What problem does it solve?
Banks charge through many systems: core banking, card platforms, loan servicing, trade finance, FX and payments engines, each with its own pricing tables, waivers and exceptions. Over time the configured prices drift from what contracts, product terms and negotiated deals say. Some customers are overcharged, which is a conduct risk that can end in remediation programmes and fines. Others are undercharged, which can leak revenue across many transactions.
Some of these errors come to light only when a customer complains, an audit samples the right accounts or a regulator investigates, sometimes years after the error began. Recomputing every charge independently was too expensive to do by hand. Cheap compute, contract reading with language models and anomaly detection on fee lines can make continuous checking practical.
- The CFPB ordered Wells Fargo in 2022 to pay more than USD 2 billion in redress across more than 16 million consumer accounts, for harms that included fees and interest improperly charged on auto and mortgage loans and incorrect charges on deposit accounts.CFPB Orders Wells Fargo to Pay $3.7 Billion for Widespread Mismanagement of Auto Loans, Mortgages, and Deposit Accounts (2022)
How does it work?
- Build the price book. Contract terms, product disclosure documents, negotiated pricing and waivers are read and turned into a structured, versioned price book. Language models help extract terms from contracts; people approve every entry.
- Recompute independently. For each account and period, the engine recomputes what should have been charged (fees, interest accruals, FX margins, spreads) from the price book and the transaction data, outside the billing systems.
- Compare and detect. It compares expected with actual charges line by line and runs anomaly detection on fee lines to catch patterns the rules miss, such as a waiver that never expired.
- Explain. For each discrepancy it produces an explanation (which term, which system, since when, how many accounts) so the product owner can decide quickly.
- Correct and remediate. Confirmed errors go to the owners of the billing configuration for a fix, and to a remediation process that refunds customers or recovers undercharges, with human approval.
- Audience
- Back office
- Autonomy
- Copilot
- Adoption
- Emerging
- 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: Risk and loss reduction, Compliance quality, Revenue growth, Customer experience.
Indicative value
A bank with USD 500 million in annual fee and commission income
USD 300,000 to USD 3 million
Fee income recovered from undercharging per year
How this is calculated
Formula: feeIncome * leakageRate * recoveryShare. The low scenario uses every low input, the high scenario every high input.
| Input | Low | High | Basis |
|---|---|---|---|
| Annual fee and commission income feeIncome, USD per year | 500,000,000 | 500,000,000 | The reference bank. Replace with your own fee and commission income. |
| Share of fee income lost to undercharging leakageRate, fraction of fee income | 0.002 | 0.01 | Editorial assumption, replace with the results of a sample recomputation on your own accounts. No verified public benchmark was found. |
| Share of leakage found and fixed going forward recoveryShare, fraction of leakage | 0.3 | 0.6 | Editorial assumption; some leakage is found but deliberately left in place, for example commercial waivers. |
What it leaves out: Undercharging only. It leaves out the benefit of finding overcharges early (smaller remediation programmes, fewer penalties), and the cost of building the price book, the platform and the remediation process.
Who already uses it?
1 public deployment, strongest evidence first. Grades: A regulator or audit, B the organization itself, C vendor case study, D anonymous or estimate.
State Bank of India
India · Banking · 2020
State Bank of India's Analytics Department builds machine learning models in house, and the bank's 2019-20 annual report lists models to identify income leakage among them, next to fraud, early warning and lead models. An article by an SBI chief manager in the journal of the Indian Institute of Banking and Finance (October to December 2021), citing the bank's Analytics Department, reports processing fees and facility fees recovered through this work in fiscal years 2018-19 and 2019-20. Neither source explains how the models work, how many accounts they cover or whether customers who were overcharged were also identified.
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
- Contracts, product terms and negotiated pricing, with their effective dates
- Transaction, balance and charge data per account from each billing system
- Waiver and exception records with approvals and expiry dates
- Past remediation cases as labelled examples
Systems to integrate
- Core banking, card, loan servicing and payments systems (read only)
- Contract and document management
- Pricing and deal management tools
- Remediation and complaints case management
- General ledger for recovered income
Complexity: High
The recompute logic must reproduce interest and fee conventions exactly (day count, rounding, tiering, value dates), and the contract terms are scattered across documents and systems. The AI helps with reading and detecting; the hard part is a trusted, versioned price book.
