Blog

Running AI agents after go-live: what do you need to see when things go wrong?

An agent is only manageable when you can see, for each task, what went in, which boundary applied, what it did and where a person stepped in.

A workspace with a monitor and a laptop side by side

An AI agent can work perfectly on Monday and respond differently on Thursday without anyone touching a setting. A source file has gained a new column. A supplier has changed an API. A new model version behaves slightly differently. Or volume rises and a time limit leaves a task unfinished. You have to design the process for changes like these. They come with any software that sits between existing systems.

A supplier who delivers an agent and then walks away leaves the process owner with an awkward question: how will I know it is going wrong before a customer, the finance team or an employee does? A dashboard is only useful if the process owner can find one faulty task, investigate it and put it right.

Follow one task from start to finish

Suppose an agent answers customer questions using data from a CRM, a contract folder and an order system. For every reply you must be able to trace which request came in, which sources were consulted, which model and which version of the instructions were used, which actions were taken and what the outcome was. If the reply goes to an employee for approval, record that handover too.

That does not mean all content should sit in a log file indefinitely. Logs are subject to the privacy rules and retention periods of the process. It does mean that an authorised colleague must be able to reconstruct which inputs, sources, configuration, actions and human decisions led to the outcome. Without that reconstruction an error stays an anecdote: someone saw something odd, but nobody knows whether it was a one-off or a pattern.

Watch quality, cost and exceptions

Running an agent starts with three questions. Does the agent do the right task? Does it use a reasonable amount of resources? And do exceptions land where someone can deal with them?

You measure quality against the process, not against how human a reply sounds. In order processing, useful signals include the share of orders passed on correctly, the reasons for corrections and the time it takes for an order to reach the ERP system. For customer questions you can look at the share of replies an employee sends back, the source that is most often missing and questions that keep landing in a queue.

Cost needs its own signal. A long attachment, a retry or a fault in an integration can make a task far more expensive than intended. So measure cost and lead time per type of task. That way you see whether a change to the source or the instructions affects the process, instead of finding out from the monthly invoice.

A queue is the third signal. Do not let it disappear by pushing every case through automatically anyway. The queue tells you where the boundary is too tight, too loose or simply unclear. A recurring exception can lead to a new rule. An exception with serious consequences stays with a person, even if it rarely occurs.

Human oversight is a job, not a button

"A human in the loop" sounds reassuring, but solves nothing when nobody knows who that person is, when they should step in and what authority they have. For each boundary, record who gets the task, how quickly that person must respond and what they may decide. For a price discrepancy that might be the order desk. For a payment above an agreed amount, perhaps finance. For a fault in a source application, IT.

Also give that person a way to stop an external action before it runs, or to divert it to a safe manual process. If an action has already run, there must be a way to correct or escalate it. An approval button without context turns oversight into extra clicking. So show the source data, the reason for the proposed action and the consequences of approving or rejecting it.

Make running the agent part of the agreement

After go-live, models, documents, suppliers and internal rules keep changing. That is why an agent needs maintenance, just like an integration or a finance application. Agree in advance who investigates anomalies, how a change is tested and what reporting the process owner receives. A monthly report need not be long. It should show what was processed, where intervention was needed, what it cost and which decisions need attention.

Whether an agent can work on its own matters less than whether the organisation can see when the agent should stop. Without that, the agent is still a demo with access to real systems.

Three ways to start

Want to set out what a live agent must show operations and IT?

Start small, with one process that comes up every week. Choose how: from half an hour behind the scenes to a session at your offices.

01

Look behind the scenes

Free

Half an hour, by invitation

See everything we build running in our sandbox, the Proeftuin. No sales pitch, unless you ask for one.

Request an invitation

02Recommended

Introductory call

Free

Half an hour on Teams

You tell us what you already do with AI. We tell you what fits, or whether it is still too early.

Call with Marcel Logtenberg

Book a call

03

Get started

From €495

At your offices, excluding VAT

An AI quick scan, an AI assessment or an AI assessment plus inspiration session, at a fixed price.

Get started