Who maintains our AI features after go-live?

AI features in your ERP need a simple operating routine after go-live: a named owner per function, a fixed review date per release wave, a small set of test cases for the automated transactions and a spot-check quota in daily business. Microsoft changes Copilot and agents with two release waves per year. Without this routine, every AI function ages within months: it keeps working, only nobody knows how well anymore.

Why AI features age differently from modules

An accounting module calculates today as it did five years ago. An AI function does not: its models, interfaces and capabilities change with the release waves, and its behaviour depends on your data, which changes as well. Together this means: the state on go-live day is not a permanent state but a snapshot.

That is a reason for routine, not for worry. The same movement continuously brings new capabilities that are already included in the licence price. Companies with an operating routine pick them up; everyone else learns from the grapevine that something no longer works.

The operating routine in four elements

1. One owner per function: the same logic as with processes. The owner knows the function's purpose, checks it after releases and decides whether new capabilities are activated.
2. One review date per release wave: twice a year, in the calendar, with Microsoft's release notes as the agenda. A few hours that prevent changes from slipping through unnoticed.
3. One set of test cases: for every automated transaction three to five typical cases plus the known edge cases. Run through once after each release, note the result.
4. A spot-check quota in daily business: a fixed small share of agent results is checked continuously. If the error rate rises, that is the early warning long before anyone complains.

The effort, honestly quantified

For a midmarket company with a handful of active AI functions, the routine amounts to a few person-days per year: two review dates, ongoing spot checks in the range of minutes, occasional readjustment. That is less than a single month of unnoticed errors costs in clean-up.

Whoever cannot or does not want to staff the routine internally brings it in as support: governance and delivery management keeps exactly this order, with steering on your side. What matters is not who runs the routine but that it exists and is named.

The connection with the template and the data foundation

Internationally, the maintenance question grows: when several countries work on a shared core, the AI review belongs in the same release routine as the global template, otherwise every country tests on its own or none does. And as everywhere: the review results are only as meaningful as the data foundation the functions work on.