What does an AI-generated bank statement actually look like?

Right now, like something made in Microsoft Excel. Perfect table alignment, transaction amounts rounded to the nearest dollar, payees labeled "groceries" instead of "Walmart." Real bank statements are messier, and that messiness is part of the signal.

I sat down with Jessica Lara, Inscribe's Risk Operations Analyst, to answer that question and the harder ones that follow it. I'm Ronan Burke, co-founder and CEO of Inscribe. Jessica reviews flagged documents every day: bank statements, pay stubs, invoices, tax forms. She has watched the AI-generated signal grow from near-zero to the highest volume Inscribe has ever recorded, and she has watched the documents themselves change along the way.

This is a conversation about what that change looks like in practice, why it matters, and why the gap between AI-generated and AI-edited document fraud may be the most important measurement problem the industry does not yet know how to solve.

What an AI-generated bank statement looks like

A year ago, the tells were obvious. Jessica describes the early documents as looking like they were made in Microsoft Excel: perfect table alignment, transaction amounts rounded to the nearest dollar, payees labeled "groceries" or "clothing store" instead of actual merchant names. Real bank statements are messier. Real transactions say "Shell" and have amounts like $47.13.

That recognition was fast. Within seconds, an experienced reviewer could call it. The formatting was wrong in ways that were immediately visible to anyone who had spent time with real documents.

Those tells are mostly gone now.

"Last year, AI documents looked almost painfully fake," Jessica said. "Now, you almost have to be a seasoned fraud investigator, and also have some sort of tool, to know it's fake."

Bank statements are the most frequently targeted document type in Inscribe's network, appearing in roughly one in four AI-generated flags. The concentration makes sense: a convincing fake bank statement is the key that unlocks high-value credit products, business financing, and mortgage approvals. "If you have a good fake bank statement with inflated income, you could really be off to the races in terms of defrauding companies and accessing high-value credit lines," Jessica said.

The harder problem: AI-edited documents

There are two categories of AI document fraud. Inscribe's AI_GENERATED detector catches one.

The first is documents generated entirely by AI: built from scratch using a prompt, a template service, or a purpose-built fraud tool. These make up the bulk of what the detector flags. The patterns, while increasingly subtle, remain detectable.

The second is real documents with AI-altered fields. A genuine bank statement with the balance changed. An actual pay stub with income inflated. The document structure is correct because it started from a real document. The alterations are targeted and precise.

"The more troubling documents I tend to come across are not the ones that have been generated using AI from scratch, but the legitimate documents that fraudsters use AI to alter," Jessica said. "Everything for the most part looks perfect and as expected, which makes them difficult to catch if you're just relying on your eye."

This category is not yet measurable in aggregate. Closing that measurement gap is a priority for Inscribe's next annual report.

Signal fraudsters try to erase

One of the most striking moments in the conversation is a specific detection story. A bank statement submitted through Inscribe's network had been deliberately stripped of its digital watermarks, a step that required knowing the watermarks existed and how to remove them. Standard visual tells were gone.

What remained was a three-digit substring that appeared repeatedly across transaction descriptions, an artifact of how the AI generation process worked. A human reviewer scanning the document might miss it. Pattern detection across Inscribe's network of genuine documents caught it.

The broader implication is one I keep coming back to: Inscribe's detection advantage comes from a large network of real documents to measure against. Fraud stands out not because individual signals are always obvious in isolation, but because fraudulent documents behave differently from genuine ones when viewed across the full corpus. The more context the system has, the harder individual fraud becomes to hide.

The arms race is changing who commits fraud

I asked the question that comes up in every version of this conversation: can AI catch AI long-term, or does it just move the competition somewhere else?

The short answer from the data is that the arms race is accelerating. Purpose-built tools like FraudGPT operate without guardrails. The open-source models that underpin commercial AI improvements are improving fraud tooling on the same curve. The barrier to generating a convincing fraudulent document has not just lowered. It has reached near zero.

"The barrier to entry for committing fraud is at an all-time low right now," Jessica said. "Something that used to take time and technical skill now can be done in minutes by somebody who has no skill."

The consequence is not only higher fraud volume. The profile of who commits fraud is changing.

"There's historically good customers that may be part of your institution that are now taking that chance to commit fraud," Jessica said. "You're going to see people who may not have engaged in fraud yesterday start doing so tomorrow."

For fraud teams, the practical implication is a timing one. Detection infrastructure needs to be in place before the queue starts showing AI flags, not built in response to them. The documents are already convincing enough that first exposure and first preparation should not be the same moment.