- 1
Start with one product and a sample
Pick a product with complex pricing and many accounts, such as business accounts or trade finance, and recompute a sample by hand and by machine to prove the logic.
- 2
Build the price book with owners
Extract terms with AI assistance, but have product owners approve every entry and its effective dates. The price book becomes a control in its own right.
- 3
Run continuously, report monthly
Recompute every account each cycle and report discrepancies by root cause, not only by account, so configuration errors are fixed once.
- 4
Connect to remediation
Agree with compliance how confirmed overcharges become remediation cases, and how customers are contacted and refunded.
- 5
Extend to more systems
Add products and systems one by one, reusing the price book structure and the recompute engine.
Guardrails
- The recompute logic is deterministic, versioned and documented; the model never calculates charges
- Every price book entry has an owner, a source document and an effective date
- Corrections to customer accounts need human approval and are logged
- Overcharges are always escalated to remediation, never netted against undercharges
KPIs to instrument
- Discrepancies found per product and root cause
- Value of overcharges refunded and undercharges recovered
- Time from error start to detection
- Share of discrepancies confirmed as real on review
- Remediation cases opened from the control versus from complaints
Human in the loop
Product owners approve the price book and decide on each class of discrepancy. Remediation teams approve refunds and customer contact, and finance approves recovery of undercharged income. Internal audit reviews the recompute logic periodically.
Common failure modes
- False alarms from convention mismatches
- The recompute uses a different day count or rounding than the billing system, so every account looks wrong. Validate conventions per product before scaling.
- Findings without owners
- Discrepancies pile up because no one owns the fix. Assign each product and system an accountable owner before switching on.
- Quietly keeping overcharges
- Commercial pressure favours recovering undercharges over refunding overcharges. Make overcharge remediation a mandatory, audited path.
What are the risks and rules?
EU AI Act
Minimal risk
Verifying charges against contracts is not listed in Annex III and is not a practice prohibited by Article 5. The system is internal, so the Article 50(1) duty to tell people they are dealing with AI does not arise; the Article 50(2) duty to mark generated text, such as the discrepancy explanations, falls on the provider of the generative model or system. It supports, but does not take, decisions about individual customers; remediation decisions stay with people.
Guidance
- RG 277 Consumer remediation (Australian Securities and Investments Commission, Asia Pacific). ASIC guidance, issued in 2022, for financial services and credit licensees on running consumer remediation, including identifying affected customers and returning money, for example after wrong fees or charges.
- FCA Consumer Duty (Financial Conduct Authority, Europe). The FCA's overview of the Duty, including the price and value outcome and its fair value assessments; charges above the agreed terms work against fair value.
Controls to put in place
- Price book under change control with owners and effective dates
- Versioned recompute logic with documented conventions and independent validation
- Log of every discrepancy, decision, correction and refund
- The control itself inventoried and monitored, with coverage reported by product
Frequently asked questions
- Why not rely on the billing systems to charge correctly?
- Because configuration can drift from contracts over years, across many systems, and errors can be found only through complaints, audits or regulators. Enforcement cases such as the CFPB's 2022 order against Wells Fargo, which covered fees and interest improperly charged on loans, show how large the consequences can become.
- Does the AI calculate the correct charges?
- No. The recompute logic is deterministic and versioned. AI helps read contracts into the price book, detect unusual fee patterns and explain discrepancies, but the numbers come from rules that product owners approve.
- Are there public examples of banks doing this with AI?
- Few, and they disclose little. State Bank of India's 2019-20 annual report lists models to identify income leakage among the machine learning models its Analytics Department built in house, and an article by an SBI chief manager reports processing and facility fees recovered in two fiscal years. Revenue assurance and pricing vendors describe similar work, but the case studies we checked either do not name the bank or do not say the detection uses AI.
How to cite this page
Blits.ai AI Use Case Library, "AI for fee and interest leakage detection", last verified 28 September 2026, https://www.blits.ai/ai-use-cases/fee-and-interest-leakage-detection. Licensed under CC BY 4.0. Method: how we verify use cases.
Changelog
- 28 September 2026: First published