Skip to content

Developer docs

File checks: fake, edited and AI-made files

When a customer attaches an ID, a bank statement, a bill, a payslip, an invoice or a photo of a person, your agent can check the file for signs that it was edited, forged or made by AI, before anyone relies on it.

The chat checks files by itself

You do not call anything extra. The agent decides to run the check when a file matters (for example, someone proves who they are or sends a document as evidence) and tells your team when a file looks doubtful.

What it looks for

What the file check looks for, by kind of file
Kind of fileWhat it can spot
ID documents (passport, driving licence, national ID)Signs of tampering, edited or pasted text, an image copied from the internet, text that does not belong on that kind of ID
Statements, bills, payslips, invoices, receiptsRe-typed or edited figures, mixed fonts and spacing, totals and dates that do not add up, editing-software traces and edit history, a template look
Photos of peopleAI-generated faces, swapped or manipulated faces (deepfakes), skin, hair, lighting and background artefacts
Any fileLabels left inside the file by AI image tools, and screenshots or photos of a screen passed off as the original

Several independent checks run together and are combined into one answer, so a single weak signal is not enough to raise an alarm.

What the agent does with the answer

Risk levels and what the agent does
ResultMeaningWhat the agent does
LOWNothing suspicious was foundCarries on. It never promises the file is genuine.
REVIEWSomething looks off and a person should lookTells the customer politely that a team member will review the file, and hands the conversation over.
HIGHStrong signs of editing, forgery or AI generationThe same careful hand-over. It never accuses the customer.
UNKNOWNThe check could not examine the file, or is unavailableSays a team member will look at it. It does not guess.
  • It is an indicator, not proof. The agent is told never to call a file “fake” or “genuine” as a fact. A person on your team decides.
  • The reasons are saved with the conversation (what was found and how serious), so your team can see why the file was flagged. Open the conversation in Conversations, or in the test chat the step shows as Checked a file for fakes.

Limits

  • JPEG, PNG and PDF files, up to the attachment size limit (5 MB). Other types are not checked, and the agent says a team member will look at them.
  • Up to 5 checks in one conversation and 100 a day for your business, so a loop or abuse cannot run up a bill.
  • File checks run on files attached in the dashboard test chat and by visitors of your website chat widget (pictures and PDFs only, up to 3 files per message).
  • No automatic check catches every forgery or every AI image. Treat HIGH and REVIEW as “look closely”, and LOW as “nothing found”.

Privacy

  • The file is sent once to our verification service and checked in memory. The check does not store the file.
  • What stays with the conversation is the file name, the risk level and the short reasons, never the file’s contents.
  • Tell your customers, in your own privacy text, that documents they upload may be checked for authenticity by automated services.