What problem does it solve?
Every bank reconciles its own books against the outside world many times a day: nostro statements against expected cash flows, card and scheme settlement files against authorised transactions, clearing and suspense accounts against the general ledger. Rule based matching engines handle the clean cases, but timing differences, partial references, split and bulked payments, bank charges and FX conversions leave a steady stream of breaks that people clear by hand, often in spreadsheets.
Those breaks are where the cost and the risk sit. Items ageing in suspense accounts can distort the balance sheet, tie up capital and liquidity, hide fraud and lead to audit findings. Writing a new matching rule for every new pattern is slow, so operations teams grow with volume instead of staying flat.
How does it work?
- Ingest every feed. Statements (MT940, camt.053), settlement files, ledger extracts and remittance advices arrive in one pipeline; document AI reads the unstructured ones.
- Match beyond the rules. Deterministic rules clear exact matches first. A learned matching layer then proposes one to one, one to many and many to many matches using fuzzy references, amounts within tolerance, value dates and FX, each with a confidence score.
- Explain and propose. For each proposed match or break the agent states why (for example "bank charge of 15 EUR deducted by the correspondent") and drafts the clearing journal or the adjustment.
- Route the exceptions. Low confidence items and anything above a materiality threshold go to an operator queue with the evidence attached. Retrieval over prior resolutions suggests how similar breaks were cleared before.
- Learn under control. Operator decisions feed back as candidate rules or training data, which a reconciliation owner approves before they change production matching.
- Audience
- Back office
- Autonomy
- Supervised agent
- Adoption
- Early adopters
- Channels
- Internal tools, API and system to system
What is it worth?
Benchmarks are computed from the public deployments below: one data point per organization per KPI, with who made each claim.
No public deployment has disclosed a measurable outcome yet.
Value drivers: Lower cost to serve, Risk and loss reduction, Speed and cycle time, Employee productivity.
Indicative value
A mid sized bank that clears 300,000 reconciliation breaks by hand each year
USD 315,000 to USD 2.2 million
Manual reconciliation effort avoided per year
How this is calculated
Formula: manualItems * automatedShare * minutesPerItem / 60 * costPerHour. The low scenario uses every low input, the high scenario every high input.
| Input | Low | High | Basis |
|---|---|---|---|
| Breaks cleared manually per year manualItems, items per year | 300,000 | 300,000 | The reference bank. Replace with the exception count from your reconciliation platform. |
| Share of those breaks the AI clears or pre clears automatedShare, fraction of manual items | 0.3 | 0.6 | Editorial assumption; no verified public benchmark for AI match rates was found. Replace with a pilot result on your own data. |
| Minutes an operator spends per break minutesPerItem, minutes per item | 6 | 12 | Editorial assumption, replace with your own time study. |
| Fully loaded operations cost per hour costPerHour, USD per hour | 35 | 60 | Editorial assumption for a blended onshore and offshore operations team. |
What it leaves out: Labour only. It leaves out the value of fewer aged items in suspense (capital, liquidity and fraud exposure), fewer audit findings, and the cost of the platform and the integration work.
Who already uses it?
4 public deployments, strongest evidence first. Grades: A regulator or audit, B the organization itself, C vendor case study, D anonymous or estimate.
Ginnie Mae
United States · Government and public sector · 2021
Ginnie Mae, part of the US Department of Housing and Urban Development, analyses the transaction data of its master subservicers every month. Since April 2021 it has used machine learning models, built in house and with Ernst & Young as contractor, to detect anomalies, data inconsistencies and exceptions in that data. It says early detection reduces manual adjustments to financial reporting, which saves cost and time. The 2025 federal inventory still lists the system as deployed. No outcome figures are published.
No outcome disclosed.
Comrade Trustee Services
Papua New Guinea · Wealth and asset management · 2026
Comrade Trustee Services, trustee of the Defence Force Retirement Benefit Fund in Papua New Guinea, went live with Smartstream's Air AI reconciliation platform, which it uses to reconcile multiple file types, including fixed length files and PDFs that need advanced matching logic. The article says Air replaced manual data collection and spreadsheet preparation, and the vendor says processing time fell from up to eight hours to under five minutes.
No outcome disclosed.
National Bank of Greece (Cyprus)
Cyprus · Banking · 2026
National Bank of Greece in Cyprus consolidated four reconciliation systems into one on Smartstream's AI enabled Air platform (the Air Cash module), replacing both incumbent and standalone systems. The bank had a fragmented landscape that needed significant daily manual effort across systems and data formats. The platform matches groups of items at once and flags data quality issues in internal data and incoming bank statements. The project was completed in three months; no operational outcome figures were disclosed.
No outcome disclosed.
World Food Programme
Global · Government and public sector · 2024
DARTS (Data Assurance and Reconciliation Tool Simplified) is a web application that uses machine learning to help World Food Programme country offices apply controls to large cash transfer datasets and generate reconciliation reports, so that humanitarian cash assistance is paid out accurately and accountably. A 2024 US federal AI use case inventory entry by the USAID Bureau for Humanitarian Assistance, dated April 2024, listed it at the implementation and assessment stage and said it was funded through the WFP Innovation Accelerator. It does not appear in the 2025 inventory, so its current status is unknown. No outcome figures are published.
No outcome disclosed.
How do you implement it?
A model agnostic playbook: what to prepare, the order to build in, and what goes wrong.
