Why AI stalls in groups
An AI pilot can look good in one entity and stall at the second. The reason is rarely the model. The same word means two things in two subsidiaries. The charts of accounts were harmonised on paper only.
An AI tool does not fix that. It reads both versions and answers with confidence. The faster it works, the faster it spreads the confusion.
So the first question is never “which AI tool”. It is “which decisions have to be made before any tool can be trusted on this data”. That is what root cause work is for, and it is where AI readiness starts.
What an AI agent does well on ERP data
Reading, cross-checking and preparing. On our own Business Central, an agent reads the month’s time entries, groups them, flags the lines that do not add up and prepares the corrected lines. Saving them is a separate step, run on purpose.
The same pattern fits the checks finance runs every month: entries to compare, totals to tie out, exceptions to flag before the close.
One warning from experience. An agent that filters the wrong field reports “nothing found” with the same confidence as a real empty result. So every number an agent reads gets checked by a second route before anyone acts on it.
The full rules are in our article on AI agents inside Business Central.
The model prepares. A person commits
The Business Central MCP server, the standard way to connect an AI to Business Central, is read-only by default. We keep it that way until a write path has been designed.
When AI has to write, it writes into a holding area first. A separate, deliberate action moves the work into your real data. Before that, the agent checks which company it is in. Tests run in a test company, never in production.
Every action runs under a named user, with that user’s permissions. No shared “AI” account. That is the audit trail your finance director and your auditors will ask for.
Not only Microsoft. Not only the ERP
Copilot is a good answer to some questions. It is not the only one. MCP, the standard Business Central uses to connect to AI, is open: Microsoft’s own tools use it, and so do other AI clients.
That keeps you free to pick the model per job, and to change it later, without rebuilding the ERP side.
Some of the value sits outside the ERP. Knowledge lives in people’s heads, emails and notes. An internal AI clone keeps it: one per person, owned by the company, still there when someone leaves. We built one for ourselves: Klyr.
Where to start: AI readiness
A written diagnosis before anything is connected: where AI creates real value on your ERP data, how to put it in place, and what must be decided first.
Sometimes the honest answer is that the data has to be fixed before any AI goes near it. We will tell you that.
To see where you stand, start with the clarity form.
