Transaction coding is the most time-consuming part of bank reconciliation for most bookkeeping practices. Manually assigning account codes and GST codes to hundreds of transactions per client per month adds up fast — and it is exactly the kind of repetitive, pattern-based work that can be largely automated without sacrificing accuracy. ReconLink's coding engine works in three layers, each adding automation on top of the last. This guide explains how each layer works and how to configure them for maximum efficiency.
Layer 1: Deterministic Coding Rules
The first and most reliable layer is the rule engine. A coding rule says: "Whenever a transaction description matches this pattern, apply this account code and this GST code." Rules match before any other processing happens, so they are instant and 100% predictable.
Rules are keyed on the vendor key — a normalised version of the transaction description with bank channel prefixes (EFTPOS, ONLINE, BPAY, PAYPAL, OSKO) and trailing reference numbers stripped away. This means a rule for "Officeworks" applies correctly whether the transaction reads "EFTPOS PURCHASE OFFICEWORKS PARRAMATTA 123456" or "ONLINE OFFICEWORKS.COM.AU REF 987654" — both normalise to the same vendor key.
Creating a rule manually:
- Navigate to any uncoded transaction and assign the account code and GST code you want.
- Click "Save as rule" — ReconLink will show you the normalised vendor key it will use.
- Confirm. The rule is immediately active and will apply to any future transaction with the same vendor key, across all clients in the practice if you choose to share it.
Rules can be scoped to a single client or shared across the practice. For vendors common to many clients — ATO PAYG, Telstra, office supply stores — practice-level rules save setup time for new clients.
Managing the rule library: Rules are listed in the Coding Rules section of the practice dashboard, sortable by date created, vendor key, and last matched date. Rules that have not matched any transaction in 90 days can safely be reviewed — either the vendor is no longer used by the client, or the vendor's transaction description format changed. Keeping the rule library clean prevents false matches.
Layer 2: Machine Learning Pattern Matching
When no deterministic rule matches a transaction, the ML model runs. It has been trained on historical coding decisions for the client (and optionally on anonymised patterns from across the ReconLink network) and suggests the most likely account and GST code based on the transaction's vendor key, amount, and day of week.
Layer 2 suggestions are shown with a confidence score. High-confidence suggestions (above 85%) appear with a green indicator; medium-confidence (60–85%) appear in amber. Below 60%, the transaction is flagged as needing manual review.
The ML model improves with use. Every coding decision you make — accepting a suggestion, rejecting a suggestion, or manually coding a transaction from scratch — updates the model's training data for that client. In the first month of use, Layer 2 accuracy is typically 70–80% for clients with consistent transaction patterns. By month three, well-maintained clients commonly see 90%+ accuracy on Layer 2 suggestions.
Key tip: Do not override Layer 2 suggestions without checking first. If the model is suggesting the wrong code for a transaction, it is likely doing so consistently — which means there is an opportunity to either create a Layer 1 rule (overriding it permanently and correctly) or to investigate whether the client's coding of similar transactions is consistent.
Layer 3: AI-Assisted Coding for Novel Transactions
When neither Layer 1 rules nor Layer 2 patterns produce a confident suggestion, the transaction goes to the AI coding layer. This uses an LLM to analyse the transaction description, the client's chart of accounts, and any industry context to suggest the most appropriate coding.
Layer 3 is deliberately conservative. It groups transactions by vendor before calling the LLM — so if a new supplier appears 12 times in a month, the AI processes it once and applies the suggestion consistently. This prevents noise and keeps the AI reasoning grounded in the actual transaction description rather than spurious one-off features.
AI suggestions always require human review before they are committed. They appear in the Review Queue with the AI's reasoning visible — you can see why it suggested a particular account code. This transparency is intentional: it lets you catch cases where the AI has misread the transaction (for example, confusing a capital equipment purchase with a consumable expense) before they are recorded.
Promoting AI suggestions to rules: When you accept an AI suggestion and the coding is correct, ReconLink asks whether you want to promote it to a Layer 1 rule. If you say yes, the vendor key is added to the rule library with the accepted account and GST code. From that point on, future transactions from the same vendor are handled by Layer 1 — instant, deterministic, and without needing further AI processing.
The promotion threshold is configurable. By default, auto-promotion only occurs when the AI's confidence score was 70% or above and you accepted the suggestion without modifying the code. If you edited the code during review, no rule is created — the model treats that as a case where human judgment was needed.
Setting the Confidence Threshold for Your Practice
The RULE_PROMOTION_MIN_CONFIDENCE setting (found in Practice Settings > Automation) controls the minimum AI confidence required for auto-promotion to be offered. The default of 0.70 is a good balance for most practices. Raising it to 0.85 or above means fewer rules are promoted automatically — appropriate if you want maximum control. Lowering it means more rules are promoted from fewer examples — appropriate if your clients have stable, predictable transaction histories and you trust the AI's suggestions.
Reviewing the Uncoded Queue Efficiently
Even with all three layers running, some transactions will need manual review each month. Prioritise the queue by:
- High-value uncoded transactions first — a $50,000 transaction miscoded matters more than a $12 one.
- Recurring vendor keys with no rule — if the same vendor appears uncoded three months in a row, it is time to create a rule.
- Low-confidence Layer 2 suggestions — these are where the model is least certain and human judgment adds the most value.
With a well-configured rule library and active AI promotion, most clients reach a state where 95%+ of transactions are coded automatically within a few months. The time that frees up is time you can redirect to higher-value work — reviewing reports, advising on cash flow, or managing more clients without adding hours.