Article

The AI already in QuickBooks or Xero, or an AI agent alongside: what is the difference?

For growing businesses: the AI in your software stays inside that software; an AI agent alongside does the work in between.

BombosBombos team··11 min read

By Piet Baudoin · September 2026

The AI already inside QuickBooks, Xero or any other accounting software stays inside that one piece of software, and Microsoft includes the one in Business Central at no extra cost for now. An AI agent alongside costs extra, handles the work that runs across two or three pieces of software in a small or midsize business, and is made to learn how things work at your company.

At Bombos, we build that AI agent alongside.

That is why this article starts with what the AI in your own software does well. Then where it stops. If you want the big picture first, read what AI agents are taking over at small businesses today.

What does the AI already inside your accounting software do?

It does the work inside its own software, and it does that well.

Copilot is the AI in Business Central, Microsoft’s accounting and business software. Microsoft writes that it “is included with your Business Central license at no extra cost”, and adds in the same breath: “However, fair-use policies, quotas, or pricing might be introduced later.” Business Central’s Payables Agent watches your company mailbox for vendor invoices, matches them to purchase orders and receipts, and “proposes posting accounts for unmatched lines”.

Intuit builds its own AI agents into QuickBooks Online. According to its help pages, they “act on your behalf with your permission.” About its payments agent, Intuit writes: “Get paid 4 days faster on average when you send invoice reminders with Payments AI.” That number comes from Intuit, not from an independent measurement. It still says something, because the maker knows its own software better than anyone.

Zoho, the maker of Zoho Books, built its own AI model for all its applications. It “runs securely on Nvidia GPUs within Zoho’s private data centers”, and the same page promises “No data harvesting” and “No shadow training”.

If most of your work lives in that one piece of software, this is the right answer. Supplier invoices that someone retypes are that kind of task.

Where does the AI in your software stop?

At the edge of that software. That is not a flaw. It is a deliberate boundary.

A breakdown report comes in on Friday afternoon as a text message. The address is in one system. The schedule is in Outlook. The technician’s work slip lands in the other system on Monday. And the arrangement you made with that customer last year lives in the head of your job planner.

The AI inside AFAS, a Dutch business software package, is called Jonas, and it does what you set up in AFAS. Irixs, a Dutch firm that sells AFAS work itself, puts it plainly: “Jonas is not a stand-alone feature you can freely experiment with. It is a component you have to set up in advance in your AFAS environment” (translated from Dutch).

Zoho’s AI stays inside Zoho. That is exactly the promise Zoho makes you about your data. You cannot have both.

The work that takes the most time sits in the workflow between your systems. Retyping, looking up where something was again, checking with a colleague who is in a meeting. The AI inside any one piece of software never reaches that far. Scheduling a service call and processing timesheets both start outside the software they end up in.

You can, and it is getting easier. So the question is not whether it can be done, but who manages it.

In September 2026, Google switched on the connections between Workspace and HubSpot, QuickBooks and Salesforce by default for everyone with Gemini. An administrator can switch that off again per group. They just have to know about it. Zapier warns in its own guide to its newest automations: for now, keep your most sensitive work out of them.

Aembit, which itself sells access control for AI assistants, wrote on 15 September 2026 about the new OWASP list of the ten biggest risks of AI applications. This time the list was checked against 6,639 incidents that were described well enough to classify. One risk climbed from sixth place to third: an AI that is allowed to touch more than its task requires.

The software makers draw different conclusions from this. AFAS wrote in June 2026 that outside AI is allowed only if it is certified. At that point not a single one was certified, and all other use was at your own risk. FreshBooks did the opposite. In a support article from September 2026, it tells its customers: “Use the FreshBooks connector in Claude to ask about your FreshBooks account’s profit, cash, invoices, clients, and more in plain language.” For what ChatGPT is allowed to do in your software, see can ChatGPT operate your accounting software.

What is the difference, point by point?

