?” | product_advice | 0.99 | low (0.19) | no | |
| M-105 “mi avete addebitato tre volte” | refunds | 0.92 | high (2.00) | no | written, in Italian |
| M-106 a compliment to the staff | other | 1.00 | low (0.00) | no | |
| M-107 trainers “maybe half a size small? I might keep them” | product_advice | 0.75 | low (0.01) | no | |
| M-108 “i fiori per il funerale di domani alle 9” | deliveries | 0.99 | high (1.98) | no |
Every answer was sure. The hesitant trainers message (M-107) is the closest call — product advice at 0.75 against a
threshold of 0.7; at 0.8 it would have been flagged for a person. Sort the queue by urgency_score and the funeral
flowers and the triple charge come first.
M-101 draft: “I’m sorry the headphones stopped working and developed a burnt-plastic smell. I’ve passed your refund request to our refunds team for review. Please stop using them.”
M-105 draft: “Mi dispiace per i tre addebiti sullo stesso ordine. Ho inoltrato la richiesta di rimborso al nostro team rimborsi, che verificherà la situazione.”
Cost: 0.052 credits — eight decisions at 0.004 and two drafts at 0.01. A thousand messages with a refund rate like this one cost about 6–7 credits.
Install it
Section titled “Install it”node madoo-install-example.mjs x02-support-inbox --api <your Madoo API URL> --key <key.json> --runChange it
Section titled “Change it”- Your teams: the options of
team, with descriptions in your words; keepother. - Your rules: a draft also for replacements (a second Predicate and Filter), or none at all; escalation with a
Predicate on
urgency_score. - Your channel: the messages can come from a form, an email integration or a webhook — anything that gives text.
- A stricter queue: raise the threshold of
teamto 0.8 and sendmodel_sure: falseto a person.