Contact Info
Unit 9, 17 Power Ave
Alexandria, NSW 2015
info@hexflow.com.au
+61 (2) 8529 4085
Follow Us

Xero Bank Rules vs an AI Agent: Where Each One Stops

Xero bank rules are the cheapest automation a practice already owns, and most firms use about half of what they can do. A rule matches a contact, an amount or a keyword in the narration, then codes the line — a well-built set clears most of a bank feed before anybody opens it. The useful question is not whether rules beat an agent, but where your rules run out, because that is the only territory an agent gets paid for.

The short version

  • A rule is deterministic, explainable and free. It is also blind to anything nobody anticipated.
  • Rules fail in four predictable places: splits, one supplier with two codes, overpayments, and tax variants.
  • Run the coverage test before anyone quotes you a build. It takes five minutes.
  • Above roughly 85% coverage, the honest answer is to improve the rules and build nothing.

What a bank rule is actually doing

A Xero bank rule is an if-statement with a memory problem. The conditions are what you see on the statement line: the contact, a word in the description or reference, and an amount that equals, exceeds or falls below a figure. The action assigns an account, a tax rate and optionally a tracking option.

That is the whole mechanism, and it is a good one: fast, free, and explainable to an auditor in one sentence. What a rule cannot do is remember the last correction your bookkeeper made. Every case has to be anticipated in advance, and written separately for every organisation. Nothing carries across a client list.

Where bank rules stop working

Most of the searches that land on this topic are variations of “xero bank rules not working”. In our experience it is nearly always one of these:

Split transactionsOne deposit covering three invoices, or a withdrawal split across two accounts. The rule codes the total; the split is done by hand.
One supplier, two codesA hardware store that is sometimes materials, sometimes office supplies. The rule picks one, wrong half the time.
Overpayments and part-paymentsThe payment is coded, the overpayment becomes an exception nobody owns, and the ledger is out by the amount that matters.
Tax variantsMixed supplies and GST-free items share a supplier with taxable ones. The tax rate is part of the action.

The fifth failure is not a rule problem: a new supplier. No rule exists, the line lands uncoded, and the review queue grows with the client.

A rule only ever works on the cases somebody thought of in advance.

The five-minute coverage test

Before you compare products, measure what you already have. Pick two clients — your messiest and your tidiest — and one full month each.

  1. Count the transactions on the statement for that month.
  2. Count how many arrived already coded by a rule, without anyone touching them.
  3. Coverage is the second number over the first. Write both down for both clients.
  4. Now count the lines that were coded by hand and later corrected. That number is the real cost — it is work done twice.

Interpretation: under 60% is usually a chart-of-accounts problem, not an automation problem, and no agent fixes it. Between 60% and 85% is where an agent earns its keep, because the remaining work is the long tail. Above 85%, fix the rules and spend the money elsewhere.

What an agent does that a rule cannot

An agent reads the line, the contact’s history and often the attached document, then proposes a coding. The differences are behavioural:

  • It learns from corrections already made on that client, instead of from a rule someone had to foresee.
  • It can hold two plausible answers and say which it believes, rather than committing silently to one.
  • It can read the invoice when the narration is useless, which is most of the long tail.
  • It can be asked why, and answer with the evidence it used — the thing an auditor will ask for.

What it does not do is approve. An agent proposes; a person accepts. Every build we do stops at that line, deliberately.

Where an agent does not pay

Three conditions end the conversation early. If rules already clear most of your lines, an agent buys the last few percent at full price. If one person handles the file in under an hour a week, there is no hour to recover. And if the chart differs between clients for no good reason, an agent learns the difference faster than you can remove it.

In each of those cases the right purchase is a day of cleanup and a stricter set of rules. We have said that to firms at the end of a scoping call, and we will keep saying it.

What we build

We build agents that sit inside Xero: they read the bank line and the attached document, propose the coding, log what they proposed and why, and stop for a person to accept. The rules stay. The agent handles the tail they were never going to reach.

Send us one month

Send one client’s last month and we will tell you what your rules already cover, and whether an agent earns anything at all.

TALK TO US

Nirjhar works on the build side at Hex Flow in Alexandria, Sydney, where AI agents and automation are put together for accounting, legal, medical and NDIS practices.

Post a Comment