AI Readiness for an ERP Group: The Decisions to Make Before Connecting Anything
In a multi-entity group, AI readiness is not a culture score or a tooling inventory. It is a short list of decisions, each with an owner: do the entities record the same thing the same way, what may the AI read, what may it write and under whose name, where may the data be processed, and how will a wrong answer be caught. Until those are made, any AI you connect reads your inconsistencies and repeats them with confidence.
Why do generic AI readiness assessments miss the point in a group?
Because they score the organisation as one block: strategy, skills, culture, infrastructure. In a single company that is a fair picture. In a group, the gap that matters sits between the entities. The same word means two things in two subsidiaries, and nobody notices until a report has to add them up.
Business Central makes this concrete. Global dimensions are company-defined: each company chooses its own two global dimensions and up to eight shortcut dimensions on General Ledger Setup. Five subsidiaries can legitimately run five different analytical schemes. And Microsoft warns that changing a global dimension later means updating every entry posted with it, which can take time, affect performance and lock tables. The cheap moment to align how entities record things is before you build on top of them, not after.
An AI layer is one more thing built on top. A readiness assessment that does not look at this gap is grading the wrong thing.
Do your entities record the same thing the same way?
This is the first decision, and it has two halves.
The accounts. Business Central can consolidate companies whose charts of accounts differ, but only once someone has mapped each posting account to an account in the consolidated company, through the consolidation debit and credit account fields. Dimension values are mapped the same way, through a consolidation code. That mapping is the closest thing a group has to a shared vocabulary. If it does not exist, or exists only for the year-end consolidation, an AI asked about a group margin has no rule to follow. Decide which mapping is the reference, and who owns it.
The master data. Customers, vendors and items: shared across entities, or local to each? Business Central can synchronise master data from one source company to subsidiaries, and Microsoft is precise about the limits: the subsidiaries pull from the source, the companies must be in the same environment, only non-transactional tables should be synchronised, and Microsoft does not recommend synchronising in both directions because of conflicts and overwrites. If the same customer exists under three numbers in three entities, an AI asked for the group’s exposure to that customer will answer for one of them. Decide the source of each master table, and who owns it.
What may the AI read, and what may it write?
Decide this per use case, not once for the whole group.
On Business Central, the standard way to connect an AI is the MCP server, and its configurations decide which API pages and queries are exposed and which operations each one allows. Microsoft’s documentation states that the built-in default configuration gives read-only access to eligible API pages and queries that the signed-in user is allowed to use. Creating, modifying, deleting and running actions must be switched on explicitly, configuration by configuration.
That gives a group three decisions to write down before any agent goes live:
- Which data each use case reads. A configuration per family of questions, not one configuration that sees everything.
- Whether it writes at all. And if it does, where: a holding area and a deliberate action to commit, never straight into posted data.
- Under whose name. Every call runs under the signed-in user’s identity and permissions. No shared AI account, so every change carries a person’s name.
How those choices hold up in front of an auditor is covered in our article on AI governance in finance.
Where may your data be processed?
This one is often left to IT, and it should not be. It is a decision for whoever answers for data protection in the group.
For Business Central’s own Copilot and agent capabilities, Microsoft uses Azure OpenAI resources it manages. For environments hosted in France or elsewhere in Europe, processing can occur across geographies within the EU Data Boundary, and the Allow data movement setting on the Copilot & agent capabilities page is on by default. Turning it off makes the affected capabilities unavailable. Microsoft also notes that AI features added by other publishers, through extensions or marketplace apps, define their own processing regions, and sends you to the publisher for details. And this setting does not govern Microsoft Copilot in Business Central, the experience available from version 29, which falls under Microsoft 365 Copilot commitments.
So the readiness question is not only “Copilot or not”. It is: for each model you intend to use, where is the data processed, and who signed off on it? A non-Microsoft model connected over MCP comes with its provider’s own terms, which have to be read with the same care.
Which use case first?
One the team already runs by hand and can describe precisely. Which report, which filter, which tolerance, what counts as an exception. If nobody can state the rule, no model can run it, and no one will be able to tell whether it ran it correctly.
Month-end checks are a good candidate in most groups, because they repeat with the same rules every period. We describe which close tasks a model can prepare and which must stay with a person in our article on AI in the month-end close.
The opposite also holds. A use case that depends on a decision nobody has made, such as which entity is right when two intercompany balances disagree, is not an AI use case yet. It is a decision waiting for an owner.
How will you know when the AI is wrong?
Decide it before the first answer comes back. Two mechanisms matter:
- A second route for every number. An agent that filters on the wrong field reports “nothing found” with full confidence. Every result it gives is compared with a figure obtained another way before anyone acts on it.
- A trace of what the agent did. From version 28.0, Business Central telemetry records each MCP tool call, with the user, the tool, the configuration and the client that made the call. Changes to MCP configurations are recorded too.
And one more owner: who can switch it off. In Business Central, administrators can deactivate individual Copilot and agent capabilities and grant or deny access per user. Name the person who will do it, and on what signal.
What does a readiness answer look like?
Not a score out of five. A table of decisions: the question, whether it is decided, who owns it, and by when. Shared account mapping, master data sources, read and write scope per use case, named users, data location per model, verification and switch-off.
Sometimes the honest outcome is that the data has to be fixed before any AI goes near it. That is still a useful answer, because it tells the board what to fund first.
The Asio Services way
We treat AI readiness as root cause work applied to AI: find the decisions a model would otherwise inherit unmade, name an owner for each, and only then pick the first use case. It is how we approach AI for multi-entity groups, Copilot when it fits, another model when it fits better, the same rules underneath. If your board is asking for an AI plan and you are not sure your entities are ready for one, start with a clarity session.
FAQ
What is an AI readiness assessment for an ERP?
A review of whether your ERP data and rules can support AI before you connect any. In a group, it focuses on decisions: shared account mapping, master data ownership, what the AI may read and write, under whose name, where data is processed and how errors are caught.
Do we need perfect data before using AI on our ERP?
No. But AI reads the data as it is. If entities record the same thing in different ways, the model repeats the difference. The assessment tells you which gaps matter for the use case you have in mind.
Is AI on Business Central read-only by default?
Over the Business Central MCP server, yes. The built-in default configuration gives read-only access to eligible API pages and queries the signed-in user may use. Writing must be enabled explicitly in a configuration.
Who should own AI readiness in a group?
Not IT alone. Finance owns the account mapping and the checks, operations owns master data, and whoever answers for data protection signs off on where data is processed. Each decision needs a named owner.