AI audit trail: who trained, published and used your model, and how

For a classification model, a useful audit trail covers two kinds of event:

• Changes to the model and its workspace. Training, publishing a version, changing settings, adding members, making API keys and access codes, filling and downloading files.
• Use of the model. Who asked it a question, through which door, and what it answered.

For both, a reviewer wants to know who did it, when, and how they reached the system. AMT keeps the first kind in the Activity log and the second in each model's recent predictions.

Every workspace has an Activity log, on every plan. It is written in plain sentences, such as "started training", "published a version", "downloaded a filled file" and "exported the activity log", so a reviewer who has never used AMT can read it without a field glossary.

What becomes an Activity entry:

• fills and downloads of filled files
• training and publishing
• settings changes
• member changes
• API keys
• access codes

Who can see it

AMT has four roles: Owner, Builder, Reviewer and Viewer. The account Owner sees the Activity log for every workspace. Everyone else sees the log for their own workspace.

Filters and CSV export

You can filter the log by action, workspace and model, which is usually how an audit request starts: "show me everything that happened to this model in the last quarter." You can export it as CSV, up to 10,000 entries at a time. Exporting is itself logged, so the log also shows who took a copy of it.

Each Activity entry records who did it and how they reached AMT.

Named API keys are what make the API rows useful. If every integration shares one key, the log can only say "the API". StayCharted's security page recommends giving integrations their own API keys and revoking keys you no longer need, and the log then shows which integration did what. The prediction API is on Essentials and Business.

When someone asks Claude or ChatGPT to start a fill or publish a model through the connector, the entry says so, and the security page notes that Activity records changes made through the assistant.

When a setting or record changes, the Activity entry records which fields changed, never their values. A reviewer can see that someone changed a model's settings, and which settings, without the log holding a copy of the data itself.

This is a deliberate choice. Activity logs get exported, emailed to auditors and stored in other systems. A log that copied values would become a second copy of your data, outside the controls that protect the first.

Teams that use the StayCharted AMT mobile app can let phones use a model through an access code. A Builder makes the code on the model's page, with limits on expiry, number of phones and daily answers. Because one code can be shared across a whole team of phones, AMT logs the code's whole life:

• making a code, pausing it and revoking it
• every phone that joins with the code
• every phone that is refused, with the reason: the code is full, paused, revoked or expired, new phones have been stopped, or the code is switched off

If a code is shared further than intended, the log shows each phone that tried to join and what happened to it.

Answering a single prediction is not an Activity entry. A busy model answers many questions a day, and logging each one would flood the log and bury the changes that governance reviews care about.

Instead, each model keeps its own recent predictions. They show each question asked through the Try box, an AI assistant, the mobile app or an access code, with who asked and through which door. An entry reads like this: "via StayCharted AMT app · access code Plant 7 technicians". The record keeps the person's name even if they later leave the workspace, so a question from last month still has an owner.

If the model was trained on masked text, the text kept in its recent predictions is masked too. See how to remove PII from training data for how masking works.

Where to look for each question

An audit trail tells you who used a model. It does not tell you whether the answers were right. AMT handles that side with confidence and review. Each answer comes with a confidence from 0 to 1, and answers below the model's cutoff go to the Review Queue, whether they came from a file, the API, Google Sheets, an AI assistant, a warehouse fill, the mobile app or the Try box. Reviewers confirm or correct them, and the confirmed answers train the next version when a Builder or Reviewer retrains. That retraining shows up in the Activity log like any other training. For how to choose a cutoff, see AI confidence scores and human review. The Review Queue is on Essentials and Business.

• Single predictions are not Activity entries. They are in the model's recent predictions, not in the Activity export.
• Values are not logged. The log shows which fields changed, not what they changed from or to.
• Exports are capped. One CSV export holds up to 10,000 entries; use filters for longer periods.
• No certification claim. StayCharted does not hold SOC 2 or ISO 27001 certification, and the security page says so. The Activity log is a record you can use in your own reviews, not a certificate.

1. Give each person a role that matches their job: Owner, Builder, Reviewer or Viewer.
2. Make a separate, named API key for each integration, and revoke keys you no longer use.
3. Make one access code per team or site, so phone use is traceable to a group.
4. Each month or quarter, filter the Activity log by model and export it as CSV for your records.
5. When someone asks who used a model, check its recent predictions.

Start in StayCharted AI Model Trainer. Plans are on the pricing page, and the full feature list is on the product page.