Why Washington Banks Still Lose Money Despite AI Fraud Detection
Washington banks lose money when alerts arrive too late to act. Trace each loss in five questions and plan for the state’s five-business-day exploitation hold.
In this guide
Banks in Seattle and across Washington can still lose money with AI fraud detection because a score is only one step between a suspicious event and an effective intervention. The system may miss the relevant signal, detect it too late, send it to an overloaded queue or trigger a response that does not stop the payment. A good model metric does not prove the entire process works.
Washington adds a deadline to one of the most useful responses. When a bank suspects a vulnerable adult is being financially exploited, state law lets it refuse the disbursement, but that refusal expires on a statutory clock. The rule is state law, so a fraud lead at a Seattle community bank needs alerts that reach someone who can decide, notify and report within days, not a queue that is worked when time allows.
The useful question is where a particular loss escaped the bank’s defenses. Broad consumer fraud totals cannot answer that question: they do not isolate a bank’s losses, exposure, model coverage or losses prevented. Nor does a rise in reported fraud prove that a detection system failed.
Explore an illustrative finance exception and review workflow.
The scale is still growing. The FTC told Congress in March 2026 that consumers filed 3 million fraud reports in 2025 and reported $15.9 billion in losses, up from 2.6 million reports and over $12 billion in 2024. Detection pays when it is tied to an action: the U.S. Treasury reported that enhanced fraud detection, including machine learning AI, helped it prevent and recover more than $4 billion in fiscal year 2024, up from $652.7 million the year before, including $1 billion recovered by identifying Treasury check fraud faster. Neither figure says where any one bank's controls failed.
IBM’s fraud overview describes risk scoring alongside limitations from false positives and data quality. It provides a starting vocabulary, not proof that every AI system outperforms rules or that a model is useless against an entire payment channel.
Trace the loss through five questions
This is a proposed operational review method, not an assessment of any bank’s actual losses.
- Was the transaction in scope? Identify the payment rail, product, customer segment and attack type the control was designed to cover. “Fraud detection installed” is not a coverage map.
- Did the needed evidence arrive? Check the data available at the scoring moment. A later dispute or image cannot retroactively improve an earlier decision.
- What did the system decide? Retain score, rule hits, threshold and version. Distinguish no alert from an alert suppressed by another rule.
- Could anyone act in time? Compare alert arrival with the operational opportunity to hold, verify, reject or investigate the transaction under the bank’s procedures.
- What happened afterward? Record the action, customer contact, confirmed outcome and any recovery. Feed corrected outcomes into review without treating every alert as proven fraud.
Assign these questions jointly to fraud operations, the payment owner and the system owner. A model team can address missing features or a threshold. It cannot, by changing the model alone, fix an unattended queue or a missing payment-control integration.
Match the control to the evidence
For a card transaction, the available evidence might include merchant, device and transaction history. For a check, reviewers may need the image, issue record, payee details and deposit context. For a customer-authorized scam, authentication can confirm the person while leaving the purpose of the payment unresolved.
Those differences call for a coverage test, not a blanket verdict about what AI can or cannot detect. Ask the provider to show which signals its product uses for your actual loss types and when they become available. Test suspicious and legitimate examples from the same channel. Keep existing operational checks in the comparison; a new model may complement them.
For each intervention, name the team authorized to take it and the customer path if it is wrong. A blocked legitimate payment needs timely recovery too. Assess false alarms by their consequences, including delayed payroll or a customer unable to access funds, rather than treating them as interchangeable inconvenience.
When a Washington alert involves a vulnerable adult
Under RCW 74.34.215, a financial institution that reasonably believes financial exploitation of a vulnerable adult may have occurred, been attempted or is being attempted may, but is not required to, refuse a transaction that disburses funds from the adult’s account, an account on which the adult is a beneficiary, or the account of the person suspected of the exploitation. Here a financial institution includes any bank or credit union authorized to do business and accept deposits in Washington under state or federal law (RCW 30A.22.040). A vulnerable adult includes a person 60 or older who has the functional, mental or physical inability to care for himself or herself (RCW 74.34.020).
A refusal comes with duties and an end date. The institution must make a reasonable effort to notify everyone authorized to transact on the account and report the incident to adult protective services at the Department of Social and Health Services and to local law enforcement. Unless a court extends it, the refusal expires after five business days, or ten if the transaction involves the sale of a security, or sooner once the institution is satisfied the disbursement will not result in exploitation. An institution and its employees that decide in good faith are immune from criminal, civil and administrative liability whether they refuse or disburse.
