Review AI prompt logs

What this achieves

Every AI action taken in your workspace is recorded, one row per call, with who triggered it, what it was for, whether it succeeded, and what it cost. This is the screen you use when somebody asks what the AI has been doing.

Required role: Administrator.

Steps

Go to Administration → AI Governance → AI Logs.
Read the figures across the top.
Narrow the list with Module, Feature, Status, From, and To.
Select a row to open the call.
Select Clear to reset the filters.

The figures across the top

Option	Description
Total logs	How many calls the current filters match.
Success rate	The proportion of those calls that succeeded.
Avg latency	How long those calls took on average.
Total tokens	The total volume processed across them.
Total credits	The credits those calls consumed.

These figures describe what the filters match, not all-time usage. Narrow to one month and they describe that month. That makes them useful for answering a specific question and misleading if you read them as totals.

Example: filtering HC Corp’s log to April and reading 340 calls does not tell you the year’s usage. It tells you April’s.

The columns

Option	Description
User	Who triggered the call.
Feature	Which AI feature it was.
Module	Which part of the product it came from.
Model	Which model answered it.
Tokens	The volume processed.
Latency	How long it took.
Credits	What it cost in credits.
Status	Whether it succeeded.
Feedback	Any rating the user gave the result.
Created	When it happened.

Reading an individual call

Opening a row shows the same fields plus an Input summary and an Output summary, and an Error where the call failed. Both summaries are hidden until you select Show.

Those two summaries are the point of the screen. A user reporting that an AI feature gave them something odd is a question you can answer here, by reading what it was asked and what it returned, rather than by reproducing the problem.

Example: if James Whitfield reports a strange course suggestion at HC Corp, find his call on this log and read the input and output summaries. That is faster than asking him to generate again and hoping for the same result.

An empty log usually means nothing has run

The screen says as much: adjust the filters, or wait for AI features to be used. A date range with no AI activity in it produces an empty log, which is a correct answer rather than a failure.

Use it for cost as well as conduct

Because every row carries its credits, this log is where a credit figure becomes attributable. When the credit dashboard shows consumption climbing, the log is where you find which feature and which people drove it.

What happens next

Logs are a record, not a control. Where the log shows a feature consuming more than it is worth, the response is a conversation about how it is being used — or a change to who has credits — rather than anything on this screen. Where it shows recommendations being rated poorly, that belongs with bias monitoring and with the teams whose data feeds those recommendations.