.png?width=1920&height=253&name=Together-Vision-Technology%20(1).png)
.png?width=300&name=Pattern%20Header%20(1).png)
Most businesses have at least a handful of administrative processes that seem obvious candidates for AI automation in Business Central, yet they never quite fit into traditional workflow tools.
It would be customer returns that arrive in different formats. Service requests that are missing information. Someone has to read emails, check records, interpret what the customer meant and decide what happens next. These tasks are not particularly difficult, but they take up valuable time because they rely on judgement rather than a straightforward set of rules.
Microsoft's standard agents, such as the Sales Order Agent, Payables Agent and Expense Agent, have demonstrated what is possible. However, many organisations quickly reach the same conclusion: the idea fits, but the process does not quite match the way they work.
Business Central's new custom agent capability is designed to address exactly that problem. It allows businesses to build agents that follow their own processes rather than Microsoft's predefined workflow. Before deciding whether custom agents are something your business should explore, it is worth understanding both what they can do and where their limitations lie.
Why the standard agents only take you so far
The AI capability in Business Central has arrived in stages. Copilot came first, handling narrow tasks such as summarising a record, suggesting values for fields or helping with bank reconciliation. These improvements were useful, but limited in scope by design.
The Sales Order Agent and Payables Agent were a step beyond that. They work autonomously in the background, chaining a series of actions together to complete a whole task rather than assisting with one. The Expense Agent went further still, adding a dedicated web portal alongside the email channel and handling a richer process with more steps.
In practice, this is often where businesses hit the limits of the standard agents. All three share the same constraint: the workflow is fixed. The instructions that drive each agent are written by Microsoft and, while those instructions have now been published so you can read them, you cannot change them. The triggers are limited too, largely to an email channel. If your process differs from the one Microsoft designed for, there has been no way to close the gap.
What a Business Central custom agent actually is
A Business Central custom agent works in the same way as the Sales Order Agent, running on the same engine but following instructions you have written. You build it inside the Business Central client and you do not write any code. It is based on configuration and natural language.
When we have been testing custom agents, we have found there are really three parts that determine whether an agent succeeds or fails. Firstly, there are the instructions, and they are the bulk of the work: a block of text setting out the agent's responsibilities, a set of guidelines covering what it should and should not do, and the steps you want it to follow. Secondly, there is the profile. This controls what the agent can see, in the same way a profile governs what a user sees and should be stripped back to the minimum the agent needs. Finally, there are the permissions. They control what it can do and, as with the profile, you should give your custom agent the least it needs to do the job.
The instructions are automatically saved as a new version every time you change them. This matters because it means you can experiment freely and roll back if a change makes things worse.
How do you know if a process is suitable for a Business Central custom agent?
One of the questions we are asked most often is whether a particular process is a good candidate for an AI agent.
In practice, the best opportunities sit somewhere between traditional automation and human decision-making. If every possible outcome can be defined in a clear set of rules, conventional automation is often the simpler and cheaper option. Equally, if a process depends heavily on context, experience or complex judgement, an agent may require too much supervision to be worthwhile.
The sweet spot is a process that follows a broadly repeatable pattern but still requires interpretation. Reading customer emails, identifying missing information, categorising requests, proposing actions and escalating exceptions are all good examples.
A useful test is to ask whether a member of your team could complete the task using only the Business Central interface and information already available within the system. If the answer is yes, there is a reasonable chance that a custom agent could help as well.
A worked example: handling customer returns
Imagine a customer email saying there is a problem with a recent delivery. They are not using your return reason codes, they are not following a neat template, and the information is not all in the body of the message. They simply explain the issue in their own words and attach the original invoice.
In this example, the customer says that a delivery received on Tuesday included two cases of Barolo with missing labels. The message is imperfect: the wine name is misspelled, the date is vague, and the invoice details are buried in an attached PDF. By using the custom agent builder functionality in Business Central, you could build a returns agent that can read the email, open the attachment, pull out the invoice number, match it to the posted sales invoice, identify the customer and find the relevant lines.
