top of page

AI in reconciliation is not one thing: where AI fits in banking operations

3 days ago
4 min read

For Banking and Finance Operations leaders · about 8 minutes · Sources: our August and July roundtables, our webinars, and the Operartis Adoption Guide


AI now sits on almost every banking operations agenda, and reconciliation is one of the places it is promised most often. Yet "AI reconciliation" covers several very different jobs, each with its own technology, its own risks and its own return. Treating them as one thing is the quickest way to buy the wrong tool, or to dismiss the right one.


This guide separates those jobs, shows where each one pays off, and explains where machine learning matching fits alongside the systems banks already run.


Why banks are looking now


The pressure comes from several directions at once. Transaction volumes keep rising, and on our BPO panel one senior delivery leader put it plainly: putting more people on reconciliation "is definitely not a scalable solution." Data is getting more complex too. New payment standards such as ISO 20022 carry far more fields than the formats they replace, and reconciliation tools have to make sense of them.


Regulators expect firms to do more, faster and with fewer errors. And according to one reconciliations consultant on the same panel, adoption started earlier and more aggressively in the US, with Europe now catching up.


There is a less comfortable driver as well. One webinar attendee called it "buzzword compliancy": adopting AI because competitors claim to have it. That is the strongest reason to be precise about what each tool actually does.


Five jobs that get called "AI reconciliation"


At our August roundtable in London, a reconciliation operations consultant said the confusion comes from collapsing very different use cases into a single phrase. Broken down, they look like this:


Job

What it does

Typical technique

When the benefit arrives

Getting data in

Reads statements, PDFs and emails and normalises them

OCR, document AI, language models

Each time a new source is added, then ongoing

Onboarding and rule-building

Sets up new reconciliations and writes their matching rules

Language models, code generation

Once per reconciliation

Matching

Pairs transactions the rules can't

Machine learning trained on past matches, with calibrated confidence

Every day

Exceptions and breaks

Routes, prioritises and helps resolve what stays unmatched

Workflow models, AI agents with human review

Every day

Root cause and anomalies

Finds why items fail and flags unusual patterns

Pattern recognition

Every day


Each has a different business case, and a different risk if it goes wrong.


One-off returns and ongoing returns


Tracey, founder of Operartis, drew the key distinction in our AI in Reconciliations Playbook webinar. Onboarding happens once per reconciliation, so making it faster captures that benefit once. Matching, exceptions and root-cause analysis happen every day, so their benefits keep adding up.


The consultant on that panel put it in finance terms: AI onboarding "shrinks the CapEx" of a project, but does nothing to the running cost. The unmatched items, and the people needed to clear them, are still there the morning after go-live.


The job in the middle


Matching is the job most people want solved and the one least often addressed. The same consultant explained why: many people believe that letting AI loose on a dataset means losing control of what gets matched with what.


That concern is fair, and it shapes how matching should be done. Machine learning for matching needs to be constrained, accurate and explainable:


  • trained on your own historical matches, so it learns your patterns rather than general ones

  • giving every match a confidence score that reflects how often matches like it are actually right

  • explaining each match in terms an auditor can follow

  • running alongside the rules engine, not replacing it


Do you need AI matching at all?


Not always, and the roundtable didn't pretend otherwise. One reconciliation technology practitioner described a large firm that rebuilt its upstream data processes around standard templates. Its rules-based match rates rose substantially, up to 99.9%.


Others in the room pointed out where that approach runs out. You can't rebuild data you don't control. External bank statements, broker files and counterparty data arrive in whatever form they arrive, often without reliable identifiers. A reconciliation that needs 40 or 50 rules, one participant noted, is usually a sign of exactly that kind of data problem.


Tracey's rule of thumb from the Playbook webinar is simple. If a reconciliation has unique references or plenty of high-quality fields, rules are perfectly good. Machine learning earns its place where you need tolerances, fuzzy matching or a growing list of exceptions to the rules.


The technique itself is well proven. As one AI executive at the roundtable noted, credit bureaus have long used statistical matching to resolve identities across records that don't share a key.


Where to start


When we asked our BPO webinar audience which use case mattered most, breaks handling and root-cause analysis came top, with 37% of votes. A senior BPO leader on the panel offered a useful map, sorting reconciliation AI into three groups:


  • Efficiency and scale: automated matching, then real-time reconciliation, then self-learning models

  • Value and cost: exception handling and predicting breaks before they happen

  • User experience: handling unstructured data and routing exceptions intelligently


For most teams, the practical starting point is the reconciliation where manual matching hurts most. Our Data Preparation Guide recommends choosing the one with the highest manual intervention rate. Bank-to-ledger, intercompany and multi-currency cash reconciliations are the most common starting points, and nostro, cash, card and GL reconciliations are where machine learning matching tends to deliver most.


Keeping control as you add AI


Whichever job you start with, the same controls apply:


  • Add to what works. Run AI alongside existing systems, so rollback means switching it off.

  • Keep people where judgement matters. A BPO leader at the roundtable described human review as the standard control for agent-driven exception handling.

  • Make confidence mean something. Confidence scores should be calibrated, so they can decide what goes straight through and what goes to a person.

  • Bring model risk in early. Rules-based and machine learning systems alike are expected to have drift monitoring and back-testing.


AI in reconciliation isn't one product or one decision. It is a set of distinct jobs. Choosing well starts with naming which one you are trying to do.


Read next



bottom of page