You ask your AI assistant for revenue per month over the past year. Back comes a tidy table: twelve months, twelve amounts, a total underneath. The shape is perfect. Only the answer is wrong, and nothing on screen says so.
Why do large analyses stall on an AI connector?
Because the limit is not the number of AI requests you may make per month, but the number of records the connector returns per question. If a connector retrieves at most 60 invoices and you have 400, the AI dutifully calculates over those 60. The result is a sample presented as a yearly total.
Where it goes wrong
Four questions, and where they break on a connector that fetches a limited number of rows per question:
| Question | What goes wrong |
|---|---|
| “Show the top 5 customers by revenue” | Works, but only over the invoices fetched first - the actual number 1 may fall outside |
| “Compare revenue per month across the whole year” | Incomplete as soon as you have more invoices per month than the limit |
| “Which product group yields the highest margin?” | Requires joining invoice lines to items; unreliable on a subset |
| “Show payment behaviour per customer group over 2 years” | Too much data and too many combinations; stalls or returns a sample |
The annoying part is not that it cannot be done. The annoying part is that the answer looks exactly like a good answer.
Two routes to the same question
There are two ways to get an AI assistant to Exact Online.
Straight through. Every question is translated into calls on the Exact Online API. Always current, and fine for lookups. But the API returns data in pages, Exact enforces a daily call budget, and so nearly every connector puts a ceiling on records per question. That ceiling is exactly where your analysis breaks. Exact AI Connect from iWebDevelopment sets that ceiling at 60 records per question in its Essentials plan.
Through a cache. The Exact Online data is kept in a dedicated database per administration. A question about revenue per month then becomes an aggregation in that database: one operation across all rows, not a hundred separate fetches. That is how Ledger Botje works.
That difference shows up in requests too. If the aggregation happens in the database, a yearly analysis costs a single AI request. If the assistant has to pull everything in page by page and add it up itself, the same question costs dozens.
So how much do you get back?
With Ledger Botje it depends on the tool. A batch or serial number history goes up to 500 rows, a trial balance up to 1,000 accounts, a stock overview per location up to 200 rows. But the important category is the aggregate: revenue per period, margin per product group, balance per ledger account. Those are calculated in the database over the complete dataset, not over a retrieved selection. The number of records you get back is then the number of groups in your answer - twelve months, eight product groups - not the number of underlying rows.
Does the cache stay current?
Fair question, because a wrong answer from a stale copy is just as wrong as one from half a dataset. Three things keep it straight:
- Through the evening the administrations are updated every fifteen minutes, with only the changes since last time.
- If a tool needs data older than a day, that data is refreshed before the question is answered.
- After every write action the relevant part is reloaded immediately, so an order you just created is in the next answer.
- And you can always force it yourself: one button in the app reloads the administration data from Exact Online.
For an analysis across two years a few hours make no difference. For the question which invoices fell due today they do, and that is what the automatic and manual refreshes are for.
Checking the answer yourself
You do not need to know how a connector works internally to notice you have half a dataset. Ask:
“How many invoices did you base this on, and what is the earliest and latest invoice date in that set?”
If a suspiciously round number comes back - 50, 60, 100 - while you expect more, you are looking at a truncated list. Ask the same per month afterwards; if months are missing, or January has forty invoices and December four, you know enough.
This is the same check you would run on a colleague who turns up with an export: how many rows were in it, and from when to when. Do it once on a new connector and you know where your limit sits.
What this means for your choice
If you use the assistant to look things up - this customer, that invoice, this item - the record limit is not your problem and you can choose on price. See what does an AI connector for Exact Online cost.
If you want to use the assistant to analyse, this is the first question to put to a vendor: how many records do I get per question, and are aggregates calculated over the complete dataset? The full comparison on this and other points is in AI connectors for Exact Online compared.