Skip to content

Audit trail and monitoring

Every action an agent takes is recorded with its inputs, its decision and its result. This page covers what the audit trail holds, how to review a single decision, what is monitored and how sampling works at Level 3 (Supervised) and Level 4 (Autonomous).

Part What it holds Example: an agent enters a sales order
Inputs What the agent had in front of it The customer’s email and purchase order as received. The customer record, prices and credit status read from the ERP, as they were at that moment. The rules and limits in force.
Decision What it decided and on what basis Which of your items it matched to each line. Which checks passed. That the order was inside its limits and could be entered without an approval.
Result What happened The sales order number the ERP returned. Or the exception it raised, with the reason. Or the ERP’s error message if the write was rejected.

Next to each agent action, the operations layer records what people did with it: who approved, what they corrected and how they resolved an exception. It also records every use of a control: a changed limit, a change of level, a stop and a restart.

Our forward-deployed engineers build every audit trail to four rules.

  • Every action is recorded, not a sample of them. If the record cannot be written, the action does not go ahead.
  • Nobody edits the trail. An agent cannot change or delete an entry. A correction is added as a new entry and the original stays.
  • It is your data. The trail lives in the operations layer, you own it and you can export it. How long it is kept is agreed in your contract.
  • A stated reason is not proof. For a judgment step, such as matching a customer’s description to one of your part numbers, the trail records the reason the agent gave. Treat that as a pointer. What you can verify are the inputs, the output and the checks the operations layer ran.

The audit trail itself comes from our reusable tooling. What it records for each kind of action in your workflows, the alerts and the sampling rules are written for you.

The audit trail is built so that a review can start from whatever you have: an ERP document number, a customer’s email or a date.

  1. Find the action. Search the audit trail by the ERP document number, by the date or by the name of the customer or supplier.
  2. Read the inputs. Open the source document as the agent received it, and the ERP values it read.
  3. Read the decision. See what it matched, which checks passed and what it flagged.
  4. Read the result. See what was written to the ERP and what the ERP answered.
  5. If it was wrong, mark it. Record the error and its cause. The transaction is put right in the ERP with the ERP’s own functions, and the error counts in the workflow’s measurements.

This is also how you answer the question a customer or an auditor asks months later: why was this order entered at this price.

During the Embedded Build and under Operate & Automate, we run and monitor the software and the agents. These are the things we watch.

What is watched What it tells you
The age of the oldest work in each queue Work is piling up, or an agent has stalled
Exceptions, by reason A jump means something changed upstream: a customer’s new purchase order layout, a price list update
Corrections and rejections at approval People are changing more of what an agent prepares than they used to
Actions stopped by a limit An agent is running into its limits, so the limits or the incoming work have changed
Rejected write-backs The ERP is refusing transactions: a changed field, a closed period
An input that has gone quiet Nothing has arrived from the customer service inbox since morning, so the connection may be broken
Sampling results An error was found in work the agent completed
Stops, level changes and limit changes Someone used a control, and the people who depend on the workflow should know
AI model usage It is billed at cost, so you should see when it moves

Alerts go to our engineers and, when a decision is needed, to the workflow owner. How quickly we respond is agreed in your contract. Stages and terms lists what your contract records, and Operate & Automate covers how support works.

Monitoring does not take the place of the stop switch. If something looks wrong from the desk, stop the agent first and ask questions after.

From Level 3, a person no longer sees every transaction before it reaches the ERP. Sampling is how people keep checking the work.

  1. The operations layer picks completed work for review. Our engineers build the picking to one rule: part of the sample is random, drawn across everything the agent did, and part is aimed at risk. Risk means larger amounts, cases close to a limit, the first cases after a rule changed and work for new customers or suppliers.
  2. A person who knows the work reviews each one against the source document, in the same view used for an approval.
  3. They mark it correct or wrong and, if wrong, give the cause.
  4. A wrong one is put right, and the workflow owner decides what follows: fix a rule, tighten a limit or move the workflow down a level.
  5. The results are counted in the workflow’s measurements. At Level 3 they are evidence for the next promotion review, and our engineers bring them to it.

How much is sampled is set for each workflow with the workflow owner and written down before the level starts.

Level 3 Level 4
What people do Handle exceptions and spot-check Set policy and audit
When the sample is reviewed As part of the day, by the people who work the exception queue As a regular audit, by the workflow owner or the people they name
What it is for Catching errors, and showing whether the workflow is ready to move up Confirming the policy is being followed and is still the right policy