AI for Microsoft Dynamics 365 Business Central
Build AI assistants and agents that work directly with your Business Central data and business processes.
Ask questions in natural language, analyse financial and operational data, and automate approved Business Central tasks with AI — inside the permissions you already run.
What AI for Business Central actually means
Business Central already holds the numbers: general ledger entries, customer and vendor ledgers, item ledger entries, sales and purchase documents, dimensions, budgets. The problem is rarely that the data is missing. It is that getting an answer out requires knowing which report to open, which filters to set and which dimension carries the split you need — so the questions that matter get asked once a month, by the one person who knows how.
AI adds a second interface to the same database. Someone asks a question in their own words, the assistant queries Business Central through the API under a controlled identity, and answers with figures that trace back to specific entries. Nothing is copied into a separate warehouse and nothing is guessed: if the assistant cannot reach the data, it says so instead of inventing a number.
That is the assistant half. The agent half goes further — it can chain several steps together and, where you have permitted it, write back: create a record, update a field, start an approval. Those are two different levels of risk and they are described separately below, because conflating them is how AI projects in an ERP go wrong.
The questions people actually ask
Not demo prompts — the questions that currently land in someone's inbox as a request for a report. Pick one to see the shape of the answer.
NBCS AI Assistant — a demo of NBCS AI capabilities for Business Central. Not Microsoft Copilot.
- Item 1140 · Pump housing€318,400
- Item 2071 · Drive belt set€204,900
- Item 1802 · Control module€171,250
- Item 1140 · Pump housing€318,400
- Item 2071 · Drive belt set€204,900
- Item 1802 · Control module€171,250
- Nordwind GmbH · 62 days€38,200
- Kessler & Co · 41 days€24,700
- Delta Logistik · 33 days€18,900
- Purchase price increases−1.6 pt
- Discount above policy−0.8 pt
- Mix shift to low-margin items−0.5 pt
- No inbound order at all6 items
- Inbound arrives too late11 items
- Longest exposure19 days
- Combined prior-year revenue€412,000
- Previously ordered monthly12 customers
- Still carrying open credit9 customers
Where AI earns its place in Business Central
Six areas where the question is asked often enough, and the data structured enough, that an assistant changes how the work is done.
AI for finance
Receivables and payables analysis, cash flow, financial reporting, profitability by dimension, and anomaly detection across ledger entries.
AI for sales
Sales analysis by customer, item and territory, performance against plan, customer inactivity detection and opportunity review.
AI for inventory
Stock analysis, slow-moving and dead stock, stockout risk against inbound orders, turnover trends and purchasing recommendations.
AI for reporting
Reporting by question rather than by report object: management summaries, KPI commentary and trend detection written in plain language.
AI business assistant
A single place to ask about Business Central: retrieve a record, explain what a number is made of, summarise a period, find the document behind an entry.
AI agents
Multi-step work with write access where you have granted it: create and update records, start approval workflows, drive a process to a defined stopping point.
Detail pages: AI assistant · AI agents · Finance · Sales · Inventory · MCP and integration
AI assistant or AI agent
These are sold as the same thing and they are not. One reads. The other acts. The security conversation is entirely different.
AI assistant
Read-only by design. It answers, analyses, searches, explains and summarises — and that is the whole surface. The worst outcome of a bug is a wrong answer, which a person can check against the source entry.
AI agent
Everything above, plus the ability to reason across several steps, call tools and execute actions you have approved. Write access is granted per operation, not as a general capability.
What we do not promise
Unattended autonomy over your ledger. An agent operates under an identity with a permission set, inside your authentication and authorisation model, with its actions logged like any other user's. Approval steps stay where your finance policy puts them. Anything an agent can do, an auditor can see it did — and anything it cannot be permitted to do, it does not do because the platform refuses, not because the prompt asked it nicely.
How the AI reaches your data
Four layers, and one boundary that decides everything: the identity the integration runs as.
Two things worth stating plainly, because vendor material tends to skip them. First, connecting AI to Business Central does not give a model access to your ERP. It gives it access to a named list of operations, executed as a named identity, subject to every permission Business Central would apply to a human holding that identity. Second, the quality of the answers depends on the quality of the underlying setup — dimensions that were never used consistently produce analysis that is confidently wrong, and no model fixes that.
Microsoft Copilot and custom AI
Not competing options. Copilot covers the scenarios Microsoft built into the product; custom AI covers the ones specific to your company.
| Question | Copilot in Business Central | Custom AI assistant or agent |
|---|---|---|
| Where it comes from | Built into the product by Microsoft, licensed with it | Built for your environment, over the API |
| Scenarios covered | A defined set: analysis assist, bank reconciliation proposals, item marketing text, forecasting | Whatever your processes require, including ones no product could anticipate |
| Company-specific logic | Standard product logic | Your margin definitions, your approval rules, your terminology |
| Data beyond Business Central | Within the Microsoft ecosystem | Can combine ERP data with other systems you run |
| Write-back and agents | Within the scenarios Microsoft ships | Agents scoped to your processes and permission sets |
| Cost model | Licensing, per Microsoft's terms | A project, then running costs for hosting and model usage |
| Effort to adopt | Turn it on and configure | Discovery, build, integration, testing |
Copilot is genuinely useful and the honest recommendation is to switch it on first — it costs a fraction of a custom build and covers real ground. The case for custom AI starts where the product's assumptions stop matching yours: company-specific business logic, industry workflows, data access rules that differ by role, specialised finance analysis, reporting the standard objects were never designed for, agents that follow your process rather than a generic one, or systems outside Business Central that need to be in the same answer.
We are not trying to replace Copilot. We extend Business Central with the AI capabilities Copilot was not built to cover. If Copilot already answers your question, we will tell you that in discovery rather than quote a project.
