AI and agents in Business Central

AI and agents in Business Central

The ERP Manager guide is written by Abakion. Read it here and download it as PDF free of charge.

Microsoft talks about almost nothing but AI these days, and it is also moving into your Business Central. That changes your role as ERP manager. You need to understand the difference between Copilot and agents, know which agents exist today, be able to build or commission agents, and you need to have a grip on the responsibility for what the agents do.

Two ways you meet AI in Business Central

You meet AI in Business Central in two different ways: Copilot and agents.

Copilot: built-in AI help at no extra cost

Copilot is already part of your Business Central and doesn’t cost extra to use. You can ask it to summarize information, generate text suggestions, or analyze data directly from lists and records.

Copilot is where most users encounter AI in Business Central for the first time. You can ask Copilot instead of searching through menus and lists. Let Copilot analyze customer ledger entries and build the sales analysis you have asked a colleague to make in Excel. It doesn’t require a technical background. It only requires that you know what you want an answer to.

Copilot is a strong helper in everyday work, but it can’t change data or perform actions. That is the agents’ job.

Agents: AI that performs actions

Microsoft builds AI agents directly into Business Central. An agent isn’t just a smarter version of a job queue or a Power Automate flow. An agent can:

• interpret a goal formulated in plain language
• reason over your data
• make decisions within the boundaries you set
• ask you for help when it is in doubt

Autonomy is clever, but you need to have control over it. Microsoft uses the phrase “human-led, agent-operated” about the direction they are moving in. Humans set the direction and make the important decisions. The agents perform the concrete work.

How the roles change

Users’ roles are changing. They will type less in Business Central and instead chat with Copilot and delegate tasks to agents.

Your role as ERP manager changes in step with that development. You become a kind of team lead for all the agents. Not because you are necessarily the manager of accounting or warehouse staff, but because you have responsibility for what happens in your ERP system, and now there are new “employees” in that system that you need to keep an eye on.

Which agents exist in Business Central today

Microsoft has started with the most obvious agents for Business Central. They improve them continuously.

Sales Order Agent

The Sales Order Agent monitors a shared mailbox for incoming emails from customers. When a customer sends an inquiry, the agent:

1. creates a sales quote with the requested items
2. checks inventory
3. sends the quote to the customer (when you have approved it)
4. converts the quote to a sales order when the customer confirms

Payables Agent

The Payables Agent works the same way for incoming vendor invoices. It:

1. monitors a mailbox
2. reads PDF invoices
3. finds the vendor in Business Central
4. suggests posting based on history
5. matches to purchase orders
6. creates a draft purchase invoice

The agent only creates drafts and only posts when you approve.

Two paths to your own agents

Beyond the two ready-made agents there are two paths to building your own agents:

The difference between automation, chatbot, and agent

Before you dive into agents, consider whether a simpler solution can do the job. The difference between the three types is:

When an agent is the right choice

AI agents perform best where you don’t have a direct “if-then” scenario. If the process is predictable with a fixed sequence, a classic automation is easier and more reliable.

If the process requires someone to assess data, handle exceptions, and adapt to variations (for example reading a vendor invoice in varying formats), then an agent makes sense.

Ask yourself whether the task requires AI

Consider whether the task requires AI at all. First investigate whether there is an app on Microsoft Marketplace that solves the task (there often is), or whether a plain automation is enough.

The responsibility is still yours

We need to talk about responsibility. It is a topic that often disappears in the excitement over the agents’ efficiency, but it returns like a hangover when a user says “it wasn’t me who posted it wrong. It was my agent.”

Nobody can offload responsibility. Your postings in Business Central are still yours, even if you have been advised by AI. That applies whether you have consulted a chatbot or let an agent post or change master data.

If you set up a posting group incorrectly, you bear the responsibility. The fact that you chose to use an agent doesn’t change the distribution of responsibility.

There are two different aspects of responsibility you must address:

• responsibility for the advice you and your colleagues get from AI tools outside Business Central
• responsibility for what the agents do inside Business Central

Responsibility for AI advice from external tools

Users who run into a challenge in Business Central will often ask an AI tool instead of calling a consultant. AI tools can give useful answers, but they can also misunderstand or be wrong.