That changes the fourth question in the loss trace. An alert that suggests exploitation of a vulnerable adult should go to a person authorized to form the reasonable belief, and the case record should carry the notice, both reports and the expiry date, because the hold ends whether or not the investigation has. Training matters as well: RCW 74.34.220 requires training on the indicators of exploitation, how to report it and how to prevent it for employees who have contact with customers and access to account information as part of their job. Have counsel confirm how the statute applies to the bank’s accounts and procedures before a workflow relies on it.
A high catch rate can still overwhelm the queue
The following numbers are illustrative assumptions, not a vendor result or a bank benchmark.
Suppose a reviewed evaluation set has 100,000 transactions, of which 100 are confirmed fraudulent. A detector catches 90 of those and flags 1% of the 99,900 legitimate transactions.
- True alerts: 90.
- False alerts: 999.
- Total alerts: 1,089.
- Recall: 90 ÷ 100 = 90% of confirmed fraud detected.
- Precision: 90 ÷ 1,089 ≈ 8.3% of alerts are confirmed fraud.
If every alert requires six minutes of review, the queue represents 108.9 hours of work. That arithmetic does not mean every flag should be handled identically; it shows why a detection percentage needs an operational denominator. If the available review capacity or payment window is smaller, the design needs a different triage or intervention plan.
The example also does not establish dollars prevented. The 90 detected cases may have different values, some may be flagged after funds move, and the 10 missed cases may contain the largest losses. Review count-based detection, dollar exposure and timing separately.
Test the response as carefully as the model
Run a historical replay to inspect coverage, then a controlled operational test that records what the system would have done at the time. Do not give it future information from a later investigation. Keep uncertain outcome labels separate from confirmed legitimate and fraudulent cases.
Track missed cases, false alerts, analyst handling time, queue age, customer friction and outcomes by payment channel. Test busy periods and outages. Ask what happens if a scoring service times out: whether the transaction continues, waits or follows another control must be an explicit bank decision.
Review changes against the same definitions. A lower alert count could reflect better precision, lost data or a raised threshold. A higher prevention estimate could reflect a changed calculation. Keep gross exposure, realized loss and recovered funds distinct in reports.
The first improvement may be a model change, a data fix or an operational handoff. Select it from the failed cases and test the full response path. That same discipline applies to finance operations: automation earns its place when the team can connect an output to an accountable, useful action.
Quick answers
Why do banks still lose money with AI fraud detection?
Because a score is only one step between a suspicious event and an effective intervention. The model may miss the signal, detect it too late, send it to an overloaded queue or trigger a response that does not stop the payment.
How much did consumers lose to fraud in 2025?
Consumers reported $15.9 billion in fraud losses to the FTC in 2025 across 3 million fraud reports, up from over $12 billion and 2.6 million reports in 2024.
Can a Washington bank hold a payment it suspects is elder financial exploitation?
It may, but is not required to, refuse the disbursement when it reasonably believes a vulnerable adult is being financially exploited. It must try to notify the account parties and report to adult protective services and local law enforcement, and the refusal expires after five business days unless a court extends it, or ten if a securities sale is involved.
Does machine learning reduce fraud losses?
It can when detection connects to an action. The U.S. Treasury reported that enhanced detection including machine learning helped prevent and recover more than $4 billion in fiscal year 2024, including $1 billion recovered from check fraud identified faster.
What is the difference between recall and precision in fraud detection?
Recall is the share of confirmed fraud the system catches; precision is the share of alerts that are confirmed fraud. In the illustrative example above, recall is 90% while precision is about 8.3%, which is why alert volume and review capacity matter.
Sources
- The FTC told Congress in March 2026 · ftc.gov
- the U.S. Treasury reported · home.treasury.gov
- IBM’s fraud overview · ibm.com
- RCW 74.34.215 · app.leg.wa.gov
- RCW 30A.22.040 · app.leg.wa.gov
- RCW 74.34.020 · app.leg.wa.gov
- RCW 74.34.220 · app.leg.wa.gov
Revision note · October 5, 2026: Refocused on Seattle; removed references to other Washington cities.
How we research and review these guides
Put the ideas to work in your business.
We provide AI consulting and implementation for finance operations teams. Bring us the process that is slowing your team down and the systems involved. We can assess the problem with you and discuss a practical implementation.
Discuss your project →Still exploring? Explore the Workflow Opportunity Workbook →