Evidence Integrity: a new fraud skill for Simplifai's AI Agent for Claims

What is Evidence Integrity?

Evidence Integrity is a verification skill built into Simplifai's AI Agent for Claims.

The agent could already read a document, understand it, extract from it and act on it. It could not ask whether the document was genuine. Evidence Integrity closes that gap. The agent now examines each file as an artefact in its own right and returns a verdict the workflow routes on.

Every AI Agent inherits the skill, the same way agents inherit Document Intelligence and Human-in-the-Loop review. It is part of the agent, not a separate application a fraud team logs into.

How does the AI Agent verify a claim file?

Four independent checks run on every file. Their results combine into a single verdict.

Metadata analysis reads camera make and model, software history, timestamps, GPS, embedded thumbnails, and the file's true type as detected from its content rather than its extension. This surfaces traces left by editing and generation tools, edit history recorded inside the file, and timestamps that contradict each other.

Content provenance reads and validates embedded C2PA Content Credentials, including manifests that declare the content was AI-generated.

Image forensics examines the pixels. It finds regions re-saved at a different compression level from the rest of the image, regions carrying the wrong camera sensor noise signature, and regions where the camera's colour filter pattern has been broken.

An AI-generation classifier runs the whole image through a trained model, catching images that were never photographs.

Each check has a blind spot. Their blind spots do not overlap. That architecture is deliberate, because most manipulation is not a fabricated image. It is a real photograph with one part changed, and a detector that scores the whole image averages that signal away.

What does the agent return?

An integrity level and a routing recommendation, plus an integrity score for analytics.

With them come every anomaly that fired, each with a stable code, a severity, a plain-language reason and separate customer-facing wording, and a one-line summary a reviewer can read in seconds.

For the forensic checks, the agent reports which region of the image the anomaly sits in, with optional visual heatmaps.

Every result carries a full audit line, including the exact model version that produced the score. A decision the agent took six months ago can be traced back to the model that took it.

Which file types does it support?

Images in JPEG, PNG, WebP, HEIC and TIFF receive all four checks.

PDFs receive metadata and provenance checks, including producer and creator chain analysis, modification-against-creation timestamps, and embedded JavaScript detection.

What happens when a file is flagged?

The work branches automatically. A clean file continues. A flagged file diverges down whatever path the carrier's process defines.

A person is alerted in the queue they already work in. An intake email carrying a flagged attachment, for example, is moved into a separate folder and marked.

Processing continues either way. A checked document still flows into Classify & Extract, and that path already includes Human-in-the-Loop review.

Where does the analysis run?

Entirely inside the platform. There is no external service call and no third-party API. Classifier weights load from local storage rather than being fetched at runtime. The file is opened read-only, nothing is stored, nothing is modified, and the verdict is the only output.

Claim evidence never leaves the carrier's tenant in order to be checked.

There is also no per-document inference cost, so the agent checks every file rather than sampling a portion of them.

What Evidence Integrity does not do

It does not determine fraud. It flags manipulation, anomalies and provenance problems in individual files, and routes them for attention. Fraud is a determination made by people and a legal process, on far more evidence than one document.

It is also not the whole of fraud. Claim-level assessment and external data corroboration are different questions, and they sit on the roadmap rather than in this release.

How does verification support straight-through processing?

A provenance manifest that validates is treated as a positive signal, and the agent can route on it to move trusted evidence straight through.

So the skill accelerates as well as flags. Clean evidence clears at machine speed, and scrutiny concentrates where it belongs.

It is also a precondition for agentic claims handling rather than a feature beside it. An agent that acts on documents without knowing whether they are real is a liability rather than an asset. Automation without verification simply pays fraud faster.

Frequently asked questions

What is Evidence Integrity?

‍Evidence Integrity is an AI Skill in Simplifai's AI Agent for Claims. It checks whether a document or image attached to a claim shows signs of AI generation, digital editing or provenance problems, and returns a routing verdict the workflow acts on.

Does Evidence Integrity decide whether a claim is fraudulent?

‍No. It produces a verdict on an individual file and routes it for attention. Fraud is a determination made by people and a legal process, on far more evidence than one document.

Why four checks instead of one AI detector?

‍Because most manipulation is a real photograph with one part changed, not a fully fabricated image. A detector that scores the whole image averages that signal away. Four checks with non-overlapping blind spots cover each other.

Does claim evidence leave our environment to be checked?

‍No. The analysis runs entirely inside the platform. There is no external service call and no third-party API. The file is opened read-only, nothing is stored and nothing is modified.

What are C2PA Content Credentials?

‍A provenance manifest embedded in a file, recording how it was created and edited. Evidence Integrity reads and validates them, including manifests that declare AI generation. A manifest that validates acts as a positive signal and can be used to fast-track a file.

Is there a cost per document checked?

‍No. There is no per-document inference cost, so a carrier can check every file rather than sampling.

Where in the claims process does it run?

‍The natural placement is the moment a file arrives by email, portal or folder, before it enters the case file. The agent can apply the skill anywhere a file needs checking.

‍