Case study 01 · n8n · AI workflow

AI email routing with a human decision boundary.

The workflow classifies an incoming email as urgent, invoice, or routine—but AI never gets the last word when confidence is low.

Public demonstration using synthetic, mocked, or fictional data; not a paid-client result. Accounts and credentials are not included; the workflow uses buyer-supplied services.

0.70human-review threshold
3explicit routing categories
4successful routes audited

// try the browser demo

Email triage · runs in this page

Write or load a message

Edit the subject lines — one per line — then run the triage. The classifier is keyword-deterministic and runs entirely in this page: no AI service is called and no inbox is connected. Write something the rules do not recognise and watch it stop at the gate instead of guessing.

Use synthetic sample data only. Do not enter confidential information, personal data, credentials, or customer records. Input stays in this page and is discarded on reload or close.

Anything below the 0.70 confidence gate stops at human review instead of guessing.
category
confidence
route
Ready. Load or write a message, then run the triage.
Public demonstration · synthetic messages · no credentials · no paid-client data

// system flow

Email triggerStructured AI classificationConfidence gateExplicit branchAudit log

The problem

Shared inboxes mix time-sensitive issues, invoices, and routine mail. A naive AI agent can route quickly but may also make a confident-looking mistake with no review trail.

The control design

The model must return a fixed JSON schema: category, confidence, and one-sentence reason. Confidence below 0.70 bypasses automation and alerts a human-review queue.

Deterministic actions

Urgent messages alert the team, invoices move through filing and ledger steps, and routine messages are labeled. The workflow—not the model—chooses the allowed action.

Handover boundary

The importable n8n JSON contains no credentials. Operators connect their own inbox, model, team-chat, storage, and sheet accounts after reviewing the nodes.

// proof available

email → {category, confidence, reason} → confidence gate → route → decision log

The complete workflow is available as an importable starter kit. The same pattern can be adapted to lead intake, support queues, document review, or approval routing without claiming autonomous judgment.

Need this mapped to your inbox?

Send the categories, exception examples, tools, and desired human-review rule.

contact@metaworld.systems