Learn · Compare

The AI in your ERP, or an AI agent of your own?

Published 31 October 2026

Short answer. Use the AI that comes with your ERP or accounting system for jobs that start and end inside that system, such as matching bank lines or turning an emailed invoice into a draft. Build an AI agent of your own for jobs that cross systems, read documents the ERP never sees, or must follow your company’s own rules and words. Often the answer is both: the ERP’s features inside the ERP, and your own agent for the work around it, delivering into the ERP.

This comparison compares the two, with Microsoft Dynamics 365 Business Central as the example, and ends with a verdict for each kind of job.

Two roads, defined

An AI strategy gives each job one of three roads: the AI licences you already pay for, the AI features in your ERP and other systems, and agents made for your own jobs on your own data. This comparison compares the last two.

The AI in your ERP	An AI agent of your own
The AI features and agents the vendor builds into the ERP or accounting system. You switch them on, set them up and review their work. The vendor maintains them.	An AI agent built for one of your jobs, on your own data, following your own rules. You own it, its tools and its data stay in your own cloud, and someone you choose maintains it. Gathena calls one an agentic knowledge worker.

On both roads, a person checks the agent’s work. On this site that person is the operator: the person in the company who instructs the agent, reviews its work and signs it off.

The first road, AI assistants such as Copilot, Gemini or Claude, is covered in What belongs in an executive AI strategy.

What the AI in Business Central does today

Business Central has two kinds of AI built in. Copilot features help a user with a task the user starts. Agents work on their own in the background and ask a person when they are unsure. Microsoft’s own documentation, checked on 31 October 2026, lists these for finance and order handling:

Feature	What it does	Where it stops
Bank account reconciliation assist	Proposes matches between bank statement lines and ledger entries, and suggests posting accounts for lines it cannot match.	It works on the bank statement and the ledger in Business Central.
E-document matching	Maps the lines on vendor e-invoices to purchase order lines and flags exceptions.	Microsoft lists English as its only supported language.
Payables Agent	Watches a Microsoft 365 mailbox for vendor invoices in PDF, reads them, finds the vendor, matches purchase orders or suggests accounts, and prepares a draft for a person to review. It can then turn the draft into an unposted purchase invoice.	It never posts the invoice. It reads only PDF attachments in the mailbox it watches.
Sales Order Agent	Reads customer requests by email, checks stock, asks follow-up questions, and prepares a quote or an order.	It works from one mailbox into sales documents in Business Central.
Expense Agent	Reads photos of receipts, checks each expense against the company’s rules and policy, and helps employees submit expense reports. Microsoft describes it as a production-ready preview.	It handles employee expenses.
Chat, analysis assist, summaries	Answer questions about the data the user may see, arrange lists into analyses, and summarise a record in a few points.	They use the data in Business Central.
MCP server	Lets AI tools outside Business Central read its data and, where an administrator allows it, create, change or post records through chosen interfaces.	Every operation runs with the signed-in user’s identity and permissions. The default set-up is read-only.

Before you rely on any of them, check five conditions:

Business Central online only. Microsoft offers Copilot only in Business Central online, not in on-premises or private-cloud installations.
Billing. Microsoft includes the Copilot features in the Business Central licence, and says fair-use policies, quotas or pricing may come later. Agents such as the Payables Agent use Copilot Credits, which Microsoft charges per use.
Control. An administrator, through your IT, in-house or your IT partner, can switch off each feature and decide which users may use it. Each agent works under its own user in Business Central, with the permissions the administrator gives it, so its actions show as the agent’s.
Language. Microsoft supports each feature in its own list of languages, and the lists change. Check that the feature supports Danish before you use it in daily work.
Custom agents. Microsoft also offers a toolkit for designing agents of your own inside Business Central. Microsoft’s documentation describes it as a preview.

Microsoft calls the person who reviews an agent’s work the agent supervisor. That is the role this comparison calls the operator.

