Margin and variance questions finance teams can ask an AI analyst
By SchemaChat (Data BI LLC) · Published 31 October 2026 · Last reviewed 31 October 2026 · 8 min read
An AI analyst is good at counting, summing, ratios and period comparisons over data it can see, and unreliable wherever the answer depends on accounting policy or judgement. This guide separates the two, walks through six practical questions, explains why you should always reconcile to the ledger, and lists what finance buyers should ask any vendor about controls. It also states what SchemaChat does not have today.
Formulas below are standard definitions. We give no benchmark values, because margins and ratios vary by industry and business model. Nothing here is accounting or financial advice.
What suits an AI analyst, and what does not
| Suited | Not suited |
|---|---|
| Totals and counts from ledger or invoice data | Deciding how to recognise revenue |
| Ratios such as margin and variance percentage | Accruals, provisions and estimates |
| Period comparisons (this month against last) | Anything that depends on a policy that is not in the data |
| Finding missing or inconsistent entries | Signing off a figure for audit |
Six worked questions
- Total spend by category, this month against last. A good answer shows both months side by side and the uncategorised amount. The trap: uncategorised spend hidden.
- Which categories moved by more than 10%? A good answer lists the amounts and direction. The trap: a threshold applied to tiny categories.
- Which vendors account for that movement? A good answer ranks vendors by contribution. The trap: the same vendor under two names.
- What is gross margin by product? A good answer shows the formula (revenue minus cost of goods, over revenue). The trap: costs booked in a later period.
- Which invoices are more than 30 days overdue? A good answer lists them with ages. The trap: paid invoices not marked paid.
- How does this month's total compare with the ledger? A good answer shows both figures and the difference. The trap: different cut-off dates.
A question it should refuse
Ask "What is our provision for doubtful debts?" If your data has no such field, a trustworthy tool says so instead of inventing a number. When you test a tool, include questions it should not be able to answer and see what happens.
Always reconcile to the ledger
Use the AI analyst to find, compare and explain. Before a number goes into a board pack, an audit file or an external email, reconcile it to the general ledger or system of record and keep the calculation with it.
Controls to ask about
- Is access to the database read-only, enforced by a database user you control?
- Is there an audit log of who asked what and who changed settings?
- Can you limit who sees which data sources?
- Which AI providers process the data, and are they listed as sub-processors?
What SchemaChat does not have today
SchemaChat does not currently offer single sign-on and has no SOC 2 report or other independent security certification. If your policy requires either, treat that as a blocker and ask any vendor the same questions. Our security page lists what we do have and what we do not.