The AI in your software An AI agent alongside
What it costs In Business Central nothing extra for now, though Microsoft says fair-use policies, quotas or pricing might come later; not checked for the other makers A separate amount per month, on top of what you already pay
Where your data stays In the software; that is the promise With the vendor you choose, so ask where its servers are
What it touches What is in that one piece of software The software you already use, wherever you give it access
Who sets it up You or your implementation partner, in advance The vendor with your people, task by task
Work across two systems Stops at the edge Is exactly what it is for
What it learns about your business What you set up in advance How things work at your company: it remembers every correction you make, and that correction becomes the rule next time
Who approves You, in the settings you fix in advance You, per message or entry, before it goes out

The real difference is in the row about work that runs across two systems, because no setting closes that gap. The second one is a row lower. The AI in your software does what you set it up to do. An AI agent is made to get to know your business: which invoices are always right, which customer calls instead of emailing, and why things work differently with that one client.

Choose the AI in your software if the tasks you want off your plate stay within its fields. For a bookkeeping firm that does everything in QuickBooks or Xero, that is the honest answer. You do have to set it up in advance.

Choose an AI agent alongside if the work starts in an inbox or a text message and ends in two systems. And if the reason things work differently with that one customer lives in someone’s head.

What happens if this pace continues?

Every piece of software gets its own AI, and each maker closes its doors in its own way.

FreshBooks put its own connector inside Claude in September 2026. AFAS wrote in June 2026 that outside AI is allowed only if it is certified, and at that point it had none. Zoho runs its own AI model within its own data centers and promises “No shadow training”. TOPdesk, a Dutch service desk maker, announced its AI teammate Robin on 17 September 2026 and keeps customer data within the customer’s region by default. Four makers, four answers, all four on the record by September 2026.

My hypothesis: by the end of 2027, every piece of software you use will have its own AI, and the scarce thing will not be that AI. It will be the knowledge of how your business moves work from one system to the next.

The reasoning goes like this. Every maker promises its customers something about their data, and so keeps its AI within its own walls. Those promises get in each other’s way, so getting the AIs of two pieces of software to talk to each other becomes harder, not easier. What remains is the workflow between your systems: which customer has a special arrangement, and why that one supplier is always booked on two lines. Right now, that knowledge lives in people’s heads and in a few files on someone’s desktop.

What would break my hypothesis: certified AI tools at AFAS, or software makers letting their AI talk to each other through one agreed door, as FreshBooks already does with Claude. Then keeping watch over the workflow between systems becomes the job of the software itself.

For a small or midsize business, that means: start now by writing down how work moves from one system to the next. And whichever side you choose, ask your vendor one question: where do you record that, and can I read and change it myself?

What can’t an AI agent alongside do today?

It knows nothing that nobody has told it. If the reason for that one arrangement lives only in your job planner’s head, it asks you in a conversation and records the answer. It asks; it does not make things up.

What it does bring is an extra vendor, a separate amount per month, and your data in a second place. Those are three questions you want answered before you start.

And it never moves faster than you. A message to a customer or a payment waits until someone clicks Approve, and that takes a few seconds every time. For why it should work this way, see should an AI assistant send email on its own. For who is on the hook if something still goes wrong, see who is liable when an AI agent makes a mistake.

How do you start small?

Switch on the AI in your own software for one task that stays inside that software. In Business Central that costs nothing extra for now.

Then, for one week, keep a tally of the work where someone clicks back and forth between two screens. That list is your real question. The overview of tasks that lend themselves to this helps you choose.

Bombos is that AI agent alongside. It reads what comes in (email, text messages, work slips and invoices), finds the matching data in the software you already use, and prepares the work with the reason attached. No message goes to a customer and no payment goes out until someone from your company clicks Approve. That is technically enforced and always on.

If every piece of software soon gets its own AI, the work between them is what remains, and that is where Bombos comes in. What Bombos learns about your business does not sit in the model. It sits in what your people approved, corrected and rejected: your correction becomes the rule next time. That memory stays when a better model is put underneath, and a better model does more with it. We can put the strongest and the most cost-efficient model underneath right away, and switch per client to whatever that client wants. That way, the knowledge of your workflows no longer lives only in people’s heads.

We run on our own servers in Europe, and we start with one task. After that, your own team teaches Bombos the next one. After three months, that work is ready every day, without anyone having to think about it. Leave your number on the contact page and we will call you back.

Start with the workflow where you retype the most.

Sources

Every source was opened on 20 or 23 September 2026 and every quote appears in it word for word.

AFAS

TOPdesk

Microsoft, Intuit, Zoho and FreshBooks

Connecting, and the risks that come with it