From there, it must make a judgement. The existing return reason codes do not clearly cover missing labels, so the agent proposes a new one called “no label” and asks for approval before using it. Once approved, it creates the return order with the correct customer, the right item, a quantity of two and the new reason code.
The wider benefit is not simply processing a return more quickly. It is reducing the amount of routine administration that prevents experienced people from focusing on customer service, exception management and more valuable work.
This is a pattern we see repeatedly when discussing AI agents with customers. The greatest value rarely comes from replacing an entire role. It comes from removing dozens of small tasks that quietly consume time every day, allowing people to focus on the work that genuinely requires experience and judgement.
That is the part that matters. The value is not simply that the agent saves someone from keying in a return order. It is that it can interpret an untidy, real-world request, connect it to the right records in Business Central and pause for human approval when the decision is not completely clear. Deciding which return reason applies when a customer describes a problem in their own words is not a deterministic task, and it is exactly the kind of work a rules-based system handles badly.
You stay in control throughout with custom agents
Whenever we speak to customers about agents, questions about control and governance are never far behind. Fortunately, the answer is much the same as it is for Microsoft's standard agents. An agent is treated as a user. It gets its own user ID, and every record it creates or updates is stamped with it, so there is a full audit trail of what it did.
Its only tool is the interface. It does not call APIs or run code directly. Like a user, it is placed into a role centre and can only navigate the pages you have chosen to expose. This is why a tightly trimmed profile matters. You also decide where it needs to check in with a person. In the returns example, the agent was told to ask before creating a new return reason and again before finalising the order. You decide whether an agent will be fully autonomous or whether it needs to stop and check at every step.
The catch: building and running custom agents are two different things
Before businesses get too excited, there is one important limitation worth understanding: the agents you build in the client this way run in a sandbox environment only. You cannot build one in the morning and have it processing live transactions in the afternoon.
This is less of a restriction than you might first think because it reflects what this capability is for. It lets you take an idea, build it quickly, break it safely and find out whether it works at all before anyone commits to it. When you have an agent that behaves the way you want, its definition is exported and converted into AL code, which is then deployed as an extension to your production environment.
That conversion is where the real capability arrives. In the sandbox you must trigger an agent manually by giving it a message. As code, it can be triggered by an email, by an event such as a sales order being created, by a job queue, by an API, or by a button someone clicks.
You also gain the governance, testing and change control that organisations typically need before introducing an agent into a live environment. This becomes increasingly important as AI capabilities evolve over time, because you need confidence that a process will continue to behave as expected.
Getting an agent ready for live use is more than a simple conversion. It is where the agent is tightened up: your profiles and permissions are narrowed and, wherever possible, the agent's decisions are replaced with code. The principle is that agents are for judgement and code is for certainty. Anything deterministic is better handled as code, leaving the agent to do the less predictable reasoning that only it can do. This makes the agent faster, more reliable and cheaper to run, because every step it takes consumes credits.
The financial costs of Business Central custom agents
Custom agents use Copilot credits in the same way the other agents do. These are not included in your Business Central licence and are bought separately, either as a pay-as-you-go consumption model or as a block of credits purchased in advance. A credit is worth one US cent.
Microsoft's current guidance is that a generative answer consumes two credits and an agent action five, though per-step rates specifically for custom agents have not yet been released.
At TVT, our testing has found that agents consume around two credits per step for almost everything they do, with five-credit actions being rare. The running total is visible on the task in Business Central as the agent works, so the cost is not a mystery.
Where Business Central custom agents work best
Custom agents work best in a repetitive, rule-based workflow with a clear beginning and end, which someone could complete through the Business Central interface without needing code. Ideally, the workflow should involve a judgement call somewhere, because that is where an agent earns its place rather than simply replacing a job queue.
Ideas that come up regularly include triaging failed job queue entries and restarting the ones that are only transient, reviewing a contact list and merging duplicates, turning inbound service emails into service orders, chasing outstanding approvals, investigating why emails have not been sent, and summarising change log entries into something readable. None of these is a large project on its own. Collectively, they are the kind of administrative work that quietly consumes a great deal of time.