What we build
Six starting points. Most engagements begin with one and extend once people trust the answers.
AI Business Central assistant
Natural-language access to Business Central information. Ask, retrieve, explain, summarise — read-only, across the data your role already covers.
AI finance agent
Receivables and payables analysis, cash flow review, margin decomposition and finance workflows that prepare work for a person to approve.
AI sales agent
Sales and customer analysis: performance against plan, inactivity detection, discount leakage, territory and item-level breakdowns.
AI inventory agent
Stock analysis and operational intelligence: coverage against lead time, slow-moving capital, stockout exposure on promised orders.
AI reporting assistant
Natural-language reporting and management summaries — the month-end narrative written from the entries rather than from last month's document.
Custom Business Central AI agent
An agent designed around one of your specific business processes, with its own tool set, permission model and stopping points.
Business Central expertise first, AI second
The hard part of this work is not the model. It is knowing what a posting group does, why the margin in one report disagrees with another, and which permission set a finance user actually holds.
We work in Business Central daily
Implementation, customisation and support are our main business. The AI practice sits on top of that, not instead of it.
We read the data model, not just the API
Item ledger versus value entries, dimensions, posting groups, the difference between an order and a posted document — the distinctions that decide whether an answer is right.
Accounting and process understanding
An assistant that computes a plausible number and calls it gross margin is worse than no assistant. We define the calculations with your finance team before building.
AL development in-house
Where the answer needs something the standard API does not expose, we can build the extension rather than route around it.
Security treated as design, not review
Identity, permission sets, allow-listed operations and logging are decided in architecture, before anything touches a production environment.
We say when AI is the wrong answer
Some requests are a report, a Power BI dashboard or a fixed configuration change. Those cost less and break less often, and we will say so.
Starting points
Ranges, not a price list. What a project actually costs depends on your Business Central configuration and the scope you choose after discovery.
Final pricing depends on scope, your Business Central configuration, the data the solution needs to reach, the permission model, the integrations involved and overall complexity. Model usage and hosting are running costs, billed by the provider and separate from the build.
How a project runs
Ten steps. The first three decide whether the remaining seven are worth doing.
- 01
AI Discovery
What you want answered, who needs to ask it, and what currently happens instead. The output is a written assessment, not a proposal deck.
- 02
Identify use cases
Candidate questions and workflows ranked by how often they come up and how much they cost today. Most lists shorten considerably at this step.
- 03
Business Central analysis
Which tables hold the answer, whether dimensions are used consistently, how your calculations are actually defined. This is where a use case is confirmed as feasible or dropped.
- 04
Solution architecture
Assistant or agent, which model and where it runs, how the integration layer is built, which identity it uses and which operations it may call.
- 05
Prototype
A working version against a sandbox environment, so the people who will use it can judge the answers before the build is finished.
- 06
Integration
Connecting to the real environment: API or MCP layer, error handling, rate limits, and the behaviour when data is missing rather than merely inconvenient.
- 07
Security and permissions
Permission sets, company scope, allow-listed operations, approval handling, logging. Tested by attempting the things the solution should refuse to do.
- 08
Testing
Answers verified against the same questions answered manually. An assistant that is right most of the time is not ready — the failure cases need to be known and visible.
- 09
Production deployment
Rollout to the real users, with the scope of what it is allowed to do written down and agreed rather than assumed.
- 10
Monitoring and support
Usage and error review, checking after each Business Central release, and extending the scope once the first set of answers is trusted.
Questions we get asked
Do we need Microsoft Copilot before adding custom AI?
No. They are independent. Copilot ships with Business Central and covers a defined set of in-product scenarios. A custom assistant or agent connects over the API and answers questions Copilot was not built for — your own margin logic, your own approval rules, data from systems outside Business Central. Most companies end up running both.
Where does the AI actually read our data from?
Through the Business Central API, using a service account or delegated user identity that you create and control. Nothing bypasses the platform: if that identity cannot open a page in the web client, the assistant cannot read the table behind it either.
Can an AI agent post documents on its own?
It can execute the actions you have explicitly permitted, inside the permission set you assign. In practice most finance workflows are configured so the agent prepares and proposes, and a person posts. Which steps run unattended is a decision made per process during discovery, not a property of the technology.
Does our data get used to train an AI model?
That depends on the model provider and the contract you sign with them, not on us. Azure OpenAI and the enterprise tiers of the major providers contractually exclude customer data from training. We put this in writing as part of the architecture document before anything connects to production.
What is MCP and do we need it?
Model Context Protocol is a standard way of describing tools to an AI model — instead of hard-coding every call, you expose a set of Business Central operations the model can choose from. It makes agents easier to extend and to audit. It is not a requirement: a narrow assistant can run on plain REST calls.
How long does a first working assistant take?
A scoped assistant over a defined set of Business Central data usually reaches a working prototype in two to four weeks after discovery. Agents that write back take longer, because permissions, approval steps and audit logging need testing against real scenarios.
Does this work on Business Central on-premises?
Usually yes, but the integration path differs. Online environments expose the standard API endpoints; on-premises needs the web services layer published and reachable, which is an infrastructure decision on your side. We check this during discovery before quoting.
What happens when Business Central updates twice a year?
Integrations built on the standard API surface survive updates the same way AL extensions do — they need checking after each major release, not rebuilding. Anything reading undocumented internals does not, which is why we stay on published endpoints.
Related reading
Copilot in Business Central: what AI actually does in an ERP
GovernanceSecurity, permissions and data residency
AnalyticsPower BI + Business Central: analytics an executive will read
SetupDimensions in Business Central
FinanceFinancial reports: how to build P&L and balance sheet
PlatformWhat Microsoft Dynamics 365 Business Central is
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