# Transaction Categorization Software: What to Compare

- Author: AutoBankConvert (https://www.autobankconvert.com)
- Published: 2025-01-07
- Updated: 2025-01-07
- Category: Choosing software
- Canonical: https://www.autobankconvert.com/blog/transaction-categorization-software

> Compare transaction categorization software on client-specific rules, learning from approvals, review queues and how it handles vendors it has not seen.
Transaction categorization software assigns each bank or card transaction to an account in the client's chart of accounts. To compare tools for client work, test four things on your own statements: whether each client keeps separate rules, whether a correction sticks next month, how clearly uncertain lines are held for review, and what it does with unfamiliar vendors.

## What does transaction categorization software decide?

Transaction categorization software decides which account each transaction belongs to, based on the description, the amount and sometimes the date. For a bookkeeper, the output is a proposed account per line plus a signal about how sure the tool is. The account is the visible part. The signal is what saves review time.

A line such as "AMZN MKTP US 4K2LJ" could be office supplies, inventory, equipment or a personal purchase. The tool picks one, or picks one and marks it unsure, or leaves it blank. Those three outcomes cost very different amounts of your time:

- **Right and marked sure.** No work.
- **Wrong and marked sure.** The expensive case, because nothing tells you to look.
- **Marked unsure, right or wrong.** A few seconds in a queue.

So the question to ask a vendor is not only how often the tool is right. It is how often it is wrong while claiming to be sure. A tool that is right less often but honest about doubt can cost you less than one with a higher hit rate and silent mistakes.

Categorization usually runs after statements are turned into transactions. If the description text arrives garbled from a poor scan, no categorizer can fix it, so clean intake comes first. The guide to [document scanners for bookkeepers](/blog/document-scanner-for-bookkeepers) covers getting readable images from paper statements.

## Client-specific categorization rules versus shared rules

Client-specific categorization rules apply only to one client's books, while shared rules apply across every client in the practice. A practice needs both, with client rules winning. The same vendor often belongs in different accounts for different clients, and a tool that only knows shared rules will code one of them wrong every month.

Take a payment to a home improvement store. For a contractor client it is job materials, probably charged to cost of goods sold. For a law firm client it is repairs and maintenance. For a sole proprietor it might be personal. One rule cannot serve all three.

What to check:

- Can you write a rule that applies to one client only?
- Can you write a firm-wide default, such as software subscriptions to one account, that each client can override?
- When a client rule and a shared rule conflict, which wins, and does the tool show you why a line was coded the way it was?
- Can each client use its own chart of accounts, with account names that differ from other clients?

Separation is also a safety matter. A rule edited for one client should never change another client's books. The post on [bookkeeping for multiple clients](/blog/bookkeeping-for-multiple-clients) covers how to keep each client's data apart across the whole workflow, not just in categorization.

## Does categorization software learn from your approvals?

Categorization software that learns from approvals turns each correction into a rule or a weighting, so the same vendor is coded right next month without a second correction. Ask to see it in the demo: correct one line, load the next month, check that vendor. If you must fix the same vendor twice, the tool does not learn.

There are two ways tools learn, and both are fine if they are visible:

- **Rule creation.** Approving a correction offers to save it as a rule, which you can read and edit later.
- **History matching.** The tool looks at how you coded similar descriptions for this client in past months and follows that.

The risk with learning is that a mistake also gets learned. If a personal charge is approved as a business expense once, a learning tool may repeat it every month. So check that you can see, edit and delete what the tool learned, per client.

AutoBankConvert combines both approaches: rules you write once, plus each client's past coding. Lines it is unsure about go to a review queue with a suggested category, and approving one teaches that client's rules for next month.

## The categorization review queue as a buying test

A categorization review queue is the list of transactions the software held back because it was not confident enough to code them alone. Treat it as the main buying test. The queue shows how the tool behaves when it does not know, and that is where your time goes each month after setup.

