Up to 98% accurate: how Fiskl measures AI categorization

LinkedInXFacebookThreads
Fi Close Agent summary of a September close in Fiskl: 12 checks run, 11 proposals awaiting the accountant's approval, 2 tasks waiting and 1 blocked
3 min. read

Fiskl’s AI categorizes every bank transaction as it enters the ledger. It assigns the ledger account in the business’s own chart of accounts, with the tax and the client or vendor.

We measured how often accountants kept those categorizations. Across 2.8 million transactions over 16 months, accuracy reached up to 98%.

Below is the method behind the figure.

What we measured

The measurement ran from July 2025 to October 2026. It covered 2.8 million transactions across 40+ currencies and seven countries. Accountants at 37 accounting firms validated the categorizations at period end.

Every categorization counts. That includes every confidence level, green, amber and red. It includes categorizations made by rules, by the business’s own history and by Fiskl’s AI models.

What counts as correct

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

Accuracy is the share of categorizations the accountant validated unchanged.

One thing is left out of the count: corrections accountants made to EU One Stop Shop (OSS) tax treatment.

Why “up to”

Fiskl’s AI learns from each business’s own accounts. Accuracy depends on the business and on how many months of records it keeps in Fiskl. The more history a business has in Fiskl, the more its categorizations reflect how that business works.

The highest accuracy reached in the measurement is 98%. That is why we publish it as up to 98%.

How it works every day

AI Categorization is on by default for every new company. There is no queue and no separate posting step. Each transaction enters the ledger already categorized.

A confidence level marks every categorization:

  • Green: 80% and above
  • Amber: 60% to 79%
  • Red: below 60%

The confidence level sets the order for review. Confirmed decisions become history and rules, so later transactions from the same merchant are categorized the same way. Corrections also feed Fiskl’s learning loop. Fiskl’s decision models are trained on more than 500 million accountant-certified transactions.

Where it fits in the close

Categorization is the first step of Fiskl’s autonomous close with human oversight. Fiskl’s AI and deterministic engine categorize, match, reconcile and calculate journals, tax and FX continuously as transactions arrive. The Fi Close Agent, live with selected accounting firm partners, then works each remaining exception. It prepares every correction with evidence for the accountant to authorize. The accountant locks the period.

Read the full method

The full method, with the figures, the rule for a correct categorization and what the count leaves out, is on our help site: How Fiskl Measures Categorization Accuracy.

To see how the layers fit together, read How Fiskl AI Works and Fi Close Agent.

Try Fiskl free for 30 days