Training Matching Rules
"Training" in REDA Financials' matching isn't a data-science project — it's three concrete habits that compound: your matches teach the AI, your patterns become rules, and your rules retire whole categories of work. Here's the practical playbook for going from "we match by hand" to "the reconciliation mostly does itself".
Access you'll need. Reconciling and creating Bank Account Rules requires the Accounting Admin v2 or Accounting Controller v2 permission set. The Accounts Payable v2, Accounts Receivable v2, and Read Only v2 sets can view but not change these.
The improvement loop
Match consistently — you're writing the textbook
Every manual match you apply joins the AI's learning context (it studies up to a thousand of your past decisions). Consistency is the curriculum: always match the merchant settlement to the receipt (not to a journal), always book fees as adjustments — and the AI's suggestions converge on your house style.
Promote repeat patterns to rules
Anything monthly deserves a Bank Account Rule: a distinctive text pattern, a document type, your GL account, optionally Auto Post. The AI even proposes rule candidates when it spots recurring shapes — accept the idea by creating the rule. From then on that line matches at machine speed, forever, no AI needed.
Audit what the automation does
Rules carry their own scorecard — Times Applied, Times Overridden and last-applied date — and every matched line records how it matched (Auto, Rule-Based, AI-Suggested, Manual, Adjustment). A monthly skim of rule hits catches a pattern gone stale before it mis-books anything.
Rule-writing craft
- Distinctive fragments beat short ones: "MONTHLY ACCOUNT SERVICE FEE" not "FEE"; use the regex form for precise formats.
- Priority = specificity: first match wins, so the narrow rule outranks the broad one.
- Auto Post is earned: new rules run reviewed (drafts) for a cycle or two, then graduate.
- One account at a time: rules belong to a bank account — the operating account's patterns aren't the payroll account's.
What "training" is — and isn't
It's worth being precise, because the word carries baggage. In REDA Financials, the AI learns from context, assembled fresh from your own records each time it runs — there is no model being retrained on your data, no training console, and nothing for an administrator to configure:
| Training is… | Training is not… |
|---|---|
| Your past manual matches (up to a thousand of them) shown to the AI as worked examples on every run | A data-science project with datasets to label or upload |
| Your active rules shown alongside, so suggestions don't duplicate what automation already covers | A model that drifts on its own — the context is rebuilt from your records each run |
| Doing the work consistently, because consistency is the curriculum | A settings page — there are no training knobs to turn, and nothing to switch on |
| Reversible: correct your habits going forward and the context follows | Permanent: no bad decision is baked in forever |
The other classroom: invoice extraction learns the same way
Bank matching isn't the only feature that studies your history. OCR invoice processing uses the identical idea on the payables side: when the AI extracts a vendor's bill, it is shown that vendor's recent posted purchase orders and invoices — including which product and GL account your team chose on every line. The result is a product-to-account map built from your own decisions: code MetroPower's supply charge to Utilities once, and the next MetroPower bill arrives pre-coded the same way.
The practical consequence is the same playbook as matching:
- Post consistently — only posted documents teach; a pile of unposted drafts teaches nothing.
- Fix coding at review time, not after posting — the correction you make in the pre-filled form becomes the example the next extraction studies.
- Expect the first bill from a new vendor to need you — there's no history yet; the second one is where the learning shows.
Common mistakes to avoid
| Mistake | Why it hurts | Do this instead |
|---|---|---|
| Two clerks matching the same pattern two different ways | The AI's textbook contradicts itself — confidence drops and suggestions hedge | Agree the house style once (fees are adjustments, settlements match receipts) and both follow it |
| Leaving everything to the AI when a rule would do | Every predictable line costs a review that a rule would have cleared instantly | If you can state the condition in one sentence, write the rule; save the AI for judgment calls |
| Granting Auto Post to a day-old rule | A pattern you haven't watched books straight to the ledger | Run new rules reviewed for a cycle or two; graduate them once their hits are clean |
| Never auditing rule hits | A stale pattern keeps firing after the bank changes its wording — silently mis-matching | Skim Times Applied and last-applied monthly; retire or refine what looks off |
| Judging the AI on week one | With no history in context, early suggestions are generic — and the tool gets dismissed just before it gets good | Give it a month of consistent matching; the learning context is built from exactly that |
Where to go next
- Bank Account Rules — the rule form in full.
- AI Matching Algorithm — what the AI learns from you.
- Completing a Reconciliation — where the payoff shows.
Frequently asked questions
The AI weighs your recent, consistent behavior — matching correctly going forward outweighs old inconsistency quickly. For rules, edit or deactivate the offender directly; automation changes the moment you do.
If you can write the condition, write the rule — deterministic, free, instant. The AI's lane is what you can't condition: garbled references, variable descriptions, judgment pairs. The healthiest orgs have boring rules and interesting AI suggestions.
No. The learning context is assembled automatically from work you were doing anyway — manual matches on the banking side, posted invoices on the payables side. Doing that work consistently is the entire training program.
No — the context shown to the AI is your org's own history, gathered fresh each run. Your suggestions reflect your team's habits and no one else's.
As soon as there's history to study. On the banking side, a few weeks of consistent matching gives the AI real worked examples; on the invoice side, each vendor improves from their second posted bill onward. Meridian's ninety-day arc above is a realistic pace.
Was this article helpful?