Look at the queue on a real client month and judge it on four points:

1. **Size.** After a few months of learning, the queue for a steady client should shrink to new vendors and genuinely odd lines.
2. **Suggestions.** Each held line should carry a proposed account, so most approvals are one click.
3. **Reasons.** The tool should say why a line is held: new vendor, amount far from usual, or conflicting rules.
4. **Speed.** You should be able to clear thirty lines in a few minutes, including batch approval of lines from one vendor.

A tool with no queue at all, which codes everything and leaves you to find mistakes, fails this test outright.

## Unseen vendors and categorization guesses

Unseen vendors are payees that have never appeared in a client's history and match no rule. They test what the software does without evidence. A good tool proposes an account from the description and similar vendors, marks the line unsure and sends it to review. A weak one picks a default account and moves on silently.

Watch for these patterns with new vendors:

- **Catch-all coding.** Lines dumped into an "uncategorized" or "ask my accountant" account without a flag.
- **Overconfident guesses.** A new vendor coded with the same confidence as rent paid for three years.
- **Name variants treated as new.** "SQ *JOES COFFEE" and "JOES COFFEE #12" treated as two vendors, so each gets its own learning.

Income lines need a separate look. Deposits from a client's customers often show only a transfer reference, and categorization software tends to guess these worst. If clients invoice through their own system, check whether deposits can be matched to invoices instead. The post on [invoicing software for bookkeepers](/blog/invoicing-software-for-bookkeepers) covers that side.

## Categorization software comparison worksheet

A categorization comparison worksheet runs the same month of client statements through each tool and counts results line by line. Use at least one month of a real client with lines you already coded by hand, so you have an answer key. The figures below are examples from a twenty-line trial, to show the arithmetic.

### A worked twenty-line trial

| Result on 20 example lines | Tool A | Tool B |
| --- | --- | --- |
| Coded right, marked sure | 13 | 15 |
| Coded wrong, marked sure | 1 | 4 |
| Held for review with right suggestion | 5 | 1 |
| Held for review with wrong suggestion | 1 | 0 |
| First-pass accuracy (sure and right, out of 20) | 65% | 75% |
| Silent errors | 1 | 4 |

Tool B looks better on first-pass accuracy, 15 out of 20 against 13. But four of its sure answers are wrong, and nothing points you to them. To find those four you would have to review all twenty lines, which wipes out the benefit. Tool A sends six lines to review, five with the right suggestion, and has only one silent error. In this example Tool A saves more time.

### Six criteria to score

Score each tool from 0 to 5 on:

- Silent error count on your test month.
- Per-client rules and chart of accounts.
- Learning from approvals, visible and editable.
- Review queue size, reasons and batch approval.
- Handling of new vendors and income lines.
- Fit with the ledgers your clients use, desktop or cloud.

On the last point, a categorizer that lives inside one cloud ledger will not help clients on other systems. The comparison of [desktop and cloud accounting software](/blog/download-accounting-software-or-cloud) helps you see how your client list splits.

## When categorization software adds work instead

Categorization software adds work when a client has too few or too varied transactions for rules to pay off, or when the setup takes longer than coding by hand would. A client with a dozen lines a month, half of them different each time, may be faster to code directly than to train a tool for.

Other cases where it can cost more than it saves:

- **One-off projects** such as a single catch-up year for a client you will not keep.
- **Very mixed accounts** where personal and business spending run through the same card and most lines need a question to the owner.
- **Charts of accounts in flux**, where the client is still deciding how they want expenses grouped. Rules written now will need rewriting.

There is also a setup cost that vendors rarely mention. The first month on any categorizer is the slowest, because the tool has no history for the client and few rules. You will approve more lines in month one than in month three. Judge a tool on its second or third month with a client, not its first, and budget the extra review time when you move a whole client list at once. Moving clients in waves, starting with the ones whose vendors repeat most, spreads that cost out and shows you early whether the learning works.

For the steady monthly client with a few hundred lines, the case usually goes the other way. If you want to test categorization on statements you already know, [request a spot with AutoBankConvert](/get-started) and compare its review queue against your own coding for the same month.