This post is not about setting up the integration - that is on the Claude page. This is about where Claude stands out once your administration is attached, and how to get the most out of that in practice.
Where does Claude stand out?
Claude is built around MCP and supports the protocol as a first party. In practice you notice that in two ways: it sustains longer analyses without losing the thread, and it is careful with figures it is unsure about.
That makes it suited to work made of several steps. Not "what is revenue", but "find the ten largest customers, look at how revenue compares to last year per customer, and show where the drop is largest". It holds chains like that together.
Dashboards from your own data
One of the more surprising uses: ask Claude to build a dashboard from your Exact Online data. It fetches the figures, structures them and delivers a working visualisation you can click through.
That works well for overviews that do not exist in that shape inside Exact Online: goods deliveries per week against order intake, hours per project next to budget, or utilisation per work center. You do not need a reporting module for it.
Three questions to start with
- "Give me outstanding items per customer, sorted by age, and flag which ones stand out." — Watch the reasoning behind "stand out"; it tells you whether it used the right fields.
- "Compare realised hours with budgeted hours per running project." — Surfaces the projects running over budget straight away.
- "Create an overview of stock value per warehouse and put it in a table." — A good test of whether your warehouse structure is right.
What you will run into
Ask for an intermediate step. With a complex analysis it is tempting to ask for the conclusion straight away. Ask for the underlying figures first, check those, and only then ask for the interpretation. It saves argument later.
Large datasets take time. Ask for every transaction in a full financial year and it will take a while. Narrow it by period or ledger account; that is faster and gives a sharper answer.
It says so when it does not know. That is a feature, not a fault. If you get "I cannot say this with certainty", that is usually justified - verify by hand rather than rephrasing the question until you do get an answer.
Where to go next
If you work with reports a lot, look at the prompt library for worked examples. If you want to know what the integration does with your data before relying on it, read the security page.
Why long analyses make the difference here
Most administrative questions are short: a balance, a status, a stock level. Any assistant is good enough for those. The difference only appears with questions made of several steps, where step three depends on what step two produced.
Think: find the projects running over budget, look per project at which hours caused it, and compare that with how similar projects went. That is not a question but a small investigation, and it stands or falls on holding context.
Working with documents alongside
A second strength: you can attach a document. A supplier quote next to your purchase history, a contract next to the invoices under it, a customer spreadsheet next to your own figures.
That opens questions your administration alone cannot answer: "Does this invoice match what the contract agreed?" Be careful here though - a document from outside can contain instructions not meant for you. With write access off that is at worst a wrong answer; with write access on it is a risk. See the security page.
What a good way of working looks like
Ask for the figures before asking for the conclusion. Then you can see where an interpretation comes from and check it.
Narrow it down. Requesting a whole financial year of transactions takes long and produces a less sharp answer than a bounded period or GL account.
New subject, new conversation. A long conversation jumping from stock to projects to invoices mixes context.
Worked examples are in the prompt library.