FlowsBook a callSign inCreate free accountSign up
AI employees and automation

Keeping a log of what each AI employee did and sent

An AI agent audit log is four records, not one: the run, each step's items, the activity page and the daily note. Where each lives, and who may read it.

Part of the guide AI employees

By the xMap team · · 6 min read

An AI agent audit log should answer four questions about any piece of automated work: what happened, on what data, who set it up or changed it, and what went out of the door. In xMap Flows those answers sit in four places. The run records what happened, each block's items record the data, the Activity page records the people, and the Team inbox holds a plain note from each AI employee about what it did and sent. You do not have to switch any of them on.

This is written for the person who will one day be asked "why did this customer get that email?" or "who told it to do that?": an operations lead, a compliance officer, an agency owner answering to a client. It walks through each record, shows how to trace one message back to its cause, and sets out who can read what.

The run: what happened, step by step

Every time an AI employee works, on its timetable, from a form, a chat message or a click on Run once, that is one run, and each run is kept in the flow's list of runs. Open one and you see how it started, when, how long it took, whether it passed, and a single line that sums it up: how many things it found, how many were new, and whether it emailed or posted them.

Under How it ran each block is a row with its own result. A sending block says what it sent and to how many. A block that failed says why in words. If a person had to approve the work before it went out, the run keeps who was asked and what they answered, including any message they edited before approving.

Each block's items: the data behind a decision

A summary line is not evidence. For that, every run keeps the full list each block passed on for 30 days. Press See all on a block to open it, search any field, sort by a column, and download it as CSV or Excel. Nothing is fetched again and nothing is charged.

The most useful part for an audit is the Removed tab on blocks that narrow a list. It lists what a Filter, Only new, Remove duplicates or Sort and limit left out, each with the reason in words, such as "rating 3.8 is under 4" or "found before on 3 Nov". So you can show not only why something was sent, but why something else was not.

For an AI agent that holds conversations, See how it answered opens its trace: the message it received, each tool it called and what came back, and the answer it gave. People who cannot manage the flow see less of this, which matters when conversations include customers' details.

Activity: who changed what

Runs tell you what the employee did. The Activity page tells you what people did to the employee: who created it, who changed its blocks, who deployed or paused it, who changed how much it may do alone, who answered an approval, and what Ask xMap did on someone's say-so. Changes made by replying to an employee in the Team inbox are listed there too, and so is anything xMap support opened or changed while your admins allowed it access.

Organization admins can download the activity log as a CSV file covering two years, which is the file to hand an auditor. Version history sits beside it: each saved version of a flow can be compared with the current one in plain words, so "what was it set to in October?" has an answer.

The Team inbox: the employee's own account

The Team inbox has a thread for each AI employee. Once a day, an employee that ran writes a note: what it did, what it found, the best few results, and anything it needs from you. A failed run or an approval waiting gets an urgent note of its own.

Whether you also get a note each time it sends depends on a single choice per employee, covered in the AI employees guide:

  • Ask me first: nothing goes out until a person approves it, and the approval is on record in the run.
  • Act and tell me: it sends, then leaves a note in its thread saying that it did.
  • Fully on its own: it sends without a note each time. The runs still record everything.

Tracing one message back: a made-up example

Say a customer of an invented pest control firm in Brisbane complains on a Thursday about an email offering a discount they did not ask for. Here is the trail.

  1. Open the AI employee that sends offers and find Thursday's run in its list. The summary line says it emailed 14 businesses.
  2. On the Email block, See all shows the 14 with the address each message went to. The customer is among them.
  3. On the Filter before it, See all shows the customer passed a rule: "opened in the last 12 months". The Removed tab shows 31 others that did not.
  4. On the block before that, a table of existing customers was meant to take current customers out. Its list shows no match, because the customer's name is spelled differently in the table.
  5. Activity shows who built the flow and that nobody changed it that week, so the cause is the table, not a recent edit.

All the numbers above are invented. The point is the order: run, sending block, the blocks that decided, then the people. Fifteen minutes later the firm can tell the customer what happened and fix the table.

Who can see what

  • People who can see the flow can open its runs and each block's items. Viewers read but change nothing.
  • A private flow that runs on its owner's personal account, such as their own mailbox, is seen only by its owner and the people it is shared with. Admins can pause or archive it without opening it, and its runs do not appear in their lists.
  • The Activity page follows the same rule: an admin sees a colleague's owner-only flow there only in what the admin did to it.
  • xMap's own support staff cannot open your organization unless one of your admins allows it for a set time, and what they open and change is listed for your admins.

Mistakes that leave gaps

  • Relying on the inbox alone. Notes are a summary for people. The run is the record.
  • Leaving the review too late. A full list of what each block passed on is kept for 30 days. Look into a complaint that week, or keep your own copy.
  • One shared mailbox for everything. When every employee sends from the same personal account, it is harder to say whose work a message was. Give each flow a clear owner.
  • Testing with real sends. Use test runs to check wording. Run once sends for real and belongs in the log as a real send.

To keep your own ledger, end the flow with two actions, not one: the message itself, and a row in a sheet with the date, the recipient and the reason. The Save to my database block does the same into a table you own.

Questions

Do I have to switch logging on?

No. Every run, each block's result, the Activity page and the Team inbox notes are kept without any setting.

How long are the records kept?

The full list of items each block passed on is kept for 30 days and can be downloaded in that time. Admins can download the activity log for the last two years. For longer, write a row to your own sheet or database in the flow.

Can I see why something was not sent?

Yes. Blocks that narrow a list, such as Filter and Only new, have a Removed tab that lists each item left out with the reason in words.

Can an admin read everything every AI employee sent?

Almost. Admins see the runs of team flows. A private flow that runs on its owner's personal account is seen only by its owner and the people it is shared with.

Is a test run part of the log?

A test run is kept like any other run and shows what it would have sent, but it sends nothing and is not remembered as results.

Templates that do this

Read next

More on ai employees and automation · Every article · Templates

Put it to work

An AI employee in xMap Flows does this on live data, on its own timetable. Start from a template or describe it in plain words.