What to look for in ERP selection when AI is the yardstick?

When AI enters the ERP selection, five verifiable criteria count: is the AI firmly included in the standard or a surcharge add-on? Does the vendor demonstrably deliver in short, reliable cycles? Does the AI work on the standard data model instead of on special layers? Does it respect the permission model? And are there productive agents rather than announcements? Added to this is the question no vendor answers: whether your own processes and data hold, because the best system AI helps nothing on an unclarified foundation.

Why AI changes the selection logic

Classic ERP selection compares feature lists: can system A do what system B does? AI shifts the question, because what matters is not today's feature level but the development curve behind it. A system that advances its AI in the standard with fixed release waves grows on its own. A system whose AI consists of acquisitions and surcharge modules grows only against invoices.

The classic selection criteria remain valid: process fit, partner quality, total cost. The AI criteria come on top; they replace nothing.

The five AI criteria, phrased verifiably

1. In the standard rather than bolted on: is the AI part of the normal licence or a surcharge catalogue? In Business Central, Copilot is included in the Essentials and Premium licences. Ask every vendor for exactly this information, in writing.
2. Verifiable release cadence: how often did AI features actually arrive in the past two years, not as a roadmap but as release notes? Two waves per year with documented additions are a verifiable yardstick.
3. AI on the standard data model: does the AI work on the same tables as the system, or does it need special layers and exports? Everything that sits outside the standard remains invisible or maintenance-heavy for the AI.
4. Permissions without exception: does the AI only see what the signed-in user may see? Have this shown in the demo with a restricted test user, not with the admin account.
5. Productive agents rather than announcements: which agents are generally available today and in use at reference customers? Announcements are marketing; availability is a criterion.

The demo traps

AI demos run on prepared data: unambiguous items, clean customers, perfect documents. Your data is different, and that is exactly where everyday value is decided. The most effective counter-check: insist on having two or three of your real, anonymised documents processed in the demo, including an ugly one.

The second trap: the demo shows the future, the contract delivers the present. Have it confirmed in writing for each function shown whether it is generally available today, in preview or announced. The difference decides your first year.

The question no vendor answers

Whether AI works for you is only half decided in the system. The other half is your processes and your data foundation: are workflows clarified, master data unambiguous, ownership named? This half cannot be bought; it has to be worked out, ideally before the system decision, because it applies to every system.

That is exactly what a pre-selection Take stock is for: it quantifies what your own house brings along and makes the vendor comparisons reliable in the first place. What AI then costs in operation and when it pays off, we have calculated separately and honestly.