Who this is for. Super Admins setting up or maintaining an Agent Co-worker. See the Roles and Access.
Wizard step 6 is where the configuration you have built is saved and the agent is switched on. The step reads: Select Run to save configuration and Activate Agent.
Figure 1. The Activate step. Steps 4 and 5 are skipped, and Run is the only action.
Reach the Activate step
Selecting Next on Define capabilities goes straight to Activate. Steps 4 (Exception treatment) and 5 (Try out coworker) appear in the progress bar but are skipped and cannot be opened.
What to check before selecting Run
Run saves the values on the wizard’s own steps and activates the agent, but it is not the first write. An inactive instance is created as soon as the work scope is complete, and the Email Categories, Processing Rules, Internal Contacts and System Senders panels save as you use them. Run is still the last point at which you can change the settings on this page before the agent starts working live traffic, so use Back to re-read step 1 in particular.
Check |
Why it matters here |
|---|---|
SOR |
Every record the agent reads or writes comes through this connection. |
Mailbox |
Binds the agent to a queue of real mail. One instance per mailbox. |
Functions |
An unselected function falls back to platform defaults rather than being configured. |
Autonomy level |
Decides which capability sections exist at all. |
Checks worth making yourself
Wizard steps 4 and 5 are not available, so there is no test run. Behavior is verified against live mail after activation, which makes this the last inexpensive moment to catch a wrong binding.
Check |
Why it matters here |
|---|---|
The system of record is the intended environment |
It cannot be changed after creation. Confirm the environment before you activate the agent. |
The integration user has the permissions the configuration assumes |
A read-only user cannot post, whatever the capability settings say. |
The mailbox is the one you meant |
A mailbox is bound to one instance, and the choice shapes every task that follows. |
Check the authorization boundary first
This check and the three that follow are on the AP Helpdesk tab of the Define capabilities step. Contact Management, Processing Rules, Email Categories, Assignee Hold, Task SLA and Email Archiving do not appear on the AP Invoices tab, so skip all four on an agent that does not run AP Helpdesk. Sender classification governs which records the Agent Co-worker can reach, so it comes before anything else.
Open Contact Management and confirm the External contact count looks right for your vendor base. A count far below expectations points to a synchronization problem upstream rather than a configuration error.
Review the Internal contacts list. Every address here, and every address on a Relaxed Matching domain, receives broad access across vendor records and internal AP data.
Confirm the Relaxed Matching domains are domains your organization controls.
Review the System Sender list for entries broader than intended. A domain entry trusts every address on that domain; a wildcard subdomain entry trusts every subdomain.
Check the Unknown Sender list for high-volume senders not yet classified. These are where automation is being lost.
For more information, see Contact Management.
Check the Processing Rules on the AP Helpdesk tab
Open Processing Rules on the AP Helpdesk tab of Define capabilities and read the enabled rules end to end. Behavior comes from the aggregate, so reading them in order is the only way to see it.
Confirm a rule exists for every action that must never happen autonomously.
Confirm the enabled and disabled state of each rule matches intent. Disabled rules stay in the list with muted styling and can be overlooked.
Where a conflict warning was raised at authoring time and overridden, confirm the override was deliberate.
For more information, see Create and Manage AP Helpdesk Processing Rules.
Check what the Agent Co-worker will say
Open Email Categories and confirm the enabled set matches the traffic your mailbox receives. Usage counts show which categories actually arrive.
For each category the Agent Co-worker will reply to, open the reply preview and read the composed reply as the vendor would receive it.
Check how missing data reads. Where a value is not on file, the reply says so.
Confirm the response guidance does not instruct the Agent Co-worker to state anything you cannot stand behind.
Check where the work goes
Confirm the Assignee Hold strategy matches how your team works threads.
Confirm the Task SLA At-Risk value is lower than the Overdue value.
Confirm the timezone is set explicitly if your team spans regions.
Confirm the archiving configuration, including the Auto-Archive day value.
Activate the agent
Select Run. On success the wizard clears, and the instance appears in the agent’s Running Instances list with status Running and replies Enabled. The list can take a moment to pick the new instance up, so refresh the Agent Co-workers page before concluding the activation failed.
Activation can fail on the first attempt. Run may return “Something went wrong activating the agent. Please try Run again.” A second Run, with no change to the configuration, usually succeeds. If the second attempt also fails, treat it as a configuration problem rather than a transient one and re-check the connection and mailbox.
Archiving and the audit log
Email archiving is an AP Helpdesk setting. Emails or email threads with NEW or OPEN tasks cannot be archived, so archiving settings never hide work in progress.
Every agent action is recorded in an audit log covering classification, record queries, draft generation, file transfers, and autonomous sends. It confirms after the fact whether the Agent Co-worker behaved as configured on messages nobody reviewed at the time.
After activation
With no test step, the first days of live traffic are the test: which rules are applying, what agent-drafted replies look like before they send, and whether tasks are routing to the people you expected.
Activation risks and safeguards
The last check before an Agent Co-worker starts working live traffic.
It matters more than the name suggests. Wizard steps 4 and 5, Exception treatment and Try out coworker, are skipped, so there is no test run. This review is the only opportunity to catch something before it reaches a vendor.
What to confirm
All six belong before the Agent Co-worker starts working live traffic, and all six are Super Admin work. The first three are set outside the agent, in the connection and the system of record; the last three are the agent’s own configuration, set in the wizard.
Table 1. Pre-activation checks
Check |
Why it matters |
|---|---|
The system of record is the intended environment |
This cannot be changed after creation. Correcting the environment after creation means building the agent again. |
The integration user holds the permissions the agent needs |
A read-only integration user cannot post invoices, whatever else is configured. |
Vendor master, contacts and purchase orders have synchronized. The connection sync runs every four hours by default, and Purchase Orders and Supplier Contracts can also be enrolled in the hourly refresh on the connection. Vendor master cannot. |
Gaps surface as cautious behavior rather than errors, often weeks later. |
Enrollment covers the intended vendors and no more |
Enrollment bounds the agent at every autonomy level. |
Approver rules cover every entity in scope |
Approver rules are set on the AP Invoices Approvers page and name the person who signs off on a vendor bill before the Agent Co-worker writes it to the system of record, scoped by entity and optionally vendor and account number. An entity with no rule gets no default approver: the approver field is left blank and annotated “No approver rules configured”. |
Scoped autonomy names the right vendors |
Only the vendors you select are written autonomously; everything else stays under review. |
Sender classification is correct, particularly Internal |
Internal is the widest access scope the agent grants. |
What to watch in the first fortnight
The Unknown sender group. Legitimate vendors landing there means contacts are incomplete.
Outbound replies. If a processing rule uses the Send reply action, read a sample of what the Agent Co-worker actually sent.
The audit log records the reason behind each decision, so this is a reading exercise rather than a guessing one.
Activation is not irreversible
An agent can be paused, and everything except the system of record can be reopened and adjusted. So the risk of activating is bounded, provided the autonomy level is not set higher than you can supervise.