For decades, paper checks have been treated as a declining risk. They are slow, manual, and protected by bank controls designed to catch anomalies. For many modern companies, checks are an edge case: used infrequently and monitored through default safeguards like Positive Pay.

That assumption is quickly becoming outdated.

AI is enabling a new class of banking fraud that works across both modern and legacy payment channels, including paper checks.

We recently encountered a fraud attempt at Secureframe that illustrates how this shift is playing out in practice. Two fabricated paper checks were successfully deposited, even though no physical checkbook was lost, stolen, or accessed.

What made the incident notable was not its sophistication, but how ordinary it looked. That subtlety is a defining characteristic of AI-enabled banking fraud.

No access, no checkbook: AI is changing how modern banking fraud works

In this case, the fraudsters never touched our check stock. There was no intercepted mail, no missing inventory, and no insider access to check issuance.

Instead, the checks were generated.

They included:

• Accurate formatting and layout
• Plausible payees and memo descriptions
• Dollar amounts calibrated to avoid scrutiny
• Valid routing and account numbers, which most businesses share legitimately for ACH and wire payments

None of this required access to internal systems or physical assets. What it required was the ability to produce something convincing enough to pass as routine.

That is exactly what AI gives malicious actors.

AI did not invent check fraud. It removed the friction.

Check fraud has existed for as long as checks themselves. What has changed is the economics of execution.

AI lowers the barrier in three critical ways.

1. Realism is no longer expensive

Tasks that once required design skills, specialized software, or insider knowledge, such as document layout, typography, and handwriting replication, can now be assisted or automated.

2. Plausibility can be optimized

AI excels at generating average outputs:

• names that do not stand out
• amounts that feel routine
• memos that sound legitimate. 

That is exactly what modern fraud requires.

3. Iteration is cheap

Variants can be produced quickly, refined based on outcomes, and scaled without significant marginal cost.

The result is fraud that does not rely on urgency or pressure. It relies on believability. 

Businesses are already seeing the impact. About two-thirds of businesses encountered check fraud in 2024, making it the most common fraud method that year, according to the 2025 AFP Payments Fraud and Control Survey Report.

Why traditional bank controls struggle with AI-enabled fraud

Banks provide important fraud protections, including tools like Positive Pay, which flags checks that do not match issued records. These controls still matter, and they did surface exceptions in this case.

But they also reflect an implicit assumption: when an exception occurs, a human reviewer can reliably determine legitimacy by inspection.

Once an exception is approved, the item is treated as customer-authorized.

AI-assisted fraud is specifically designed to exploit that boundary by producing artifacts that look routine enough to clear human review, especially in environments where exceptions are frequent and time-boxed.

This is not a failure of any one bank. It is a mismatch between legacy verification models and modern content-generation capabilities.

The most effective AI-enabled fraud now looks normal

One of the clearest lessons from this incident was what wasn’t present.

• There were no misspellings.
• No implausible amounts.
• No unusual formatting.
• No urgency or social engineering.

The checks looked like something that could reasonably exist.

AI does not need to create perfect forgeries. It only needs to create artifacts that are good enough to blend into operational noise to impact businesses today.

As AI improves, the line between suspicious and routine continues to blur, especially in systems built on visual inspection and pattern matching.