Classify emails, leads and tickets in Make with a model trained on your own decisions

• StayCharted: an Essentials or Business workspace, which includes the API, with at least one published model. Training works on every plan, Free included, so you can check the model's report first.
• An API key: in StayCharted, Settings → API keys → Create key. A key belongs to one workspace and can use that workspace's published models. It's shown once, so keep it somewhere safe.
• The model's ID and the columns it reads: open the model in StayCharted, choose Use → Prediction API. It shows the model's ID and how many columns it was trained on. The column names are the ones in your training file (for example Subject and Message).
• Make: any Make account. The Free plan has a 15-minute minimum interval for scheduled runs. Credit use depends on the modules and how often they run. Check current Make plans.

Make HTTP module: API key and Bearer token

Make keeps the key in a keychain, so it never sits in the scenario itself.

1. In a scenario, add the HTTP app's Make a request module.
2. Set Authentication type to API key.
3. Next to Credentials, click Add and fill in the keychain:
   • Name: something you'll recognize, for example StayCharted – Support workspace.
   • Key: the word Bearer, a space, then your API key: Bearer amt_live_…
   • API key placement: In the header.
   • API key parameter name: Authorization
4. Save the keychain.

To stop access, delete the key in StayCharted: every scenario using it stops at once.

Check the connection (optional). With the keychain chosen, set URL to https://amt.staycharted.com/api/v1/models, Method to GET, Parse response to Yes, and click Run once. The output lists your published models with their IDs and the columns each reads, and names the workspace the key belongs to. It spends no predictions.

1. Trigger: the app and event where your records start, for example:

   • Gmail or Microsoft 365 Email: Watch emails (for sales leads or support requests).
   • Google Sheets: Watch new rows.
   • Typeform, Jotform or Google Forms: Watch responses.
   • HubSpot, Pipedrive or Salesforce: Watch new contacts or leads.
   • Your help desk: Watch new tickets.

   Run the trigger once so Make has a sample record to map.

2. HTTP → Make a request, with the keychain from Part 1:

   • URL: https://amt.staycharted.com/api/v1/models/<model ID>/classify
   • Method: POST
   • Body content type: application/JSON
   • Body input method: Data structure
   • Parse response: Yes

3. Describe the record once. Next to the data structure, click Add, name it after the model (for example Ticket record) and click Generate. Paste a sample with the model's own column names, spelled exactly as in your training file:

   Show JSON example

   {"input": {"Subject": "", "Message": ""}}
   

   Click Generate, then Save. Make now shows an input group with one field per column.

4. Map each field from the trigger: Gmail's Subject into Subject, the email body into Message, and so on. Fill as many as you have. At least one must contain a value.

5. Run once. You'll see the model's answer for the sample record. Then turn the scenario on and set its schedule.

Why a data structure, not a JSON string: an email body is full of quotes and line breaks. Typed into a JSON string, the first quote breaks the request. The data structure lets Make write the JSON for you, whatever the text holds.

Map the same kind of text the model was trained on. A model trained on short subjects and the first paragraph of each email answers best when it gets the same.

With Parse response on, these values can be mapped in later modules:

Send unsure answers to a person. Add a Router after the HTTP module. On one route, set a filter needsReview equal to false and act on the answer. On the other, filter needsReview equal to true and send the record to a person, for example a Slack message or an email with the model's answer and a link to the record. Records under the cutoff also wait in StayCharted's Review Queue. Pick one place to review; checking both is wasted work.

Act on the answer. Map label (or a fields value) into the next module: a Gmail label, a CRM owner or stage, a help-desk group, a spreadsheet column. To give each answer its own route, add one Router route per answer with a filter on label.

• Sales leads from Gmail: classify each inbound email as a lead type (Enterprise, Small business, Partner, Not a lead). Create the contact in your CRM with that type and owner; send Not a lead to a label instead.
• Support requests: classify by category and team, then create the ticket in your help desk with both filled in. Unsure ones go to a triage channel.
• Form entries and spreadsheet rows: write the answer to a column in the same row, so your team sorts by it.

Each starts with a model trained on decisions your team already made: past emails with the lead type you gave them, resolved tickets with their final category and team.

See AI email routing for a shared inbox for training examples and a review workflow.

Paid-search review: Classify rows from an exported search terms sheet and send suggestions to your team for approval. Sort Google Ads search terms with your trained model →. Applying account changes is a separate step.

• One record per run of the module. For thousands of existing rows, fill a file or a Google Sheet in StayCharted instead. Developers can send up to 100 records in one call with the predict endpoint.
• It doesn't learn from the scenario. Corrections your team makes in the other apps reach the model when you retrain on a fresh export of those decisions, the same way you trained it.
• Models that read pictures take a link to the picture instead of text fields. See the API reference for the request.

Make shows StayCharted's own message in the HTTP module's output. The code field says which: