Explaining your company data to the AI: how the instruction approach works

The AI connector knows exactly how Exact Online works: which tables exist, which fields they contain, how you retrieve data. But it does not know how your specific administration is set up. And that is exactly where things go wrong.

An example. At many companies the account manager is not stored in a field called "account manager". Sometimes it is saved in the "ShippingMethod" field, sometimes in "Remarks", sometimes in a free field. Without an explanation, the AI searches in the wrong place and finds nothing, or worse, gives a confident but incorrect answer.

The solution: a company instruction

You write a short briefing once about how your administration differs from the standard. You save that briefing as a system instruction (ChatGPT) or project instruction (Claude). From that moment on, the AI knows it in every new conversation.

Step 1: map out your differences

Work through these questions:

• Which standard fields do you use for something other than intended?
• Do you have free fields (custom fields) with company-specific data?
• Do you work with multiple administrations? Which one is the main administration?
• Do you have agreements about VAT calculation (inclusive/exclusive)?
• Do you use specific general ledger accounts for reports?
• Are there customer groups, item groups or project codes that have a special meaning?

Step 2: write your instruction

Below is a fill-in template. Copy it, complete it and remove lines that do not apply.

COMPANY CONTEXT FOR EXACT ONLINE AI

Company: [name]
Industry: [e.g. wholesale, services, manufacturing]
Main administration: division [number]

--- DEVIATING FIELD NAMES ---

For Accounts (customers/suppliers):
- Field "ShippingMethod" = account manager
- Field "Remarks" = internal notes about payment behaviour
- Field "Classification1" = customer group (A/B/C)

For Items (products):
- Field "ExtraDescription" = internal product code
- Field "SearchCode" = EAN code

--- STANDARD SETTINGS ---

- Amounts always in euros (AmountDC), unless I ask otherwise
- Revenue always excluding VAT
- For outstanding items: use the Receivables table, filter on status 20 and 30
- Our financial years run from January through December

--- REPORTING PREFERENCES ---

- Show customers with customer number and name
- Compare periods with the same period last year
- Top lists default to top 10 unless I ask otherwise

--- TERMS THAT MEAN SOMETHING DIFFERENT AT OUR COMPANY ---

- "Projects" = we internally call these "assignments"
- "Cost centre" = we use this as location (Amsterdam, Rotterdam, Utrecht)

Step 3: save it

In ChatGPT

1. Click on your profile picture > Settings
2. Go to Personalisation > Custom instructions
3. Paste your company instruction under "What should ChatGPT know about you?"
4. Click Save

Every new chat automatically receives this context.

In Claude

1. Go to claude.ai > Projects in the sidebar
2. Create a project: "Exact Online [company name]"
3. Paste your instruction under Project instructions
4. Start new conversations from this project

Step 4: test it

Open a new conversation and ask a test question that touches your deviating setup:

"Who is the account manager of customer [name]?"

If the AI consults the correct field and gives the correct answer, your instruction works.

Not working? Check whether you used the exact Exact Online field name (case sensitive). If needed, ask: "Which fields are available for Accounts?" to find the correct name.

Common mistakes

For multiple roles

You can create different instruction sets for different users:

• Bookkeeper: focus on general ledger accounts, VAT, transaction lines
• Sales manager: focus on customers, revenue, outstanding items, account managers
• Controller: focus on margins, cost centres, budget comparisons

In Claude you do this with separate projects. In ChatGPT you can adjust the instruction per conversation role. Read more about the differences between Claude and ChatGPT in the next blog.

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