Confidence scores explained

What is a confidence score?

Every transaction classification AccBooks makes is accompanied by a confidence score — a number from 0 to 100% that indicates how certain the AI is about its suggested nominal code.

• 90–100%: Very high confidence. The AI has seen this merchant many times and the amount is within the expected range.
• 60–89%: Moderate confidence. The classification is likely correct but worth a quick check.
• 0–59%: Low confidence. The AI is guessing — manual classification is recommended.

How the score is calculated

AccBooks uses a multi-factor scoring model:

Merchant recognition (40% weight)

The AI maintains a merchant database of over 2 million UK businesses, matched against the bank description using fuzzy string matching. A clean match with a known merchant contributes the most to a high score.

Historical posting pattern (30% weight)

For merchants you’ve classified before, AccBooks looks at how consistently you’ve used the same nominal code. If you always post Amazon purchases to 6100 (IT equipment), that history adds confidence. If you sometimes post to 6100 and sometimes to 6050 (office expenses), the score is lower.

Amount pattern (15% weight)

Recurring transactions of similar amounts score higher. A monthly £49.99 charge from a software subscription is highly predictable; a one-off payment of £7,432.17 is less so.

VAT consistency (10% weight)

If the suggested nominal code’s default VAT treatment matches the amount’s implied VAT content, the score increases. Discrepancies (e.g., a supplier who should be VAT-registered but the amount doesn’t suggest VAT) lower the score.

Context signals (5% weight)

Time of month, account, and category context. For example, payroll transactions on the 28th of the month, from a known payroll provider, get a boost.

Improving your scores over time

Confidence scores improve as AccBooks learns from your corrections. Every time you:

• Change a classification, AccBooks updates its model for that merchant.
• Accept a suggestion, AccBooks reinforces it.
• Create a reconciliation rule, AccBooks applies it with 100% confidence.

Most businesses see scores improve significantly within the first 3 months of use.

Using scores in your workflow

High-confidence batch (90%+): Review the total value and spot-check 2–3 transactions. If they look right, use Approve all high-confidence to clear them in one click.

Medium-confidence (60–89%): Open each transaction and review the suggested code. Accept or correct, then approve.

Low-confidence (0–59%): Classify manually. Consider creating a reconciliation rule for recurring low-confidence transactions.

Viewing score details

Click any transaction in the reconciliation queue and scroll to the AI explanation section. This shows:

• The confidence score
• The top factors that contributed to it
• Alternative classifications the AI considered (with their scores)
• A link to the merchant’s database entry

Exporting score data

Go to Reports → Reconciliation analytics to export a CSV of all transactions in a period with their confidence scores. This is useful for auditors who want to understand how much of the ledger was AI-classified vs. manually reviewed.