Short answer: AI does not "analyse finance" in the abstract. It runs specific queries against specific Business Central tables and assembles an answer from them. So the useful lens is not capabilities but six typical questions — and what each of them demands from your setup.
"Who owes us and how much"
Reads: customer ledger entries, open documents, payment terms from the customer card.
The simplest class of question and the one where AI delivers immediately. An ageing report does this too — the difference is that an assistant adds context without extra configuration: not just "62 days overdue" but "62 days overdue, from a customer who paid on time for two years, with two open orders currently in the pipeline".
Where it breaks: almost nowhere. This is the most reliable category. The one requirement is that payment terms are filled in — otherwise "overdue" is calculated from the document date and produces nonsense for anyone on individually agreed terms.
"Why did margin fall"
Reads: value entries, item ledger entries, posted document lines, dimensions.
The most valuable question and the hardest. The answer has to be a decomposition rather than a single figure: how much of the drop came from purchase price movement, how much from discount above policy, how much from a mix shift toward low-margin items.
The main caveat in this article. Gross margin in Business Central has at least three defensible calculations — cost at posting, adjusted cost after the cost adjustment routine, or standard cost. If nobody states which is yours, the assistant picks one and disagrees with your management report. Agree this before the build.
Where it breaks: inconsistent dimensions, and figures pulled before cost adjustment has run. The assistant should flag the second itself.
"What happens to cash in 60 days"
Reads: open receivables with payment behaviour, open payables, sales and purchase orders.
Here AI does what a person does in Excel: takes commitments and expectations and spreads them across dates. The advantage is that the forecast can rest on actual customer behaviour rather than contract terms. A customer on 14-day terms who consistently pays on day 35 should appear in the forecast on day 35.
Where it breaks: some obligations live outside Business Central — payroll, taxes, loan schedules. A forecast built from ledger data alone will be confidently optimistic. Those inputs get added deliberately.
"Which postings look unusual"
Reads: G/L entries, vendor ledger, history per counterparty and account.
An unusual amount for a vendor, a posting to an account that rarely receives one, a document posted outside working hours, a repeated invoice number from the same supplier. The assistant produces a list of candidates for review.
Where it breaks: in expectations. This is pattern detection, not fraud detection. False positives are permanent, and the value is in shortening a controller's review rather than replacing it.
"Compare this month to last and explain"
Reads: G/L entries by category, dimensions, budgets.
This is the commentary somebody writes by hand every month for the management report. AI assembles it from entries: what changed, by how much, driven by what.
Where it breaks: if the chart of accounts or G/L categories changed during the period, a naive comparison shows a variance that did not happen. The assistant needs to know about such changes, or they have to be excluded from the comparison. The underlying process: period close in Business Central.
"How much did we make on this customer"
Reads: sales lines, value entries, dimensions, and cost allocation where it exists.
It looks simple and is actually a question about accounting policy. Is customer profit gross margin? Including delivery? With indirect costs allocated? Each variant gives a different number, and all three are correct in different conversations.
Where it breaks: in the definition. This is the second place after margin where the assistant has to be taught your logic rather than left to the model's judgement.
When the answers can be trusted
A four-point check. If all four hold, financial analysis will behave predictably.
- Dimensions are filled in consistently and the rules did not change mid-period — or the change is known and accounted for.
- Posting groups are disciplined: similar transactions do not scatter across accounts depending on who entered them. See posting groups.
- Cost adjustment runs regularly, and the assistant marks provisional figures as provisional.
- Key terms are defined in writing: margin, revenue, overdue, active customer.
If any point fails, the honest sequence is to fix the setup first. That produces better reports whether or not AI is ever involved. What this looks like as a service: AI for finance in Business Central.
Where this sits against reports and dashboards
A question worth settling early, because it decides what you should actually be buying.
A Power BI dashboard answers questions you anticipated when you built it, and answers them at a glance, identically every time. That is exactly what you want for the numbers you monitor weekly.
AI answers the follow-up question nobody built a visual for. The dashboard shows margin down 2.9 points; the useful next question is which territory, then which salesperson, then whether it is discount or mix. Each question emerges from the previous answer, and that chain was previously impossible — because every report cost somebody an hour.
Most companies want both. What they rarely need is a third dashboard: if a question repeats unchanged every month, it is a report, and building an assistant for it is the expensive way to get one.
Frequently asked questions
Will the AI figures match our reporting?
Only if the calculations are defined the same way, which is why definition is part of the build rather than an assumption. Margin and customer profit each have several defensible definitions in Business Central; without agreement, a discrepancy is guaranteed and will look like an AI error when it is actually an undefined term.
How much history is needed?
For balances and aggregation, whatever you have. For trends and seasonality you need two to three years, otherwise the assistant will describe a coincidence as a pattern — convincingly.
Can AI find a problem we did not ask about?
Partly — that is what anomaly detection does: it surfaces what differs from the norm. But that is a list of candidates for review, not a conclusion. Deciding whether it is a problem stays with a person.
Do we need Power BI for this?
They are different tools. Power BI shows the metrics you decided to monitor, at a glance. AI answers the follow-up nobody built a visual for. Most companies benefit from both.
Who is responsible if AI gives a wrong figure?
Whoever made the decision. That is why the assistant must show what the answer is made of and let you trace it to specific entries. If a figure cannot be verified, it should not be used for decisions.