# Categorization Confidence Score: When to Trust It

- Author: AutoBankConvert (https://www.autobankconvert.com)
- Published: 2024-11-21
- Updated: 2024-11-21
- Category: Statement automation
- Canonical: https://www.autobankconvert.com/blog/categorization-confidence-score

> What a categorization confidence score means on an automated transaction, which score to trust without review, and when a low score points at a bad rule.
A categorization confidence score is a tool's estimate of how likely its suggested account is right for one transaction. It measures agreement between the line and the evidence behind the suggestion, not truth. Trust high scores on vendors with a steady history, review everything below a threshold you set per client, and treat repeated low scores as a rule problem.

## What a categorization confidence score measures

A categorization confidence score measures how strongly the evidence points to one account for a line. That evidence is usually a matching rule, the client's past coding of the same payee, or both. A score near the top means the evidence agrees; a middling score means it is thin or split between accounts.

The score is easy to misread as a probability that the books are correct. It is narrower than that. A tool can be very sure a line from a hardware store belongs in Supplies because every past line from that store went there, and still be wrong this month because the owner bought a personal grill. The score describes the match, not the purchase.

Tools also differ in how they build the number. Some show a percentage, some show high, medium and low bands, and some just route uncertain lines to review. AutoBankConvert's homepage does not mention a numeric score; it says transactions the rules aren't sure about go to a review queue with a suggested category. Whatever form your tool uses, the decisions below are the same.

## Which categorization confidence score can skip review?

No single categorization confidence score is safe for every client. A workable starting point is to let only the top band skip review, and only for vendors the client has used for several months. Then sample those approved lines each month and lower the threshold only after the samples come back clean.

The table shows how one client's lines might be handled. The scores, vendors and actions are examples.

| Line | Score | Suggested account | Action |
| --- | --- | --- | --- |
| COASTAL POWER AUTOPAY | 98 | Utilities | Accept, steady monthly vendor |
| ADOBE CREATIVE CLD | 95 | Software subscriptions | Accept |
| HOME DEPOT 4410 | 91 | Supplies | Review, mixed-use vendor |
| AMZN MKTP US | 64 | Office supplies | Review, shared marketplace |
| ZELLE FROM J RIVERA | 41 | Sales | Review, ask client |
| SQ *KT EVENTS | 22 | none | Review, first appearance |

The third row is the one to notice. A high score on a vendor known for mixed business and personal spending is not a reason to skip review. Vendor type matters as much as the number, so keep a short list of payees that always go to a person regardless of score.

Sampling keeps the threshold honest. Pick ten accepted lines at random each month and check them against receipts or the client's answer. If every sample is right for three months, the threshold is doing its job.

## Low confidence scores that point at a bad categorization rule

Low confidence scores point at a bad categorization rule when the same vendor scores low month after month. One low score is a hard line. A pattern of low scores on a familiar payee means the rule is not matching the description text the bank actually prints, or two rules are pulling the line toward different accounts.

Three causes account for most of these patterns:

1. **The description changes each month.** A store number, date or reference code inside the text stops an exact rule from matching. Cleaning payee names into one stable form fixes the cause; the post on [vendor name normalization on bank statement lines](/blog/vendor-name-normalization-bank-statements) covers how.
2. **Two rules overlap.** A broad rule such as "AMAZON" and a narrow one such as "AMAZON WEB SERVICES" both fit, and the tool splits its confidence between them.
3. **History disagrees with the rule.** If past coding sent a vendor to one account and a new rule sends it to another, the evidence conflicts until history catches up.

When you see the pattern, fix the rule, not the lines. Approving the same low-scoring vendor by hand every month treats the symptom. The step-by-step guide to [automation exception handling in bookkeeping](/blog/automation-exception-handling-bookkeeping) shows how to sort held lines by cause so these rule problems stand out from one-off purchases.

## Can a high categorization confidence score still be wrong?

A high categorization confidence score can still be wrong whenever the history behind it is wrong or the purchase changed while the payee stayed the same. The tool is confident because the evidence agrees, and evidence built on a past mistake agrees with itself perfectly.

Watch these cases even at the top of the scale:

- **Inherited miscoding.** If a previous bookkeeper coded owner draws as an expense for two years, history will score that expense highly forever.
- **Payment processors.** A deposit from a card processor looks the same every month, but it is net of fees and refunds, so the account may be right while the amount still needs splitting.
- **Transfers.** A transfer between the client's own accounts can score high as income if earlier transfers were miscoded that way.
- **New services from an old vendor.** A software vendor that sold one subscription may now bill a second, unrelated one under the same name.

None of these shows up as a low score. They show up in a balance sheet that drifts, or in a client asking why a number looks wrong. A review of the top-scored vendors when you take on a client is the defence. The overview of [automated bank statement conversion, stage by stage](/blog/automated-bank-statement-conversion) shows where categorization sits after the balance check, which is why a correct statement can still carry wrong accounts.

## Setting confidence thresholds per client account

Confidence thresholds work better per account than per client. A utility account can accept a lower score without harm, because errors are rare and small. Owner draws, loan payments and sales should demand the top band or always go to review, because a wrong line there changes tax figures and the balance sheet.

A simple tiering that many practices could adapt:

- **Always review**: owner draws and contributions, loan payments, sales and income, anything over a dollar amount you choose.
- **Top band only**: supplies, meals, travel, marketplace purchases.
- **Middle band and up**: rent, utilities, software subscriptions, insurance.

Write the tiers into the client's notes so a second bookkeeper applies the same standard. If a client wants every line reviewed whatever the score, that is extra work worth naming in the engagement; the guide to [scope creep in bookkeeping](/blog/scope-creep-in-bookkeeping) covers how to spot and price requests like that. To see how uncertain lines reach a queue with a suggestion attached, [AutoBankConvert](/) shows the flow from upload to export.