Accounting & Finance · Growing businesses
The right account on every bill, including suppliers who sell more than one thing
Bill coding for accounting practices on Xero and MYOB, tested on 59,886 real supplier payments against the supplier default bookkeepers correct today.
Executive summary
A learned coder sets the right account on 79% of supplier bills, against 67% for the supplier default, and posts 63% without review at 91% accuracy.
Situation
An accounting practice’s bookkeepers check and correct the account on every captured bill; the supplier default misses whenever a supplier sells more than one thing.
Approach
A coder from supplier name, history, amount and cost centre, fitted on 187,269 real payments and tested on 59,886 later ones against the supplier default.
Result
79% coded right against 67%; a 0.8 confidence bar posts 63% of bills without review at 91%.
The problem
In a bookkeeping practice the supplier, date and amount on a bill are now read reliably by Dext, Hubdoc or Xero’s own capture, and supplier rules post the simple ones automatically. The time that is left is the account code. Xero and MYOB suggest the account the supplier was coded to last time, a bookkeeper checks it, and the suppliers that sell several kinds of thing get corrected by hand, bill after bill. Practices that remove that step say it is what lets one bookkeeper look after two or three times as many client files.
I wanted to measure how often the supplier default is right on a large real ledger, how much a coder that also reads the description and amount adds, and how many bills could post without a review at an accuracy a practice principal would sign off.
- Team
- Practice principal, client managers, bookkeepers
- Scale
- Hundreds to thousands of bills a month across many client files
- Systems
- Xero or MYOB Business, Dext or Hubdoc, Xero Practice Manager or FYI
- Measure
- Bills posted without an account edit; bookkeeper minutes per bill
What I built
A coder that proposes the account for each captured bill from the supplier name, the supplier’s coding history, the amount and, where known, the client’s cost centre or tracking category, with a confidence score. Confident bills go to Xero or MYOB as drafts with the account set; the rest arrive in the bookkeeper’s queue with the top suggestions ranked.
Results
I tested on real supplier payments from a public ledger: 247,155 payments with the account chosen by the finance team across 131 expense accounts, fitted on payments before October 2025 and tested on the 59,886 that followed. The supplier default coded 67% of bills correctly. The learned coder reached 75% from the supplier and amount alone, and 79% when the cost centre was known.
Posting without review
At a confidence bar of 0.8 the coder posts 63% of bills with the account set, and 91% of those match the finance team’s choice. The bookkeeper reviews the rest. A practice sets the bar per client: tighter for a client with GST-sensitive coding, looser for a simple file.
| Confidence bar | Bills auto-coded | Accuracy on those bills |
|---|---|---|
| 0.60 | 78% | 83.7% |
| 0.80 | 63% | 90.8% |
| 0.90 | 53% | 93.2% |
| 0.95 | 46% | 96.3% |
Repeat and new suppliers
Repeat suppliers carried 94% of bills in the test year. For them the coder adds about seven points over the supplier default. For new suppliers there is no default at all; the coder still gets 25% right from the name alone, and 33% with the cost centre, which turns a blank field into a ranked shortlist.
Data and method
- Data
- London Borough of Camden, payments to suppliers over £500, Camden Open Data (Open Government Licence): 247,155 payments from April 2022 to October 2026 with supplier, expense account, cost centre, amount and date. Payments to individuals and redacted names were removed, and no supplier is named here.
- Target
- The expense account recorded by the finance team; accounts with fewer than 50 bills grouped
- Split
- Fitted on payments before October 2025; tested on October 2025 to October 2026
- Methods
- Supplier last-used account; supplier most common account; logistic regression on supplier-name character patterns, supplier history, amount band and, in a second version, cost centre
- Measures
- Share coded to the recorded account, by supplier history and amount band; accuracy and coverage at confidence bars
Limitations
- A council ledger stands in for a practice’s many client files. Its chart of accounts is larger than most small businesses’, which makes coding harder, and a practice’s per-client histories are shorter, which makes it harder in a different way. The coder is tested on the practice’s own files before use.
- The data has no line descriptions. Bills captured by Dext or Hubdoc carry line text, which adds signal.
- The recorded account is the finance team’s choice, which is sometimes inconsistent; part of the remaining error is disagreement rather than mistakes.
- GST treatment is checked separately; the coder sets the account only.
How I would take it live
- Export two years of coded bills for a pilot group of clients from Xero or MYOB
- Fit the coder per client with a shared supplier-name model, and agree confidence bars with each client manager
- Run in shadow for a month: the coder suggests, bookkeepers decide, and the edit rate is tracked
- Post confident bills as drafts with the account set; review the edit rate monthly in Xero Practice Manager or FYI