Connect an AI Agent to Dynamics 365 Business Central: Copilot vs Custom
· CompaniesAutomation
Native Copilot vs custom agents in Business Central: integration routes (REST API, OData, AL), finance use cases with payback and mistakes to avoid.
Connecting an AI agent to Dynamics 365 Business Central is one of the most rewarding ERP integrations you can build: Microsoft's ERP exposes well-documented REST and OData APIs, and on top of them an agent can read and write customers, vendors, invoices, orders and journal entries with proper security and traceability. The real decision isn't technical — it's about approach: when is the Copilot Microsoft ships inside the product enough, and when do you need your own agent operating your processes end to end?
Business Central (and its bigger sibling Dynamics 365 Finance) is the default ERP for small and mid-sized companies living in the Microsoft ecosystem, often inherited from Navision. We connect agents to our clients' ERPs every week — this article mirrors the analysis we already published for SAP and for Odoo — and Business Central is, alongside Odoo, the one with the least integration friction. Let's take it in order: what ships built in, how a custom agent connects, and which finance use cases pay first.
What AI does Business Central include out of the box?
Business Central ships Copilot features embedded in the product: bank reconciliation assistance, sales line suggestions, marketing text generation for items, natural-language data analysis and chat over product help. These are in-screen assists: useful for the user already working inside Business Central, mostly included with the license, and requiring no project budget.
What native Copilot doesn't do is operate processes: it doesn't watch the vendor inbox, chase delivery notes, decide approval routes or chain steps across systems (bank, email, CRM, portals). That's the practical border: Copilot assists the person working inside the ERP; a custom agent works the process, whoever logs in. Microsoft also pushes its agent layer (Copilot Studio and the agents announced for Business Central), which can cover middle-ground cases if your organization already pays for the Power Platform ecosystem — with the usual trade-off: license/credit-based consumption and logic tied to the platform.
How do you connect a custom AI agent to Business Central?
Through three main routes, ordered from most to least recommended:
- The standard REST API (v2.0). Business Central online exposes a REST API covering the main business entities (customers, vendors, sales and purchase invoices, orders, items, journal lines). Modern authentication via Microsoft Entra (OAuth2), fine-grained per-application permissions and audit of who did what. For 80% of agent use cases it's sufficient, and it's our default route.
- OData over pages and queries. When you need custom fields or tables the standard API doesn't cover, any Business Central page or query can be published as an OData service. It's the natural escape hatch in installations carrying years of Navision-era customizations.
- Custom AL extensions. For operations that must run inside the ERP with its business logic (complex postings, processes with custom validations), a small AL extension exposes exactly the operation the agent needs — and nothing else. More cost, maximum control.
On older on-premises installations (Navision/NAV, Business Central on-prem without APIs enabled) the menu shrinks: SOAP/OData web services that must be enabled, or file-based integration. It works, but it adds weeks to the project — better to know before budgeting.
Whichever route you take, the security model is the same: the agent gets its own application identity, its own permission set limited to the entities it touches, and its own audit trail. That's not bureaucracy — it's what lets you answer "what exactly can the AI do in our ERP?" with a list instead of a shrug.
Finance use cases: where to start with Business Central
The highest-return cases mirror what we've covered for other ERPs, with the advantage that in Business Central the integration is direct:
| Use case | What the agent does | Payback signal |
|---|---|---|
| Vendor invoices | Email capture, extraction, matching against PO and receipt, posting into BC | From 100-200 invoices/month |
| Bank reconciliation | Imports movements, proposes matches, posts the safe ones, escalates doubts | Weekly admin hours |
| Collections | Staged reminders reading due dates from BC, payment promises logged | High DSO, irregular follow-up |
| Monthly close | Close checklist, pending entries verification, accrual drafts | Close taking more than 5-7 days |
| Account summaries | On-demand customer/vendor briefing: position, risk, history, open issues | Management asking for data "for yesterday" |
They all share the pattern: the agent reads and writes Business Central through the API, also works outside the ERP (email, bank, PDFs) and leaves a trace of every action. The full map of finance cases is in AI agents in the finance department.
Microsoft Copilot or custom agent? The honest answer
Our recommendation — the same we give for SAP and Salesforce: it's not either/or, they're different layers.
- Use the included Copilot for what it already does well: assisted reconciliation, natural-language analysis, item texts. It's paid for with your license; don't rebuild that.
- Consider Copilot Studio if your company is already invested in Power Platform, the case is conversational/internal, and you accept the platform dependency and its consumption model.
- Build a custom agent when the process crosses systems (email + bank + BC + portals), needs supervised autonomy (posting, ordering, chasing), or is a differentiating process you don't want tied to a vendor's roadmap. The architecture also stays agnostic: change AI model or even ERP tomorrow and the agent adapts.
In SMB numbers: a first scoped agent on Business Central (say, vendor invoices end to end) runs €3,000-8,000; a broad finance deployment, €8,000-15,000, plus 10-20% yearly maintenance. Business Central licenses are separate and you already pay them: the agent doesn't add user seats — it connects as an application.
On recurring spend — the part of the cost that isn't project work: inference for a typical finance agent on Business Central (hundreds of documents a day) runs from tens to a few hundred euros per month depending on model and volume. A minor line compared with the hours it replaces, but one worth budgeting and monitoring from day one.
Mistakes we see in AI + Business Central integrations
- Giving the agent a SUPER user. The agent should carry a minimal permission set: the entities it needs and nothing more. Business Central supports this precisely; use it.
- Ignoring customizations. Most long-lived BC installations carry years of extensions and custom fields. Discovering those particulars is week one of the project, not a month-two surprise.
- Posting without a shadow phase. For the first weeks the agent proposes and a person confirms; autonomy is earned with a demonstrated accuracy rate, not granted on day one.
- Forgetting API limits. Business Central online throttles requests; a well-designed agent works in batches and queues, not with a hammer of individual calls.
How we'd approach your case
The path we follow: a diagnosis session on your actual Business Central (version, online or on-prem, extensions, volumes), selection of a first process with measurable return, and a 4-8 week deployment including the shadow phase. If you want that diagnosis for your installation, it's the first conversation we have at our artificial intelligence agency in Madrid.
Frequently asked questions
Does this work with old Navision or only Business Central online?
Both, with different effort. Business Central online ships the APIs ready; NAV/Navision on-premises requires enabling web services or integrating via files, which adds project time. If a migration to BC online is on your horizon, the agent is designed to survive it: the process logic doesn't change, only the connector.
Does the agent need a Business Central user license?
Not as a named user: it connects as an application registered in Microsoft Entra with its own permissions. Check licensing details with your partner for your specific contract, but the usual pattern is that the agent doesn't consume a paid user seat.
What about the data — does it leave Microsoft?
That depends on the design. The agent can run in your own Azure tenant, use models under enterprise agreements that don't train on your data, or run models on private infrastructure when the requirement is strict. What we never recommend is sending financial data to consumer APIs without a data processing agreement.
Doesn't Copilot Studio already do all this?
Part of it: conversational cases and flows inside the Microsoft ecosystem, under its license and consumption model. Where it falls short is processes with complex logic, integrations outside the ecosystem, and fine control over costs and models. Many clients end up with both: Copilot Studio for light internal cases, a custom agent for the processes that move money.
How long until the first process is live?
Between 4 and 8 weeks for a scoped process on Business Central online: two of discovery and design, two or three of build, and the rest in shadow mode with real data. On-premises, add 2-4 weeks of integration groundwork.