Other ERP and accounting systems

Other systems are adding AI of their own. e-conomic, for example, reads scanned invoices and receipts and suggests how to post them in its Smart Inbox, and its Autoforslag function suggests postings for bank lines from how the company has posted before. Uniconta has added an AI Assistant that answers questions about the data from the screen the user is working on.

Each vendor decides what its features do, which plan includes them and where they stop. Read your own system’s documentation with the five questions below.

Five questions that decide

Ask these for one job at a time. A company dictionary, used below, is a shared record of what your company calls things, what each thing means, and how the things connect.

Question	The AI in your ERP	An AI agent of your own
Does the job stay inside one system?	It fits best when the job starts and ends in the ERP.	It can work across the ERP, the bank, the webshop, supplier portals, mailboxes and files.
Does it need documents or data from outside the ERP?	It reads the data in the ERP and what arrives through the channel the vendor built, such as one mailbox of PDF invoices.	It reads what the job needs, through connections you approve, and keeps each document with the entry it supports.
How much do your own rules and words matter?	It follows the vendor’s process and the settings the system offers. Some features learn from your history in that system.	It applies your written rules and your company dictionary, and the logbook shows which rule it used.
How much control do you need over access, rules and the logbook?	You control access with the ERP’s own permissions. The ERP records what its agent did. The vendor decides how the feature works and tests it.	You give the agent its own user in each system and can withdraw it. Your logbook covers every system it touches. You version its rules and test it on your own cases before and after every change.
What happens when you change systems?	The feature and its set-up stay with the old system.	The rules, the company dictionary and the tests move with you. The connection to the new system must be rebuilt.

Two more differences matter for the budget and the work:

Who maintains it. The vendor maintains the ERP’s features and updates them with the system. An agent of your own needs someone to maintain it: you, or whoever you choose.
What it costs to run. The ERP’s features come with the licence or are charged per use by the vendor. An agent of your own is built first, and then costs what its parts cost to run: a language model priced by the token, a document service by the page, and the cloud it runs in.

The verdict

The AI in your ERP is enough	An agent of your own is worth it	Use both
The job starts and ends in the ERP, the vendor built the feature for that job, your process matches the vendor’s, and the feature supports your country and language. Switch it on, give it the right permissions, and let the operator review its work. You do not need Gathena, or anyone else, for this.	The job crosses systems, or needs documents the ERP never sees, or depends on your own rules and words, or you need one logbook and your own tests across every system. Also when you want the work to stay yours if you change ERP.	The ERP’s features handle what happens inside the ERP. Your own agent handles the work around it and delivers into the ERP, where the ERP’s own approval and posting take over.

Start with what the ERP already does. Build only for the part it does not cover.

How the two work together

Take supplier invoices. The ERP’s agent handles invoices that arrive as PDFs in the mailbox it watches. Some invoices never arrive by email: they wait in a supplier’s portal, and someone has to sign in and fetch them.

Your own agent does the part outside the ERP. It signs in to the portal under its own identity, fetches the invoice, and reads it with your rules.
It delivers into the ERP. It creates a draft entry with the invoice attached, through the ERP’s own interface. For Business Central, that can be the MCP server Microsoft provides, with the permissions an administrator chose.
The ERP takes over. The draft waits for the same approval and posting as every other invoice. The operator approves it in the ERP.

The books stay in the registered bookkeeping system. An agent that reads from it and delivers drafts or entries into it is not a bookkeeping system itself; one that keeps the books or the only copy of a document elsewhere would be. Who is accountable when an AI agent makes an entry? explains the rule. Fetching invoices from vendor portals walks through the portal step.

The same rules apply on both roads

Whichever road a job takes, the company stays accountable for the result. Every step of an agent’s work runs in one of three sign-off zones: “acts alone”, “asks first”, or “the operator signs off”. The limits that matter most are enforced outside the model, in access rights and approval steps. An agent inside the ERP needs these as much as an agent of your own.

Checked against the vendors’ documentation on 31 October 2026.

Sources

Product and company names are trademarks of their respective owners, used to identify the products referred to. No endorsement or partnership is implied.