Limit the automated decision
Treat the AI result as a proposed queue, not an approved answer or account action. Keep the source in an access-controlled register with request_id, category_candidate, review_status and reviewed_by. Set needs_review until a person checks it. An AI category alone should not suppress urgent or sensitive enquiries.
Define categories before choosing a model
Use billing for payments and invoices, setup for connections and configuration, and other for mixed or unclear requests. Keep an example that asks about both an invoice and an integration. An explanation saying the model is confident is not a calibrated probability or evidence that its category is correct.
Minimize the classification input
Read a practice intake sheet with Watch New Rows. Send only text needed for classification, excluding unnecessary names, addresses and account identifiers. Route an empty or oversized input to manual review before calling AI. Treat instructions embedded in an enquiry as untrusted content; they must not be allowed to choose recipients or alter the workflow.
Configure Categorize text
Add Make AI Toolkit > Categorize text and map the approved body to Text to categorize. Enter category names and definitions, choose Assign only one category = Yes and request an explanation if useful. Record the connection, selected model and definition version. Inspect the actual returned structure before mapping its category into downstream conditions.
Use fixed, mutually exclusive review routes
Add a Router with exact-match routes for billing and setup. Mark the general review route as a fallback so other or unknown values reach it when neither condition matches. Overlapping conditions can produce multiple notifications. Keep destinations configured by the operator; do not map an AI-generated email address or channel name directly into a recipient field.
Make the review status visible
A notification can say “TEST-A001 | Candidate: billing | Needs review | Source link.” Avoid copying confidential bodies into a broad channel. The reviewer reads the original and records the confirmed category and owner. Preserve both the candidate and correction so later analysis can distinguish a model error from a routing or mapping error.
Test the difficult cases
Prepare billing, setup, mixed, underspecified, injected-instruction and replay cases. Have a person set the expected category and queue first, then compare candidate, confirmed category and notification count. A handful of successful examples does not establish an accuracy claim. Continue tracking corrections by category after launch.
Budget AI usage and retain a manual fallback
Save the source and needs_review state before the AI step so a failure cannot erase the enquiry. With Make’s AI Provider, review consumption for the selected model and tokens. A custom provider connection has a separate provider bill. Check the processed version and notification ID before retrying to avoid repeated classification charges and duplicate alerts.
Start a small workflow in Make
Try the required workflow with practice data, then choose a plan around its usage and features. Check current prices, promotions and eligibility in the official signup flow.
Sign up on MakeAffiliate linkEstimate costs for your workload
Enter monthly jobs, billable steps and retries to see how different billing units affect your estimate.
Open cost calculatorOfficial sources
Pricing, availability and feature limits can change. Check the linked official pages before subscribing.
- Make AI Toolkit modules and connectionsInformation checked: 2026-10-29
- Router and fallback routesInformation checked: 2026-10-29
- FilteringInformation checked: 2026-10-29
- Make creditsInformation checked: 2026-10-26
- Google Sheets module referenceInformation checked: 2026-10-29
- Make pricing — Core at 10,000 monthly creditsInformation checked: 2026-10-28
- Make scenario settingsInformation checked: 2026-10-26