STAYCHARTED AMT / HOW IT WORKS

From your team's past work to a model that does the next ones

Upload examples your team has already handled, with the answer they gave. Check them, train a private model, see how often it's right on examples it never saw, and put it to work. Answers it's unsure about go to a person.

A file returned with predicted categories and confidence
A file returned with predicted categories and confidence · Sample data

What can it learn?

Examples you bringThe answer it learns
Support tickets, emails, documentsA category: queue, topic, document type
Inbound leads and messagesThe reply your team would send, and who sends it
Files and SharePoint sitesWho has access, and at what level: a list of groups
Invoice linesSeveral fields at once: GL account and cost center
Product and inspection photos, with or without a noteA category: product type, defect, damage

The walkthrough below uses support tickets; the same steps apply to all of these.

STEP 1

Bring examples with answers

A spreadsheet (CSV, Excel or JSON) with the columns that describe each item and the answer your team gave. For photos: folders of pictures, or a spreadsheet of picture links, optionally with a text column such as the inspector's note.

  • SharePoint and file permissions: upload your permission export as it is (PnP PowerShell report, SharePoint's sharing report or Microsoft Graph JSON). StayCharted combines it into one row per file, with the groups that can open it.
  • Data in Snowflake or Databricks: on Essentials and Business, choose a table and the columns to read instead of uploading a file. [How →](/guides/snowflake-databricks)
  • No answers yet? On Essentials and Business, StayCharted groups your rows for you to name, suggests the rest, and you review them before anything trains.
  • Not sure where to start? Pick what you categorize on Start here and try a sample first.
Choose the input columns and the answer column
Choose the input columns and the answer column · Sample data

STEP 2

Check what the model will learn

Before training, StayCharted finds duplicate rows, answers that look like the same answer, answers with too few examples, and personal information such as email addresses and phone numbers. Fix each in a click: mask, merge, or leave rows out. Pattern-based checks don't catch everything; for photos, review duplicates and pictures that may be in the wrong category.

Review detected personal information before training
Review detected personal information before training · Sample data

STEP 3

Choose a model and train

  • Classifier Model: learns from the words. On every plan.
  • AI Classifier Model: also reads meaning, so it copes with different wording. Free includes one.
  • Dedicated AI Model: an AI fine-tuned on your own examples. On Business.
  • For photos (Vision add-on): the AI Image Classifier Model, and on Business the Dedicated AI Image Model. Both can read a text column with each photo.

Or choose Your AI Model to use your company’s own AI without training; it is measured the same way. Setting it up and testing it are included on Essentials and Business; publishing it needs the Your AI Models add-on.

Choose the training approach available on your plan
Choose the training approach available on your plan · Sample data

STEP 4

See how often it's right

Every training run is tested on examples it never saw. The report shows accuracy overall and per answer, which answers it mixes up, and the examples it got wrong. For a list or several fields, it scores each part: for each group, how often it was added wrongly or missed.

  • Choose what gets checked: for each confidence cutoff, the share of answers a person would check and the share of the model's mistakes among them.
  • Test on a file whose answers you know, and compare runs side by side before you publish one.

The scores in the screenshots belong to sample data; they're not a promise for yours.

Check accuracy by category and inspect mistakes
Check accuracy by category and inspect mistakes · Sample data

STEP 5

Put it to work

Publish the version you want, then use it wherever the work happens:

  • Files: fill a CSV, Excel or JSON file; you get it back with an answer and a confidence on every row.
  • Google Sheets: the add-on fills the empty rows in your sheet.
  • Snowflake and Databricks: fill new rows from a table; every answer goes to a results table in your warehouse, with its confidence and whether it needs review.
  • API: send records from your own systems, or from a Power Automate flow, and get the answer, its confidence and whether it needs a person.
  • Claude and ChatGPT: ask your model from your AI assistant.
  • Download it: take a trained AI Classifier or Dedicated model with you.

The mobile app: ask a published model from an Android phone, signed in or with an access code. More →

Get the Google Sheets add-on ↗ · Model download terms →

Fill new rows with answers and confidence
Fill new rows with answers and confidence · Sample data

ALWAYS

A person checks what the model isn’t sure about

Answers below your cutoff, from files, Sheets, the API and assistants, wait in the Review Queue. Beside each one: the model’s top three answers and similar examples your team labeled. Confirm or correct them, and the next version learns from the corrections. An optional daily email says what’s waiting. (Essentials and Business.)

Built for teams, improved over time

Roles

Owners, Builders, Reviewers and Viewers, each with what their work needs.

Activity

A record of who changed what, in each workspace and model.

Versions

Every run is kept. Publish the one you choose; the published one keeps answering while a new one trains.

Explore the details

Try it on work your team has already done.

Free includes text Classifier Models and one AI Classifier Model. Essentials adds API access, the Review Queue and starting without answers; Business adds Dedicated AI Models. Vision is an optional add-on on paid plans.