Connecting ChatGPT, Copilot and Gemini to Exact Online

What you are really asking

The question reaches us almost weekly, in every possible variant. Can ChatGPT talk to Exact Online? Does Copilot work with Exact? Can Gemini update my bookkeeping? Behind all of them sits the same wish: too much time goes into retyping orders, invoices and tickets, and you hope a language model can take that over.

That is a fair wish, and the answer is more nuanced than yes or no. A chat assistant can do a surprising amount, but almost never the part where the time actually goes. It helps to separate two things clearly: retrieving and understanding information, versus writing a transaction into your administration.

What a chat assistant can do

Drop an export of your general ledger or an open items report into a chat window and the assistant will do useful things with it. It summarises, spots outliers, explains why a VAT code looks odd and drafts a tidy dunning email. For one-off thinking work, that is genuinely valuable.

Something is happening on the Microsoft side too. Copilot lives in Teams, Outlook and Excel, and helps there with summarising and writing. Drag an export in and you get the same kind of help. Fine for analysis, but it stays at the edge of your process.

• Explanation: what does this ledger account mean, why is this item still open?
• Summary: which customers are furthest behind, where is the largest outstanding amount?
• Draft: write a reminder, a supplier email or a note alongside the monthly figures.
• One-off analysis: line up two exports and name the differences.

Notice what these four have in common. They all end with text on your screen. Nobody has put anything into Exact Online yet.

Where it breaks down: reading is not posting

The gap between talking about your administration and working in it is wider than it looks. Posting an order in Exact Online is not a text task, it is a sequence of checks and actions that all have to hold before anything enters the books.

• Recognising which account this is, even when the mail comes from a different address than the customer record in Exact.
• Matching items to your item numbers while the customer uses their own descriptions.
• Applying price agreements, tiers and discounts as they exist in the system, not as the email describes them.
• Setting the VAT code, cost centre and journal correctly, because otherwise that is where next month's corrections begin.
• Knowing when it does not know, and stopping instead of guessing.

That last point matters most. A chat assistant is built to always produce an answer. In conversation that is pleasant. In an administration it is dangerous, because a plausible but wrong item number simply becomes a posting that someone trips over three months later.

Then there is the practical side. Exact Online has an API with authentication, tokens that expire and limits on the number of calls. A chat window does not talk to that on its own. And even if you build that bridge, the questions remain: who is allowed to do what, and where do you record what happened.

Copilot, Gemini, Claude or Mistral: does the model matter?

We see the question in every flavour: Copilot with Exact Online, Gemini with Exact Online, Claude, Mistral, Perplexity. The honest answer is that the choice of model matters far less than people assume.

Today's models are all comfortably good enough to read an order confirmation and pull the right fields out of it. What it fails on is not language comprehension but everything around it: the connection to Exact, your company's rules, memory of earlier cases, oversight and logging. You do not build that in a chat window.

For us the model is therefore a component you swap, not a foundation. If a better or cheaper model arrives next year, we switch it out without your process changing. What stays is the connection, the logic and the oversight. So do not pick a model, pick a way of working.

What you need to genuinely post in Exact Online

If you want orders, purchase invoices or sales invoices to actually land in Exact Online, you need more than a clever model. You need an agent that is allowed and able to work in the system, with the same kind of arrangements you would make with an employee.

• A real connection to the Exact Online API, with its own credentials, refreshing tokens and sensible handling of rate limits.
• Permissions that fit the task: what may the agent create, what may it only read, and what never.
• Business rules: your tolerance on price differences, your journals, your cost centres.
• An explicit uncertainty threshold: when in doubt, no posting, but a task for a person with the reasoning attached.
• Logging at posting level, so you can see afterwards, line by line, what happened and why.

How an AI coworker does it differently

An AI coworker does not live in a chat window but in your process. The email with the order arrives at your normal order address. The agent reads the attachment, recognises the customer, matches items to your item numbers, checks prices against the agreements in Exact Online and either stages the sales order or posts it, depending on what you agreed.

If something is off, the order does not quietly proceed. The agent sets it aside with the reason attached: this item number was not found, this price deviates from the agreement, this address does not appear under this customer. Your colleague decides. That is the human in the loop, and it is a design choice, not an emergency brake.

The result is not that you have an assistant to chat with, but that the pile which used to sit on someone's desk every morning is no longer there. What remains are the cases that genuinely deserve attention.

What you can do today

If you are currently experimenting with ChatGPT, Copilot or Gemini: by all means continue. It is the cheapest way to discover which documents in your business are hard and which are not. Paste a few order confirmations into a chat window and see how far the model gets in finding item, quantity and price.

What you learn there is immediately useful. If nine out of ten documents go well, you know reading is not your bottleneck and the gain sits in the connection. If it often goes wrong, you know exactly which suppliers or document types need attention first.

What you should not do is give a chat assistant access to your production administration without permissions, oversight and logging. You pay for today's time saving with next quarter's corrections.

Getting started

If you want to know what is possible in your Exact Online environment, a Quick Scan looks at your real document flow: which orders, invoices and tickets come in, in what form, and what share of them an AI coworker can take over. Within a week you know where the gain sits.

Typically an AI coworker is live within eight weeks in your own environment, connected to Exact Online, with your rules and your people in the loop. Plan your go-live whenever it suits you.