Index

はじめる

See what your documents aren’t telling you

Run fileForensics on your own invoices, claims or statements. Book a demo or get a free review of a sample of your recent files.

Book a demo

Quick answer

fileForensics is fileAI’s trust check for documents. It inspects every file at intake, including invoices, claims, bank statements and photos. It looks at structure, metadata, edit history, pixels and content logic to flag tampering, AI-generated content and duplicates. Each finding comes with evidence a reviewer can check in seconds. Teams can use it through the fileAI platform, API, MCP server or developer playground.

fileForensics is fileAI’s trust check for documents. It inspects every file at intake, including invoices, claims, bank statements and photos. It looks at structure, metadata, edit history, pixels and content logic to flag tampering, AI-generated content and duplicates. Each finding comes with evidence a reviewer can check in seconds. Teams can use it through the fileAI platform, API, MCP server or developer playground.

Making a convincing fake document used to take skill. Today it takes a free phone app or a single prompt. An amount on an invoice changes from $1,500 to $15,000, a date moves by a week, a photo gains a cracked bumper. Nothing about the result looks wrong.

Checking a document still works the old way: a person glances at it, or a rule checks one field. That gap is widening every quarter, and it is about to matter far more, because the next reader of most documents won't be a person. It will be an agent.

This month we launched fileForensics, the trust check for every file your teams, your software or your agents touch.

Extraction tells you what a document says. It doesn't tell you whether to believe it

For five years, fileAI has turned unstructured files into structured, reliable data for insurers, banks, accounting firms and government teams. We have processed more than a billion files. And in almost every deployment, once extraction went live, someone asked the same question: "But what if the file has been tampered with?"

It is a fair question. An accurate extraction of an edited invoice is still an edited invoice. Perfect OCR on a manipulated bank statement just moves the problem downstream faster. Once a forged claim or invoice has been extracted, it stops looking like a suspicious file and starts looking like clean, structured data, ready to be approved and paid.

Until now, the tools that answer that question have mostly been sales-gated, priced for large enterprises and separate from extraction. Teams ended up stitching a fraud tool onto an extraction tool onto a workflow tool, and paying for the integration in time and gaps. A startup adding an upload step, or a developer wiring a document into an agent, had two options: write brittle ExifTool scripts, or do nothing.

What fileForensics does

fileForensics checks every file at the point of intake, not a sample, before anyone acts on what it says. It looks at the file's structure, metadata, edit history and pixels, and at whether the content holds together. It covers eight check areas: file information and quality, editing signals, AI-generated content, duplicate and reused content, file structure and hidden content, security and provenance, content integrity, and a visual heatmap that shows which parts of an image drove the analysis.

In plain English, it works in four automated stages: inspect the file, examine its history, test its logic and support the decision.

Generative AI created much of this risk. fileForensics meets it with AI-native checks, rather than relying on manual spot checks to keep up.

What comes back is not a black-box score. Each finding says where, what and why: "the total field on page 1 was set in a different font from the rest of the invoice", "this photo's modified date is earlier than its created date". A reviewer can check it in seconds.

What this means for your business

The teams who feel document fraud most are rarely the ones writing code. They are claims handlers, credit and onboarding teams, AP and finance operations, and the compliance and audit functions behind them. fileForensics is designed for the way they work:

  • Insurance. Check every claim document and supporting photo at intake, so low-risk claims can move straight through and the review team spends its time on the files that warrant it.

  • Banking and financial services. Add a trust check to onboarding, lending and compliance workflows, where payslips, bank statements and supporting documents are now easy to fabricate.

  • Accounting and finance. Catch edited or duplicated invoices and receipts before they reach matching, reconciliation and payment.

  • Review teams and auditors. Every finding comes with evidence and an exportable record of what was checked, so you can show not just what a document said, but what you did to check it.

One engine, four ways in

fileForensics runs on one engine, with four ways in: the fileAI platform, for teams that want routing, review queues and audit trails; the developer portal, where you can drop a file into the playground and see findings in about five seconds; the API, with Python and TypeScript SDKs; and the MCP server, for agents and AI clients. Whichever you use, you get the same findings, so an operations team's live workflow and a developer's prototype work from the same evidence.

Built into our pipeline, open to builders

For teams on the platform, fileForensics is built directly into the fileAI extraction pipeline, so every file is checked in the same pass that reads it. If your team already uses our fileLedger or fileShield products, there is nothing new to adopt: switch the capability on, and integrity signals appear inside the screens you already work in every day, next to the data you already review.

