Short answer: AI over Business Central is not a "smart ERP". It is three separate things: Microsoft's built-in Copilot, a custom assistant that reads your data and answers questions, and an agent that performs approved actions. The first is switched on; the other two are built. The biggest cost variable is not the model — it is the state of your setup.
Three levels of AI over Business Central
Most of the confusion in this area comes from one word covering three different things. Separating them makes the rest of the decision straightforward.
Level 1. Microsoft Copilot
Ships with the product. Covers a defined set of scenarios Microsoft chose: analysis assistance, bank reconciliation proposals, item description generation, forecasting. It is configured, not developed. Our detailed look at it: Copilot in Business Central.
Always start here. It is the cheapest ground to cover, and if it answers your question, the rest of this guide is unnecessary.
Level 2. A custom AI assistant
A separate solution connecting to Business Central through the API, answering questions in natural language. Read-only. The point is that it knows your definitions: your margin, your terminology, your dimensions. Copilot does not know these and structurally cannot — it is a product feature, not yours.
Level 3. An AI agent
Everything above plus actions: create a record, update a field, start an approval. It executes only what you have explicitly permitted, inside a permission set. Boundaries and controls are covered on our page about AI agents for Business Central.
What AI actually does with ERP data
The useful test is not whether a question sounds impressive but whether the data exists and is entered consistently. On that basis, questions fall into four groups.
| Type | Example | Difficulty |
|---|---|---|
| Look-up | "What is this customer's credit limit and how much is used?" | Easy |
| Aggregation | "Revenue by item category this quarter against last" | Easy |
| Filtered list | "Customers with no orders beyond their own usual interval" | Moderate |
| Causal | "Why did gross margin fall this month?" | Hard |
The first three are translation from a question into API queries. The fourth is where projects diverge from expectations, because a "why" question requires a definition. Gross margin in Business Central has at least three defensible calculations depending on whether you use cost at posting, adjusted cost after the cost adjustment routine, or standard cost. If nobody specifies which one is yours, the assistant picks one and disagrees with your management report — and everyone concludes the AI is broken.
Practical consequence: causal questions are agreed with the finance team before the build, not after. That is an hour at design time against months of lost credibility.
Four questions before you start
Worth answering before any technology conversation. Most projects that should not begin are filtered out here.
- What specific question are you asking today, and what does it cost? Not "we want AI" but "every month someone spends two days assembling an overdue receivables report split by salesperson". If you cannot write that sentence, the project is premature.
- Who will be asking? An assistant used by one person is almost always cheaper to replace with a report. Value appears when several people ask, and ask differently.
- Where does the data live? If the answer needs pipeline data and the pipeline lives in another CRM, this is an integration project, not an AI project.
- Is this actually a report? The honest answer is often that you need a report, a Power BI dashboard, or a configuration change. Those cost less and break less often.
What it needs from your Business Central
The most important part of this guide and the one usually skipped in vendor material. Answer quality is determined by the state of your setup, not by the model.
Consistent dimensions
Analysis by department, project or region only works if dimensions are filled in reliably. The classic trap: a dimension became mandatory at some point and was optional before. Any analysis crossing that date is wrong — and looks entirely plausible. No model detects this.
Posting group discipline
If similar transactions land in different accounts depending on who entered the document, aggregate figures will not reconcile — not in a report and not in an assistant's answer. The mechanics: posting groups in Business Central.
Cost adjustment run regularly
Anything touching margin or inventory value is provisional before cost adjustment runs. The assistant should flag this rather than present it silently as final.
Fields published through the API
If some of the data you need lives in custom fields added by AL extensions, those need publishing as API pages. Small work, but worth discovering at the start rather than mid-build.
Cost and timelines
| Stage | Order of cost | Timeline |
|---|---|---|
| Discovery: use cases, data, permissions, architecture | from $750 | 1–2 weeks |
| Assistant over a defined data set | from $3,000 | 2–4 weeks to prototype |
| Agent with approved actions | from $6,000 | longer, due to testing |
| Solution around a specific process | from $10,000 | depends on scope |
Running costs are separate: model usage and hosting for the integration layer, billed by the provider. The order of magnitude here is rarely decisive, but it needs budgeting.
Agents cost more than assistants for one reason — testing. Verifying that a system refuses what it should refuse takes longer than verifying that it answers what it should answer.
Five common mistakes
- Starting with an agent. First prove the answers are trusted. Write access comes after that, not instead of it.
- Not defining terms. "Revenue" might mean shipped, invoiced, or either excluding intercompany. Every such term gets fixed in the build.
- Letting the assistant guess. An assistant that produces a plausible number instead of "no data" loses trust the first time somebody verifies it. This is a build requirement, not a preference.
- Treating the prompt as a control. The security boundary is the Business Central permission set and the list of available operations, not the wording of an instruction.
- Building analysis on top of a mess. If dimensions are inconsistent, fix the setup first. That produces better reports whether or not AI is involved.
If after this guide you know which question you want to close, the next step is the AI for Business Central page, with solutions and process.
Frequently asked questions
Will AI replace our accountant or finance manager?
No. It removes data gathering — the part that takes time but requires no judgement. Decisions about credit terms, customer risk and accounting policy stay with people. The realistic effect is not headcount reduction but the same people answering more questions, faster.
Do we need Copilot before building custom AI?
They are independent, but start with Copilot: it is already in the product and costs a fraction of a custom build. If it covers your question, you do not need a custom solution. Most companies eventually run both.
How much data does this need?
For look-ups and aggregation, whatever you have. For trend and seasonality analysis you need two to three years, otherwise the assistant will describe a coincidence as a pattern. Consistency of dimensions matters more than volume.
Is this safe for financial data?
The AI operates under a Business Central identity with a permission set and sees exactly what that identity would see. Access is logged. The separate question is which model provider processes the data and under what contract — that gets settled in writing before production.
How long until the first result?
One to two weeks for discovery, then two to four for a working assistant prototype, then testing against real questions. Most of the time goes into defining calculations and testing failure cases, not into connecting the API.
Does this work with Business Central on-premises?
Usually, but the integration path differs — web services need publishing and reachability, which is an infrastructure decision on your side. We confirm this before quoting.