Insurance Claim Photos and Invoices: Build an Evidence Matrix
Turn a vague concern about a claim into specific, traceable questions: which file supports the described damage, which repair the invoice covers and what evidence is still missing.
Give media findings and claim consistency their own columns
A detector examines a photo or short video for AI-generation and manipulation signals. A comparison examines readable or visible details across the files and description supplied. Neither is a decision about fraud, insurance coverage, repair value or eligibility. A genuine image can be unrelated to the reported event; a matching invoice does not certify a photo’s origin.
The evidence matrix is our practical review method. It is a worksheet, not an insurer-approved standard or a claim adjudication system. Use it to define questions, trace answers to files and request missing evidence. The fictional example below has no customer material, model scores or live comparison result.
| Question | Record | Who resolves the next step? |
|---|---|---|
| Does the submitted media contain AI-related signals? | Recorded detector findings and unavailable checks | A reviewer examines signals and source context |
| Do photos, invoice and description refer to the same work? | Cited matches, discrepancies and insufficient evidence | A reviewer asks for the specific missing evidence |
| Should the claim be accepted, valued or paid? | Company decision with its evidence and policy basis | The authorised claims team under its own rules |
Prepare the source bundle before submitting a paid comparison
Keep the exact originals in your authorised evidence system and use clear, relevant copies for processing. Write a short claim description and questions the supplied files can answer. Include a close view of the reported damage, a wider view that locates it and the relevant readable invoice pages when those materials exist. Do not turn an unreadable label, cropped date or missing view into an asserted fact.
Cross-check currently accepts up to six files: JPEG, PNG, WebP, PDF, CSV or XLSX, subject to the published bounds. Each file may be up to 5 MiB and the combined bundle up to 10 MiB; PDFs have a fifty-page bound. Table parsing is bounded and formulas must be exported as values. The technical preflight checks suitability and price; it cannot certify that an image shows the needed damage or that a serial number is readable.
The JSON download below is a fictional review worksheet, not a POST /v1/comparisons payload. The filenames describe imagined evidence; the download does not supply those images or an invoice. Your actual comparison requires your files, a valid company workspace or key, and explicit submission.
- Define one review question per issue rather than asking whether the entire claim is fraudulent.
- Keep source filenames and page references in the worksheet; use the actual report citations after processing.
- Mark not visible, not readable or not supplied explicitly.
- Use required supporting files rather than adding unrelated personal documents.
Worked example: the invoice supports a different repair item
Fictional case CLAIM-DEMO-001: a description reports water damage to kitchen laminate flooring on 30 September 2026. The imagined photo is a close view of a wet floor without a room-identifying view. The imagined invoice lists ceiling painting and carries an earlier date. No detector or comparison was run on this example; the matrix is an editorial illustration of questions to ask.
A repair-item difference is concrete; it does not establish why the difference exists. The invoice could concern other work, a quotation could have preceded the event, or the description could be incomplete. Ask for the relevant itemisation and timing instead of labelling the claim fraudulent. The photograph’s room remains unresolved even if a floor is visible.
| Question | Supplied observation | Open point / next request |
|---|---|---|
| Does the photo locate the reported damage in the kitchen? | photo-a.jpg: wet floor close-up; no room-identifying view | Location not established; request a wider relevant view |
| Does the invoice cover flooring replacement? | invoice.pdf, page 1: ceiling painting | Different repair item; request the flooring itemisation |
| Does the invoice date explain the reported timeline? | invoice.pdf, page 1: 29 September; description: 30 September | Clarify quotation / work / loss dates; a date difference alone does not establish fraud |
| What did a detector find in the photo? | No live media check was run for this fictional example | No score or authenticity finding is available |
| What is the claim decision? | Not assessed in this worksheet | Authorised claims team decides after reviewing the required evidence |
Read the recorded comparison before adding a reviewer decision
For an actual completed comparison, inspect each finding’s cited source and page. The available result can contain a match, a discrepancy or insufficient evidence. A readable invoice line may support a specific repair item, while a photo’s visible features support a different question. Keep missing and omitted material explicit.
Export JSON, text or PDF while the result is retained. Exporting another format reproduces the recorded findings and does not run a new analysis. Company comparison originals are removed after processing; recorded results with extracted values and short excerpts expire seven days after submission. The worksheet is not a backup or a substitute for your own evidence record.
Append your reviewer decision and its reasoning without changing the recorded findings. If evidence is missing, request it and submit a new comparison with all required files. The old result should remain distinguishable from the follow-up. Completed insufficient-evidence comparisons are billable; failed processing is not.
Use the smallest workflow that answers the question
For AI-related findings on selected claim photos, use a company media batch. For details across a photo, invoice and description, use Cross-check. For repeated incoming submissions, connect the appropriate API workflow. A batch of image checks does not automatically become a photo-document comparison.
The blank CSV provides a place to record your own question, observations, report citations, missing evidence and decision. It includes no predetermined findings. Do not treat the fictional JSON or example CSV as an automated report. Start with a selected case and review whether the output produces useful, specific requests before expanding the process.
Frequently asked questions
Does a photo-invoice difference establish fraud?
No. It identifies a question about the supplied evidence. Request clarification and apply your own claims procedures before making a decision.
Are the worked-example observations actual DeepfakePolicy findings?
No. The example is fictional editorial material. No live detector or comparison was run, and no measured score or customer outcome is claimed.
Can technical preflight confirm that the necessary evidence is visible?
No. It checks technical suitability and price. Your team must still select relevant, readable evidence and review any missing details.
Automated results require source and context review.
Continue with independent verification.
Explore the photo and document comparison workflowPrimary reading
We use original standards, regulators, public institutions and research papers wherever possible. Sources were last checked on 7 October 2026.