If you are building your own product or agent, the developer portal, API and MCP server give you the same engine to work with. With fileForensics, you can:

  • Add trust checks wherever you work. One API call returns clean JSON with findings and evidence.

  • Get extraction and integrity from the same call. No second vendor, no second integration, no second bill.

  • Give your agents a check before they act. Through our MCP server, ask Claude or Cursor "can I trust this invoice?" and get the evidence back in the chat.

  • Show reviewers and auditors exactly why. Every finding points to the field, page and region it concerns.

Evidence for people, not verdicts about them

We built fileForensics as decision support. Findings route a file to review; they never reject a claim, a loan or an application on their own. A review flag is not a fraud finding, and a clean result is not a guarantee of authenticity. The decision stays with your team.

That matters, because behind every flagged file is a person, and most of the time that person is honest. A wrongly flagged claim means a delayed payout and an unhappy customer. Evidence a reviewer can check in seconds is how you clear a false positive quickly, and how you keep automation running without treating every customer as a suspect.

We asked Singapore: How good are people at spotting a fake?

From 2 to 5 November, our Keeping It Real van took fileForensics out of the browser and into the CBD: two days outside Asia Square Tower 1, then two at Ocean Financial Centre, from ten until four each day. We asked passers-by a simple question: how much fraud can you catch in a game of ‘Spot the Fake’? Ten cases, 45 seconds on the clock, and in each case you tap the image you think has been tampered with. Every answer locks the moment you tap, and everyone who finished spun the prize wheel, whatever their score. Hundreds of people put themselves to the test, comparing genuine and manipulated images and finding out what we already knew. It’s hard!

The engagement was fantastic, and the results were telling. Here is how they did:

Result

Share

Verified human

45.5%

You found some signs of fraud

11.2%

You're a victim of fraud

43.3%

Probably not human

0%

Total

100%

Nearly half of players (45.4%) earned a “Verified human” status: they ran out of time and were fooled more than once along the way. Almost as many (43.3%) finished all ten cases but ended as victims of fraud. Another 11.2% spotted some fakes but let others through, despite taking their time.

Put another way, 88.8% of players fell for at least one fake. And these were close to ideal conditions. Every player knew there was a fake in every case, was actively looking for it, and had nothing competing for their attention except lunch. A claims handler or AP team working through a queue gets none of those advantages. A tampered file usually arrives looking routine, attached to an email that says “Please see attached.”

That is not a case against human judgement. Noticing that something looks off is often how fraud gets caught. But an AI-generated image is not automatically fraud, and an edited document is not automatically fraudulent: context decides. What reviewers need is evidence about the file itself, gathered before its data is extracted and acted on, so their instincts have something concrete to work with. That is the job fileForensics does.

Thank you to everyone who played, to our team who ran the van and the demos all week, and to Asia Square and Ocean Financial Centre for hosting us.

Built in the open: our Product Hunt launch

On 4 November, we also launched fileForensics on Product Hunt. It was a great experience, and the community engagement was energising: builders testing files in the playground, sharp questions about how individual checks work, and plenty of ideas for where a trust check belongs next. Thank you to everyone who upvoted, commented and tried it. Your feedback is already shaping what we build next.

Looking ahead

Within two years, AI agents will read, approve and pay on documents with no human in the loop. A tired reviewer might notice an odd font. An agent won't, unless it is told to check. The agent reading a document and the check examining it are not the same thing. Regulators, auditors and reinsurers are already starting to ask not just what a document said, but what you did to check it.

We think "check before trust" will become a standard part of every document workflow, as ordinary as validating an email address. fileForensics is how we are making that check easy, cheap and explainable from day one, whether the next reader of a document is a developer's pipeline, a claims team or an agent.

Try it today

Developers: drop a file into the playground at https://developer.file.ai/, connect the MCP server in Claude or Cursor, or grab an API key and add fileForensics to your upload flow.

Business teams: see fileForensics on your own documents. Book a demo or request a free review of a sample of recent claims at www.file.ai/fileforensics.

今すぐはじめる

実行するほど、組織の記憶が育つ

実行するほど、組織の記憶が育つ

今すぐはじめる

実行するほど、組織の記憶が育つ

Building the foundation for agentic AI

Building the foundation for agentic AI

Building the foundation for agentic AI