Lack of context is the biggest problem

The most problematic part is if the AI tool doesn’t have enough context. When AI doesn’t know your full setup, it will try to answer the concrete question on incomplete assumptions.

It takes professional expertise to assess whether an AI answer is correct. If an inexperienced employee asks an AI tool for help changing a configuration, they have no basis for judging whether the answer is right. Common sense requires professional grounding.

Set a policy for AI use in the organization

As ERP manager you should have a position on how your organization uses AI tools in connection with ERP:

• when is it OK to ask an AI tool
• when does it require a professional
• how thoroughly should testing be done
• who is responsible for testing

Responsibility for agents’ actions: human-in-the-loop

Microsoft’s agents are designed so you can choose how much control you want. The principle is called “human-in-the-loop”. You are in the process, and you can approve, reject, or adjust what the agent does.

The principle isn’t new. You know it from other areas in Business Central. If you set up a job queue that posts automatically, you typically run the process manually a number of times first to make sure it does the right thing. Only when you trust that it works do you let it run automatically. The same thinking applies to agents.

Take ongoing positions on which control points you can remove, because you have sufficient trust in the agent. Be ready to introduce new control points if you start seeing errors.

Examples of control points in Microsoft’s agents

The Sales Order Agent sends notifications to your Role Center every time it needs your attention, for example when it wants to send an email to a customer, when it lacks data, or when it needs an approval.

The Payables Agent only creates drafts and asks you to review before anything is posted.

Permissions for agents

You already know permission management in Business Central. You know not to give every employee super permissions, and that audit will flag it if you do. The same logic applies to agents.

How to assign permissions in Agent Playground

When you set up an agent in Agent Playground, you give it a profile and a set of permissions, exactly as you give a user. The profile determines which pages and menu items the agent can see. The permissions determine what it may do: read, insert, modify, delete.

How to assign permissions in Copilot Studio

With Copilot Studio it works differently. You connect the agent to Business Central through an MCP connection, and the agent doesn’t have access to the user interface or menu items. Access is controlled by whether the MCP connection is set up with read, edit, or delete rights.

Think of the agent as a new employee

When you hire a new employee, you give them access to certain parts of the system, explain the work tasks, and keep an eye on whether things are done correctly at the start. You don’t give a new warehouse employee super user permissions. You give them a scanner and a bounded process.

That is also how you should approach agents. You give them a role, a set of permissions, and some instructions. And then you keep an eye on whether they do what you have asked them to do.

Agents have a different personality than people

Agents behave in many ways like ordinary users, but they have a different personality.

The difference from a human employee is that an agent always follows the rules you have given it. It doesn’t take a quick shortcut because it is getting late, and it doesn’t forget a procedure because it is busy.

Agents don’t have common sense on their own

The agent doesn’t have the common sense an experienced employee has. If you have asked your agent to always suggest the item that has historically sold best, it will do exactly that. It will always suggest the red T-shirt to all customers while the blue and green ones gather dust in the warehouse.

A salesperson would, after three or four orders, start to wonder whether it was right. The agent just plows on.

An agent’s “common sense” is predictable because it has exactly the sense you ask it to have. By contrast, it is more unpredictable in humans, where experience, situational awareness, and personal relationships come into play.

Test agents differently than people

You must test agents differently than you test people. An employee can become tired, distracted, or take a shortcut, which is why you test whether they can perform a process correctly every time. An agent doesn’t have that problem. If it does something correctly once, it also does it correctly the next time.

Agents make other kinds of mistakes:

• they can misunderstand an instruction
• they can handle an exception incorrectly
• they can draw a wrong conclusion from your data

That kind of error you don’t catch by testing the same scenario twice. You catch them by testing with variations: other data, other combinations, other exceptions. You must test much more thoroughly at the start than you think is necessary.

What does it cost to use agents

The ready-made agents in Business Central and your own agents built in Copilot Studio consume credits. We dare not put concrete prices in a guide like this because the pricing model is likely to change over time.

As ERP manager you must have a grip on how much it costs to use agents, and whether it is worthwhile. It is by no means free. Keep an eye on consumption, especially at the start, so you can assess whether the costs are in a reasonable relationship to the time savings.

