Skip to main content

How AI Categorisation Works

Fiskl categorises every imported bank transaction with a proprietary AI decision system. This page explains how the system decides, how it learns from the decisions you confirm, and what stays under your control.

How Fiskl Decides​

Matching runs first. A bank transaction that matches an invoice, bill, payment or transfer is linked to it before categorisation starts. See How automatic matching works.

For every other transaction, Fiskl's decision system works through three layers in order. The first layer that can place the transaction decides it.

  1. Rules. A rule is the system's permanent memory for a merchant. Fiskl proposes rules from decisions you confirm, and you can write your own. A rule applies the same way every time.
  2. History. When you have confirmed a merchant before, the system repeats your decision.
  3. Fiskl's AI models. For merchants the system has not learned yet, Fiskl's proprietary models propose a ledger account with a confidence score. The models are trained on more than 500 million accountant-certified transactions. They categorise with up to 98% accuracy.

Fiskl's models decide first. For the remainder they cannot decide accurately, Fiskl asks an external language model for suggestions. Each suggestion goes through the same confidence test, and Fiskl writes it to your books only at 80% or higher.

Rules and history make results repeatable. The same transaction gets the same answer every time.

How the System Learns​

Every transaction you finish is a confirmed decision, and the system learns from it.

  • Finish a merchant once, and history repeats your decision for its next transactions.
  • Finish the same merchant the same way twice, and Fiskl proposes a rule.
  • Approve the rule, and every future transaction from that merchant arrives categorised and finished.

Each merchant moves from the AI models, to history, to a rule. The share of transactions that arrive already categorised grows each month, and you spend less time on review. The review strip on Accounting > Transactions is built to get your recurring merchants to a rule quickly. See Finishing Transactions for the steps.

Calibrated Confidence​

Each AI answer carries a confidence score, shown as a colour: green at 80% and above, amber from 60 to 79%, and red below 60%.

Fiskl writes green answers to your books automatically. The 80% threshold was calibrated on real books. Answers between 70 and 79% were wrong about one time in three, so writing them would add more errors than it saves work.

Answers below 80% are kept as proposals. When you run Categorise on a selection, they appear for you to accept or change. See Categorising a Selection.

What Finish means​

Finish confirms a transaction is right. Until you finish it, a transaction the system categorised belongs to the system. After you finish it, the decision is yours, and the system learns from it.

This distinction matters for two reasons:

  • History learns only from finished transactions. A category the AI models wrote is not repeated for the next transaction from that merchant until you finish it. One unconfirmed prediction never becomes the answer for later transactions.
  • Rule suggestions come only from finished transactions. Fiskl proposes a rule when you have finished the same merchant, the same way, at least twice. AI-categorised transactions you have not finished do not count, however confident the models were.

Transactions a rule places need no Finish. A rule is your own instruction, so there is nothing to review.

Where the work happens​

Day to day, everything happens on the transactions list. The switches and the full record of rules live in Settings.

What you want to doWhere
Review what the automation did, finish it, and turn a merchant into a ruleAccounting > Transactions: the review strip above the list and the selection bar
Categorise a chosen set of transactions nowAccounting > Transactions, select rows, then Categorise
Turn a step on or offSettings > AI & Automations > Automations
Re-run rules and history over everything still uncategorisedSettings > AI & Automations > Automations > Retry categorisation
Describe your business and bookkeeping preferences for the AISettings > AI & Automations > AI profile
See every rule across all accounts, edit or delete one, dismiss a suggestionSettings > AI & Automations > Rules
info

Auto Categorisation runs rules and history. AI Categorisation runs Fiskl's AI models. Each has its own switch under Settings > AI & Automations > Automations, and AI Categorisation is included in every plan. See AI & Automations Settings.

What Stays Under Your Control​

Some things are deliberately left to you:

  • Tax rates. Tax rates come from your rules and from merchants you have already confirmed. The AI models do not set them.
  • Locked periods. Nothing is written to a transaction dated inside your locked period. It waits until the lock moves. See Locked Period.
  • Uncertain answers. Below 80% confidence, the transaction waits for your decision. Use Categorise on a selection, or Retry categorisation under Settings > AI & Automations > Automations.

Audit Trail​

Every automated categorisation is recorded in the Audit Log as a System change, naming the process responsible.

Common Issues​

The same merchant keeps coming back uncategorised

Fiskl looks at each uncategorised transaction once. If nothing could place it, it is left for you and not retried on its own. To settle the merchant for good, open Accounting > Transactions, select its transactions, set the account, and use Finish and create rule. The rule places every future transaction from that merchant. To catch the ones already waiting, select Retry categorisation under Settings > AI & Automations > Automations, or select the transactions and use Categorise.

A categorised transaction has no tax rate

Tax rates come from your rules and confirmed history, so a merchant's first transaction can arrive without one. Add the rate and finish the transaction. History carries the rate to later transactions from the same merchant, or you can add a rule that names it.

A transaction was categorised to an account that makes no sense for my business

Correct the account on the transaction and finish it. If it happens repeatedly for one merchant, use Finish and create rule so the AI is never asked about it again. A description on the ledger account under Chart of Accounts also helps Fiskl's AI tell similar accounts apart, and so does a fuller AI profile.