Agentic AI is now part of Business Central. Microsoft ships agents that draft sales orders, process supplier invoices and handle expenses, and since April 2026 any compliant AI agent can connect to BC through Microsoft's MCP server. What agentic AI is in plain terms, what BC can do with it today, what it means for developers, how to get your system ready, and where it is heading.
Agentic AI is now part of Business Central. Microsoft ships agents that draft sales orders, process supplier invoices and handle expenses, and since April 2026 any compliant AI agent can connect to BC through Microsoft's MCP server.
For the last couple of years, "AI in ERP" mostly meant a chat box: you asked Copilot something, and it answered or drafted some text. That is useful, but it still leaves a person doing the work. Agents change that. They take a task, carry it out across several steps, and come back to you only when they need a decision.
This article explains what agentic AI is in plain terms, what Business Central can actually do with it today, what it means for developers, how to get your system ready, and where I think it's heading. I'm writing as an independent BC consultant who spends a lot of time on implementations, NAV upgrades and project rescues, so the focus is practical.
Agentic AI is software that pursues a goal by deciding and taking its own next steps, rather than producing one answer to one prompt.
A chatbot works in a single exchange: question in, answer out. An agent works in a loop. It reads the goal, plans a step, acts, checks what came back, and decides what to do next. It keeps doing this until the job is done or it reaches a point where a person needs to decide.
The loop repeats without anyone stepping in. The person comes in at the start, with the goal, and at the end, to approve the work.
Think of a capable new starter in your accounts team. You don't tell them which buttons to press. You say "process this week's supplier invoices", and they open the inbox, match each invoice to a purchase order, ask you about the odd one that doesn't fit, and leave the rest ready for you to approve.
Agentic systems are built from four parts:
The last part matters most in an ERP. An agent that can post to the general ledger needs the same level of control you would give a person with that access, or more.
Business Central has three Microsoft-built agents in use today, plus a common set of controls for supervising them. Version 29 (the 2026 release wave 2, rolled out in October) mostly improved these agents and their controls rather than adding new ones.
| Agent | What it handles | Where it stands in BC29 |
|---|---|---|
| Sales Order Agent | Watches a sales mailbox, reads customer requests for quotes and orders, checks item availability, and drafts the quote or order with a reply for a person to review | Available; started from the Role Center |
| Payables Agent | Reads supplier invoices from an AP inbox, handles PO and non-PO invoices, and learns from the mapping corrections users make | Generally available; BC29 improves purchase order matching using line amounts and expected receipt dates |
| Expense Agent | Processes employee expense claims, including mileage, VAT, withholding tax, duplicate checks and approvals | Public preview; BC29 adds AI-driven approvals and more countries and languages |
The controls around them matter as much as the agents. BC29 made these generally available:
The pattern is human-in-the-loop. Agents prepare the work and people approve it. That is the right starting point for financial data, and it is how I would introduce agents to any client.
On cost, Copilot's assistive features come with Business Central. The task agents are charged on consumption (Copilot Credits) rather than included in the licence. Check Microsoft's current licensing before you estimate a business case, because the pricing details are still changing. (We worked through some numbers in the real cost of AI agents.)
The Business Central MCP server lets AI agents built outside BC read and act on your data through a standard, secured connection. It has been generally available since April 2026.
MCP (Model Context Protocol) is an open standard that defines how AI agents discover and use tools. Before it, every AI integration needed its own custom connector. With MCP, an agent connects to BC's endpoint, sees which actions and data it is allowed to use, and works with them, much as a person uses the menus their permission set allows.
What Microsoft has shipped:
mcp.businesscentral.dynamics.com) using the latest MCP specification and OAuth 2.1.This is the change I find most significant. Microsoft's own agents cover common scenarios. MCP lets you build agents for your own: an agent in Copilot Studio that chases overdue customers, a tool that answers "what's our stock exposure on this range?" from Teams, or an assistant such as Claude working through a month-end checklist against live data. Because the security model is BC's own permissions, it doesn't open a new hole in your system. (See permissions in Business Central.)
AI agents are now a supported part of the AL development toolchain. BC29 includes a set of generally available features that let coding agents work with your extensions directly.
In practice, the routine parts of AL work get quicker: scaffolding a table extension, writing an API page, tracing an event subscriber chain or reading an unfamiliar ISV codebase. The judgement doesn't change. Someone still has to decide whether a customisation should exist at all, where it belongs, and how it will survive the next upgrade.
For NAV upgrade work, this helps with the slow part: understanding years of C/AL modifications and deciding what becomes an extension, what returns to standard and what is dropped. An agent can map that estate quickly. An experienced consultant still has to decide what to do with it. (See upgrading from NAV: the code side.)
Agents perform only as well as the system they work in. Most of the preparation is ordinary good ERP practice that now pays back twice.
This is where NAV customers need to plan carefully. Moving a NAV database to BC unchanged, with its old customisations and messy data, gives you a cloud system that agents can't use well. An upgrade is the best opportunity you'll get to bring processes back towards standard and clean up the data. Treat it as preparation for agents as well as a technical migration, and you'll get the benefit for years.
Rescue projects are similar. When I'm brought in to recover a struggling implementation, the problems are nearly always data, setup and process. Those are the same things that decide whether agents will help or just automate the mess.
AI in Business Central will no longer arrive in two scheduled waves a year. BC29 is the last release announced under the release-wave model. From September 2026, Microsoft publishes new capabilities continuously on its AI at Work roadmap, and the old release planner closes on 15 November 2026.
That shows how quickly Microsoft expects this area to move. Based on what has shipped and the direction of the platform, these are the developments I expect over the next 12 to 24 months. They are my views, not Microsoft commitments.
What won't change is the need for people who understand both the technology and how the business actually runs. Agents remove the repetitive work. Someone still has to design the processes they follow.
Pick one bounded, high-volume process, prepare the data and permissions behind it, and run an agent with a person approving every action until the results earn more trust.
Purchase invoices are usually the best first candidate: high volume, clear rules and a measurable error rate. Sales order intake is a close second if your customers order by email. (See what the Payables Agent actually does.)
If you're on NAV or an older BC version and wondering how agents could fit your business, or you have a BC system that isn't delivering what it should, I'm happy to talk it through. At BBXD I work on Business Central implementations, NAV-to-BC upgrades, rescues, data migration, AL development and integrations, and getting systems ready for what comes next is part of that work. Book a 30-minute call.