Introducing our AI Proof of Concept service for Business Central
The truth is that not every AI idea survives contact with reality. Some turn out to be straightforward and valuable. Others prove impractical once you look at them properly, and it is difficult to tell which is which from a discussion alone.
Many organisations have no shortage of AI ideas. The challenge is knowing which are genuinely worth pursuing.
We regularly see businesses identify processes that appear to be ideal candidates for automation, only to discover that the effort involved outweighs the benefit. Equally, we sometimes find that seemingly minor administrative processes offer surprisingly large gains once the manual effort is removed. The difficulty is that it can be hard to tell the difference from a discussion alone.
That is why we have introduced our AI Proof of Concept service for Business Central. It is a fixed-price, fixed-scope engagement at £2,500 that takes one idea and tests whether it genuinely works inside Business Central. We start by defining the problem and the success criteria with you, build a first working prototype using your own data and scenarios, then review and refine it with you and your stakeholders before producing a second draft.
What you get at the end is not a finished solution. It is a working prototype and a clear recommendation. This will cover what a full implementation would involve, the expected benefits and limitations, the technical considerations and risks, and time and cost estimates. In other words, all the information you need to decide whether to take the idea further, and something tangible to show stakeholders rather than a hypothetical.
Talk to us about your custom agent ideas
Most organisations have at least one process that everyone agrees is frustrating but nobody has found time to improve.
Custom agents provide a new way to approach those processes, particularly where traditional automation struggles because some degree of judgement is still required. The opportunity is not simply to automate more work. It is to remove low-value administration while keeping people involved where experience and decision-making matter most.
The most important lesson is not to start with the technology. Start with the process. Understanding about where people spend time, where decisions need to be made and where information needs to be interpreted will usually tell you far more about the opportunity than any discussion about AI features.
If you have a process in Business Central that feels like a candidate for an AI agent, it is worth having a conversation. Book a short discovery call to discuss whether your process is suitable for an AI Proof of Concept for Business Central. We will help you assess whether a custom agent is likely to add value and, if it is, our AI Proof of Concept service for Business Central provides a practical and low-risk way to explore the idea further.
Frequently asked questions
What is a Business Central custom agent?
A custom agent is an AI agent you build yourself inside Business Central, following instructions you write in plain language rather than a fixed workflow defined by Microsoft. It runs on the same engine as the Sales Order Agent and Payables Agent.
Do you need a developer to build a Business Central custom agent?
Not to build and test one. Custom agents are created through configuration and natural language instructions inside the Business Central client. Development work is needed only when you want to move an agent into your production environment, which involves converting it to code.
Can a custom agent run in my live production environment?
Not directly. Agents built in the client run in a sandbox environment.
How do you move a Business Central custom agent into production?
A Business Central custom agent created in the client runs in a sandbox first. To use it in a live environment, its definition is exported and converted into AL code, then deployed as an extension. This gives you the opportunity to tighten permissions, add testing and governance, and decide which parts should be handled by code rather than by the agent.
How much do Business Central custom agents cost to run?
They consume Copilot credits, which are purchased separately from your Business Central licence, either as pay-as-you-go or as a pre-purchased block. One credit is worth one cent. Consumption is visible on each task as the agent works.
Is there an audit trail of what an agent does?
Yes. Each agent has its own user ID, and every record it creates or updates is stamped with it. A task log also records each step the agent took, what it did and why.
What is included in TVT’s AI Proof of Concept service?
Our AI Proof of Concept service is a fixed-price, fixed-scope engagement that takes one practical AI idea and tests whether it can work inside Business Central. We define the process and success criteria with you, build a first working prototype using your scenarios, review it together, then produce a refined second draft and a clear recommendation on whether the idea is worth taking further.
How much does an AI Proof of Concept for Business Central cost?
The service costs £2,500. It is designed to give you a practical way to test one idea before committing to a larger project, so you can make a decision based on a working prototype rather than a theoretical discussion.
Will the proof of concept give us a finished AI agent?
No. The purpose of the proof of concept is to prove whether the idea is viable, not to deliver a production-ready solution. At the end, you receive a working prototype, an assessment of the expected benefits and limitations, and guidance on what a full implementation would involve.