Agent Playground: build agents with natural language

With Agent Playground you can design agents yourself directly in Business Central. If you know the Sales Order Agent or the Payables Agent, you know the interface they live in: a small icon in the upper right, a task pane on the side of the screen, notifications on the Role Center. Agent Playground gives you the ability to build your own agent in the same style.

You write instructions in plain language: what the agent should do, what it must not do, and when it should ask for help. If you can explain a process in words, you can in principle have the agent perform it. There is no code involved.

Playground does not mean production

It is called Playground for a reason. It only exists in sandbox environments, and it isn’t ready for production. Microsoft has launched it to let users and partners experiment and collect experiences.

You don’t have an operational responsibility for Agent Playground because it only runs in the sandbox. You can do yourself a favor by starting to play with it now. When the technology matures and is ready for production, you are well prepared to quickly put process optimization in motion.

It is Microsoft’s vision that agents will become the backbone of building business processes going forward. Whatever you think of that vision, it is wise to understand what is on the way.

Agents outside Business Central via Copilot Studio and MCP

Agents don’t have to live inside Business Central. With Copilot Studio you can build agents that live elsewhere: in Teams, on a website, in Outlook. They can still read and write in Business Central.

MCP server: the connection between Copilot Studio and Business Central

The connection is made via an MCP server (Model Context Protocol). MCP is an open standard that allows AI agents to understand data structure and perform actions in other systems.

In practice this means that in Business Central you configure which APIs the agent may use, and in Copilot Studio you build the agent and its logic.

How to configure MCP in Business Central

The setup in Business Central is fairly simple:

1. Find the “MCP Server Configuration” page
2. Create a configuration
3. Add the API pages the agent should have access to
4. Choose for each API whether the agent may read, insert, modify, or delete data

In Copilot Studio you add the MCP connection. From there the agent has access to the data and functions you have exposed. You build the agent with instructions, triggers, and any Power Automate flows.

Your role: ask the right questions

As ERP manager you don’t have to build agents yourself. You should know what is possible, and be able to ask the right questions to your partner and to colleagues working on AI initiatives elsewhere in the company:

• can this process be automated with an agent
• what would it take
• what are the risks
• how do we maintain control

Think bolder but keep control

With AI agents it is easier and faster to get started with automations than ever before. You don’t need a developer to write code. You can formulate a process in natural language and have an agent perform it.

Easier to build does not mean easier to control. An agent that can click menu items and update data can also make errors at scale if it isn’t set up correctly. Errors scale at lightning speed because the agent doesn’t pause and think “this looks wrong”.

If you build many agents, you risk having different agents work in opposite directions, with no one having the full overview. One agent lowers a value, another raises it, and one day a customer calls and asks why they are receiving 32 emails a day from the agents.

Practical action plan for AI agents

Use the following action plan as a starting point:

1. Understand what exists
2. Test in the sandbox
3. Involve your partner
4. Lay out a strategy
5. Optimize continuously

1. Understand what exists

Know the ready-made agents from Microsoft. Know the possibilities in Agent Playground and Copilot Studio. You don’t have to master the technology, but you need to know enough to ask good questions. Keep an eye on whether Microsoft releases new standard agents that are relevant to your company. The technology moves fast, and a standard agent six months from now may cover a need you handle manually today.

2. Test in the sandbox

Copy your production environment to a sandbox and experiment there. It is the same recommendation as for apps and updates, and it applies even more strongly to agents.

3. Involve your partner

Agents in Copilot Studio require knowledge of both Power Platform and Business Central. It is a collaboration that typically requires skills from both worlds.

4. Lay out a strategy

When are AI agents the answer? When is an app better? When is a Power Automate flow enough? When is the right solution to call a consultant? You shouldn’t use AI for everything, but you should have a position on when it makes sense.

5. Optimize continuously

An agent isn’t something you implement and forget. You must revisit your agents, assess whether they are still doing the right thing, and adjust instructions and permissions when your processes change.

Keep an eye on whether new functionality or new agents from Microsoft give reason to change your own agents. Test, evaluate, adjust, repeat. A good process with agents is a process you continuously improve, exactly as you do with all other processes in your Business Central.