Skip to main content

How Fiskl Measures Categorisation Accuracy

This guide explains how Fiskl measures the accuracy of its AI transaction categorisation. It sets out the result, the transactions measured, the rule for a correct categorisation and what the count leaves out.

The Result​

Fiskl's AI categorisation reached up to 98% accuracy in a measurement of 2.8 million transactions. Accountants at 37 accounting firms validated the categorisations at period end, over 16 months from July 2025 to October 2026.

MeasureValue
Highest accuracy reached98%
Transactions measured2.8 million
Measurement period16 months, July 2025 to October 2026
Accounting firms37
ValidationBy accountants, at period end
What drives accuracyThe business, and how many months of records it keeps in Fiskl

What Fiskl Measures​

Fiskl categorises every bank transaction as it enters the ledger. Each categorisation assigns the transaction to a ledger account in the business's own Chart of Accounts, with its tax treatment and its client or vendor.

Every categorisation applies as the transaction enters the ledger. The AI and history mark each one with a confidence level, and a rule marks its transactions with an R badge instead. Green is 80% and above, amber is 60% to 79%, and red is below 60%. The confidence level shows you which categorisations to check first. See how AI categorisation works.

The measurement counts every categorisation in the period, at every confidence level. It includes categorisations made by rules, by history and by Fiskl's AI models.

How the Figure Is Calculated​

A categorisation counts as correct when the accountant validated it at period end, as Fiskl assigned it. A categorisation the accountant changed counts as incorrect.

  • Correct categorisations: those the accountant validated at period end, as Fiskl assigned them
  • All categorisations: every categorisation in the measurement, across 2.8 million transactions
  • Accuracy: correct categorisations divided by all categorisations

Fiskl's AI learns from each business's own accounts. Accuracy therefore depends on the business and on how many months of records it keeps in Fiskl. The highest accuracy reached during the 16 months is 98%, which Fiskl publishes as up to 98%.

What the Count Leaves Out​

Corrections that accountants made to EU One Stop Shop (OSS) tax treatment are left out of the count. OSS is the EU scheme for declaring VAT on cross-border sales to consumers in other EU countries.

Training Data and This Measurement​

Fiskl's decision models are trained on more than 500 million accountant-certified transactions. That figure describes the training data. The up to 98% figure comes from the period-end validation described on this page.