Data you need
- Twelve months of history of matched items and cleared breaks, with the resolution chosen
- Clean static data for accounts, counterparties and correspondent banks
- Documented tolerances and materiality thresholds per reconciliation
Systems to integrate
- Reconciliation platform or matching engine
- General ledger and subledgers
- SWIFT or bank statement feeds (MT940, MT950, camt.053)
- Card scheme and acquirer settlement files
- Workflow or case tool for exception queues
Complexity: Medium
The matching logic is well understood; the work is data. Feeds arrive in many formats and timings, references are inconsistent across systems, and every automated journal needs a control design that auditors accept.
- 1
Start with the reconciliations that hurt most
Rank reconciliations by manual breaks per month and by aged value in suspense. Pick two or three with high volume and clear ownership, such as a card settlement or a busy nostro.
- 2
Baseline the current rules
Measure what the existing engine already matches so the AI is credited only for the increment. Many teams find quick wins by fixing static data before any model is involved.
- 3
Run in shadow mode
Let the AI propose matches and journals next to the operators for several cycles and compare. Only promote a match type to automatic once its precision on your data is proven.
- 4
Set materiality and maker checker thresholds
Agree with finance and audit which proposals may post automatically and which need a second person, by amount, account type and confidence.
- 5
Close the learning loop
Capture why operators accept or reject a proposal and review those patterns monthly with the reconciliation owner before they become new rules.
Guardrails
- No automatic posting above the materiality threshold; those journals need a maker and a checker
- Every match and journal carries the evidence and confidence it was based on
- Tolerances and thresholds are configuration owned by finance, not learned by the model
- Unmatched items are never forced to clear; low confidence always goes to a person
KPIs to instrument
- Auto match rate per reconciliation, on top of the rule engine baseline
- Precision of automated matches from a monthly sample
- Number and value of items aged over 30 days in suspense
- Operator minutes per break
- Audit findings related to reconciliations
Human in the loop
Operators own the exception queue and approve every journal above the threshold. The reconciliation owner approves new match types and tolerances, and internal audit samples automated matches each quarter.
Common failure modes
- False matches that hide a real break
- A plausible but wrong match clears an item that should have been investigated. Sample automated matches and keep tight tolerances on amount.
- Drift after an upstream change
- A new file format or reference convention quietly lowers match quality. Monitor match rate per feed and alert on sudden drops.
- Credit for work the rules already did
- Benefits are overstated because the baseline was not measured. Report the increment over the existing engine.
What are the risks and rules?
EU AI Act
Minimal risk
Matching entries between internal financial records is not a use listed in Annex III and is not a practice prohibited by Article 5. Operators knowingly use an internal AI tool, so no Article 50(1) disclosure is needed. If a generative model drafts the explanations or journals, the provider of that system may have to mark its output as AI generated under Article 50(2). The AI literacy duty of Article 4 applies to the bank as deployer.
Rules that apply
Guidance
- Principles for effective risk data aggregation and risk reporting (BCBS 239) (Basel Committee on Banking Supervision, Global). Principle 3 expects risk data to be reconciled with the bank's sources, including accounting data where appropriate, and aggregated on a largely automated basis.
- SS1/23 Model risk management principles for banks (Bank of England, Prudential Regulation Authority, Europe). Applies to UK banks, building societies and PRA designated investment firms with internal model approval for regulatory capital. A learned matching model that suggests or posts journals falls under its model definition and belongs in the model inventory with validation and monitoring.
- SR 26-2 Revised Guidance on Model Risk Management (Board of Governors of the Federal Reserve System, OCC and FDIC, North America). Issued on 17 April 2026, it supersedes and replaces SR 11-7 and asks for a risk based approach tailored to each bank's model risk profile, size and complexity. The letter says it is most relevant to banking organizations with over $30 billion in total assets regulated by the Federal Reserve. A US bank that treats its learned matching model as a model under its policy validates and monitors it on this basis.
Controls to put in place
- Model inventory entry with an owner, validation and drift monitoring per reconciliation
- Maker checker on journals above the materiality threshold
- Immutable log of every proposal, the evidence used and who approved it
- Quarterly sample of automated matches reviewed by an independent team
Frequently asked questions
- How is AI reconciliation different from a rule based matching engine?
- Rules clear exact and near exact matches and should stay. AI adds matching on partial references, amounts within tolerance, one to many and many to many combinations, and a plain language explanation of each break, so fewer items reach an operator and those that do arrive with a suggested resolution.
- Can the AI post clearing journals on its own?
- Only within limits agreed with finance and audit. A common design lets low value, high confidence matches post automatically and keeps a maker and a checker on anything above a materiality threshold, with every proposal logged with its evidence.
- Who is already using AI for reconciliation?
- Public evidence is still thin on measured results. National Bank of Greece in Cyprus consolidated four reconciliation systems onto an AI enabled platform in 2026, and Comrade Trustee Services in Papua New Guinea went live on the same vendor platform, which says processing time fell from up to eight hours to under five minutes. A 2024 US federal AI inventory entry listed a World Food Programme machine learning tool for reconciling cash transfers at the implementation and assessment stage. (Ginnie Mae uses machine learning to find anomalies and exceptions in subledger transaction data before financial reporting, which is a related data quality use, not reconciliation matching.)
How to cite this page
Blits.ai AI Use Case Library, "AI for ledger and payment reconciliation", last verified 27 September 2026, https://www.blits.ai/ai-use-cases/ledger-and-payment-reconciliation. Licensed under CC BY 4.0. Method: how we verify use cases.
Changelog
- 27 September 2026: First published