AI assistant for Business Central
Ask Business Central a question in plain language and get an answer built from live ledger entries — without opening a report, setting a filter or waiting for someone in finance.
Short answer: an AI assistant is the read-only half of AI for Business Central. It answers, analyses, searches, explains and summarises. It changes nothing in your database, which is why it is almost always the right place to start — the failure mode is a wrong answer a person can check, not a wrong posting somebody has to reverse.
What it actually answers
The useful test is not whether a question sounds impressive but whether the data behind it exists and is entered consistently. These four categories cover most of what companies ask in the first six months.
| Question type | Example | What it reads |
|---|---|---|
| Look-up | “What is the credit limit on Nordwind and how much of it is used?” | Customer card, customer ledger entries |
| Aggregation | “Revenue by item category for Q3 against Q2” | Posted sales invoice lines, item categories |
| Filtered list | “Customers with no posted sales document in 90 days” | Customer ledger entries, sales headers |
| Decomposition | “Why did gross margin fall this month?” | Value entries, item ledger entries, dimensions |
The first three are straightforward: the assistant translates a question into API queries and formats what comes back. The fourth is the one worth scrutinising, because “why” questions require a definition. Gross margin has at least three defensible calculations in Business Central depending on whether you use the cost at posting, the adjusted cost after cost adjustment runs, or a standard cost. An assistant that picks one silently will disagree with your management report, and the person reading it will conclude the AI is broken when in fact nobody told it which definition you use.
This is why decomposition questions are agreed with the finance team before the build rather than after. It is also why the same question can be cheap on one Business Central setup and expensive on another.
How it reaches the data
The assistant does not connect to your database. It calls a defined set of operations through the Business Central API, under a service or delegated identity that you create, inside the permission sets you assign. If that identity cannot open a page in the web client, the assistant cannot read the table behind it. The integration layer holds that list of permitted operations and is where the security work sits.
Nothing is copied into a separate warehouse. Each question is answered against live data at the moment it is asked, which means the assistant is always current and also that it inherits every gap in your setup — see the limitations below.
Where it gets things wrong
Four failure modes, in the order we see them:
- Inconsistent dimensions. If a dimension was mandatory from 2024 but optional before that, any analysis crossing that boundary is wrong in a way that looks plausible. The assistant cannot detect this; a person who knows the history can.
- Ambiguous business terms. “Revenue” might mean invoiced, might mean shipped, might exclude intercompany. Every such term needs a definition fixed in the build.
- Cost timing. Margin questions asked before cost adjustment has run return provisional numbers. The assistant should say so rather than answer confidently.
- Questions the data cannot answer. If nobody records lost-sale reasons, no assistant can tell you why you lost the deal. The correct behaviour is to say the data is not there.
That last point is worth stating as a requirement rather than a caveat. An assistant that guesses is worse than no assistant, because it costs trust the first time somebody checks it. We build them to refuse.
Assistant or agent
An assistant reads. An AI agent can additionally chain steps and write back — create records, update fields, start approvals. The two are frequently sold as one product and the security conversation is entirely different. Most companies get more value than they expect from the read-only version, and the honest sequence is to prove the answers are trusted before granting anything write access.
Permissions and security
The assistant runs as an identity with a permission set, so it is restricted exactly as a user with that permission set would be. Company scope applies. Access is logged. If you need different people to see different data, that is not solved in the prompt — it is solved by having the assistant act under the asking user's own identity rather than a shared service account, which is a design decision made up front because it affects the whole integration. The wider picture is covered in security, permissions and data residency.
What it costs
An assistant over a defined set of Business Central data starts from $3,000, after an AI Discovery from $750. The variables are how many question types you need, how much definition work your calculations require, and whether your dimensions are consistent enough to support the analysis you want. Model usage and hosting are running costs billed by the provider, separate from the build.
Frequently asked questions
Is an AI assistant the same as Copilot in Business Central?
No. Copilot is Microsoft's built-in feature set, licensed with the product and covering scenarios Microsoft chose. A custom assistant connects over the API and answers the questions specific to your company — your calculations, your terminology, your data. Running both is normal.
Can it see data the user is not allowed to see?
Not if it is configured correctly. The assistant operates under an identity with a permission set, and Business Central applies those permissions the same way it does for a person. If you need per-user restrictions, the assistant should act under each asking user's identity rather than a shared service account — decide this before the build, because it shapes the integration.
What happens when it does not know the answer?
It should say so. We build assistants to distinguish between a query that returned nothing and data that does not exist, and to report both rather than produce a plausible number. An assistant that guesses loses the users' trust the first time somebody verifies it.
Does it work with our customisations and AL extensions?
Yes, if the fields and tables are exposed through the API. Custom fields added via AL extensions can be published as API pages. If something your questions depend on is not exposed, publishing it is a small piece of AL work rather than a blocker.
How current are the answers?
Live. Each question queries Business Central at the moment it is asked; nothing is cached into a separate store. The exception is figures that depend on cost adjustment having run, which the assistant should flag as provisional rather than silently report.
How long does it take to build?
Typically two to four weeks after discovery to a working prototype against a sandbox environment, then testing against real questions before production. The time is spent on defining calculations and testing failure cases, not on connecting the API.
Ready to modernize your business?
Leave your details and we'll send the details of a free process audit. We reply within one business day.
or write directly: info@nbcs365.com · +380 98 107 5878