What can you ask the AI, and what not? Concrete examples

With the Exact AI Connect connector you ask questions about your bookkeeping in plain English. That works surprisingly well, but not every question gives you a useful answer. Below are the most important patterns.

Checklist met voorbeelden van goede en slechte AI-vragen

With the Exact AI Connect connector you ask questions about your bookkeeping in plain English. That works surprisingly well for most daily tasks. But not every question gives you a usable answer. Some questions are too vague, others are about things that simply are not in your records.

Below are the main patterns, so you know right away what to expect.

Questions that work well

Specific lookup questions

The AI excels at targeted questions with clear criteria:

  • “Which invoices from customer Jansen are still open?”
  • “What is the phone number of supplier Bakker?”
  • “Show the last 5 entries on general ledger account 8000”
  • “How much stock do we have of item X?”

Why this works: the question is specific enough to determine the right table and the right filter.

Standard financial reports

  • “Show the profit and loss statement for Q1”
  • “How does the balance sheet look?”
  • “Give an overview of the outstanding accounts receivable”
  • “Prepare the VAT return”

Why this works: these are fixed report formats that Exact Online supports as standard.

Comparisons with concrete periods

  • “Compare the revenue of March with February”
  • “Is the revenue for Q1 this year higher than last year?”
  • “Which month had the highest revenue?”

Why this works: the AI can interpret dates and perform calculations on the retrieved data.

Actions with clear instructions

  • “Create a new customer: Bakkerij Bloem, Amsterdam”
  • “Create a sales invoice for customer X: 3 hours of consulting at 150 euros”
  • “Update the email address of contact Y to info@voorbeeld.nl”

Why this works: concrete data, no doubt or ambiguity. The AI always shows a preview before it makes any change.

Questions that work poorly

Too vague, no concrete criterion

  • “How is the business doing?”
  • “Are there any problems?”
  • “Give me an overview”

Why not: the AI does not know where to look. “An overview” could be outstanding items, revenue, cash flow or a hundred other things. Rephrase: “Give an overview of the outstanding accounts receivable, sorted by amount.”

Advice and interpretation

  • “Should I write this off?”
  • “Is this customer reliable?”
  • “What should I do to improve my cash flow?”

Why not: the AI shows facts, not judgments. It can show that a customer pays 3 invoices late, but the conclusion “this customer is unreliable” is up to you.

Data that is not in Exact Online

  • “What does the customer think of our service?”
  • “How many visitors did the website have last month?”
  • “What are the market prices for raw material X?”

Why not: the connector only gives access to your Exact Online records. Information from other systems (CRM notes, website statistics, external market data) is not available through this connector.

Predictions about the future

  • “How much revenue will we make next month?”
  • “When will customer X pay their invoice?”
  • “How will our margin develop in the coming months?”

Why not: the AI can analyse historical data but cannot predict the future. What it can do: “Show the revenue per month for the past 12 months”, so you can spot the trend yourself.

The borderline cases

Some questions sit in between. They work, but only if you phrase them well:

Question (vague) Phrased better Why better
“Show the sales” “Show the sales invoices for this month, sorted by amount” Period + sorting makes the question specific
“Who pays badly?” “Which customers have invoices that have been open for more than 60 days?” Concrete criterion instead of a value judgment
“Give the margins” “Calculate the gross profit per product group based on SalesInvoiceLines and Items” Specific tables and calculation named
“How much VAT?” “Prepare a VAT overview for Q1, with supplies and acquisitions listed separately” Period and level of detail specified

Rules of thumb

  1. Ask questions whose answers you can check. Start with something you know, then build on it.
  2. Name a period. “Revenue” is vague. “Revenue for March 2026” is concrete.
  3. Name the location in Exact if you know it. “This stops the AI from spending a long time searching in the wrong data.
  4. Do not ask for opinions. The AI gives facts, not advice.
  5. Split up complex questions. Three simple questions in a row give more reliable results than one all-encompassing question.

Read in this blog more about what AI can mean for your business records.

Want to try it yourself? Start for free with Exact AI Connect →