AI Reconciliation Agents Inside ERP: How the Pattern Works for Australian Finance Teams

The rule in our AI practice is simple: the agent drafts, the human approves, the system posts. That sequence is non-negotiable for anything that touches the general ledger or the Activity Statement.

A reconciliation agent makes that sequence faster, not optional.

Bank statement reconciliation is the use case most Australian finance teams tackle first when adding AI to their ERP. According to MYOB’s 2025 ERP Trends report, 45% of decision-makers say disconnected data is their biggest barrier to better business decisions. Reconciliation sits right at that disconnect. The agent does not remove the decision, but it removes most of the searching.

What the pipeline looks like

A reconciliation agent runs a five-stage pipeline on a schedule your finance team sets.

Ingest. The agent pulls the bank statement, either via a direct feed or a file import, and reads the outstanding items in your ERP: open invoices, vendor bills, pending journal entries, and payments in transit.

Normalise. Dates parse to a standard format. Amounts strip currency symbols. Reference numbers extract from transaction descriptions using pattern matching on common Australian bank statement formats.

Match. Deterministic rules run first. Exact amount, matching supplier, reference number present, date within the payment terms window. A large proportion of routine transactions match on these rules alone. Most mature deployments see 85 to 95 percent of high-volume items matched deterministically before the LLM processes a single line.

Reason over exceptions. The transactions that did not match cleanly go to a reasoning layer. The agent checks the supplier history, compares the amount against the purchase order within a configured tolerance, and looks at prior resolutions to similar exceptions. It classifies each one: timing difference, amount variance under threshold, missing invoice, or possible control risk.

Draft and route. For every exception, the agent produces a classification, a draft recommendation, and routes it to the right person. It does not post. The agent has no write access to a finalised period.

What Odoo 19 has today [shipped]

Odoo 19’s bank reconciliation module includes configurable reconciliation models that apply rules automatically as bank statements import. The models learn from your team’s choices over time: if you consistently match a particular transaction type to a specific account code, the model starts suggesting that pairing. Unmatched items hold in a review queue and do not block the rest of the import.

For the deterministic matching layer, the native module handles routine volume well. Extending it with an agent, to handle the exceptions that fixed rules miss, requires connecting via Odoo’s server actions and the Model Context Protocol. That is a configuration and development task, not a toggle. The server actions define what the agent is permitted to do; MCP controls how an external LLM reads and writes within the Odoo environment. Set the permission scope first.

For the full list of what Odoo 19 server actions expose to an agent, and what the v20 roadmap adds on top, the Odoo AI agents guide covers this in detail.

Direct reconciliation write-backs via Odoo’s Ask AI feature, where you instruct the agent in natural language to reconcile specific items, are on the v20 roadmap [roadmap]. They are not available in Odoo 19. Ask AI in the current release can read and report; it does not execute ledger write-backs.

What MYOB Acumatica has today [shipped/preview]

MYOB Acumatica has automated bank reconciliation as a live, shipped feature. The platform matches bank feeds to ERP transactions, handles multi-currency feeds from multiple banks, and uses historical transaction patterns to suggest account codes for new entries. Finance teams on a live tenant see the exception queue shrink each month as the system learns their specific chart of accounts and supplier behaviour.

The 2026 R1 release added AI Advisor, which flags anomalies and discrepancies in real time and routes them to the relevant reviewer rather than leaving them buried in a period-end report. This is available for ANZ customers on the 2026 R1 release [shipped].

AI Studio, which connects MYOB Acumatica to an external LLM via MCP, was in early-entry preview from May 2026 [preview]. Once at general availability, AI Studio will allow custom reconciliation agents to operate inside the MYOB Acumatica security boundary, with role-scoped permissions and no data leaving the system. For ANZ customers, allow some lead time between global release dates and local rollout.

How an AI reconciliation agent handles exceptions

If your agent matches 90 percent of transactions and routes the remaining 10 percent as undifferentiated exceptions, you have not solved the problem. You have moved the pile from the bank statement to a different inbox.

The value of the agent is in how it classifies that 10 percent before it reaches the reviewer.

There are three categories worth distinguishing for an Australian mid-market business.

Timing differences are the most common. Payment sent before the bank clears. Invoice dated end-of-month, payment received the next business day. The agent identifies these by comparing the expected payment date to the actual receipt date within a configurable window, typically three to five business days, and confirming the amount and supplier match.

Amount variances under a threshold cover rounding, bank fees, and currency conversion differences on international payments. The agent drafts a minor variance journal entry and routes it for a single confirmation. Most finance teams set this threshold at 1 percent of transaction value or $50, whichever is lower. Above the threshold, the item goes to the standard exception queue for manual review.

Missing invoices are the control risk category. Payment received, no matching invoice in the system. The agent flags this for AP review. It does not attempt to resolve it independently, and it should not.

A well-configured agent presents each exception with a classification, the evidence it found, and a draft action. The reviewer sees context, not a raw transaction line.

“The question I ask before scoping any reconciliation agent is: what is your current exception rate, and what happens to those exceptions today? If they pile up until the BAS deadline, the agent’s real value is in the exception workflow, not the match rate.”

Josh Craig, Director, Auboros

The Australian compliance layer

Every GST-registered business has an additional constraint the agent must respect: the GST code on every transaction determines its BAS treatment.

The agent can suggest a GST code based on the supplier and transaction type. It should not apply that code without a human confirming it first. A GST input tax credit posted on a transaction that is actually GST-free does not just create a reconciliation difference. It misrepresents your BAS and your tax liability.

The ATO requires businesses to retain records of their GST transactions for five years. An agent that posts autonomously, without a log of the human approval that preceded the posting, creates an audit trail problem that outweighs the time saved.

The practical control is a GST code allowlist. Common codes such as GST (standard rated), FRE (GST-free), and CAP (capital acquisitions) sit on the allowlist; the agent can suggest these without escalation. Less common codes, including IMP (imported services) and specific input-taxed categories, require mandatory human confirmation before the entry is drafted. When the agent is uncertain, it escalates rather than guessing.

A worked example with an approval gate

A wholesale distribution business runs Odoo 19 with a reconciliation agent configured over the API. On a Friday afternoon, a $4,680 payment arrives from a regular customer. The ERP has one open invoice for $4,680 dated 12 days ago.

The deterministic matching layer identifies the exact amount, the same debtor, and the invoice date within terms. It assigns the payment to the invoice automatically. No one reviews this; it matched cleanly.

Two hours later, a $4,950 payment arrives from the same customer. The ERP has a second open invoice for $4,680, still unpaid, and no invoice for $4,950.

The agent checks the customer’s history. This debtor has paid $270 over invoice face twice before in the same month, both times followed by a freight credit note processed separately. The agent classifies this as a probable split: $4,680 against the open invoice, $270 as an unapplied receipt pending the credit note.

The agent routes it to the AR manager with that classification and draft entries. It does not post.

On Monday morning, the AR manager opens the review queue, reads the agent’s reasoning, confirms the credit note is on its way, and clicks approve. Both entries post. The agent’s suggestion was right; the human confirmed it.

That is what the approval gate looks like in a working deployment.

Three configuration decisions before you go live

The exception threshold sets what the agent routes versus what it tries to resolve. Start conservative, around 1 percent of transaction value or $50, and tune it after the first quarter of data. Setting it too high on day one means the agent approves items your finance team should have seen.

The GST code allowlist constrains what the agent can suggest without escalation. Implement this explicitly before go-live. Failing to define the allowlist is the most common configuration oversight in reconciliation agent deployments.

The locked-period rule prevents the agent from drafting entries into a closed accounting period. Code this into the agent’s permission set explicitly. The ERP’s own period-close controls are not a substitute; the agent needs to know your close schedule.

What to do next

For businesses on Odoo 19, the Odoo bank reconciliation setup guide covers the native configuration in detail. Extending it with an agent layer starts with scoping your transaction volumes, exception rate, and chart of accounts.

For businesses on MYOB Acumatica, the existing automated reconciliation feature is already handling part of this work. AI Advisor adds the real-time anomaly layer, and AI Studio (preview) opens the path to custom agent configurations once it reaches general availability for ANZ.

For a broader picture of what agents can do across the full scope of an ERP deployment, the overview of AI agents for Australian businesses covers the range of use cases and the governance model behind them.

Our AI agents service covers both Odoo and MYOB Acumatica and includes pattern design, exception taxonomy, and approval-gate configuration as part of the engagement scope.

Book a session if you want to estimate what your current reconciliation workflow would look like with an agent layer.