AI Disclosure

EU AI Act Article 50

uses AI to help you work. This page lists every AI capability you can interact with, what it does in plain language, and whether it asks for your explicit consent before it acts. Capabilities marked consent required stay off until you turn them on.

Capabilities that require your consent

AI-composed email

Consent required

The assistant can draft and send email on your behalf, in the voice of the mailbox that sends it, and reads what an email carries -- its documents, the text in its images and recordings, and an attached video through its frames -- on this device. AI-composed messages carry a clear AI-generated disclosure (EU AI Act Article 50).

Autonomous bookkeeping

Consent required

MyMoney reads invoices and tracks your P&L. Any operation that changes your financial records requires your explicit approval.

Multichannel agent control

Consent required

Reach and command your assistant across messaging channels, and let it settle matters with another person's assistant over signed, sealed email. Messages routed through connected accounts are processed locally.

Live call and meeting captions

Consent required

Transcribes speech to on-screen captions while a recorded call or captured meeting is in progress, using a speech-to-text engine that runs entirely on your own machine. The live transcript is not persisted; only the final transcript is stored. Recording, capture and import stay behind the one-time acknowledgment.

Standard capabilities

Intelligence Nucleus orchestration

Two models work together -- one plans, one executes -- with automatic quality checks and self-correction on each step. Every step is recorded so you can see exactly what the assistant did.

Source-grounded answers

Answers are checked against your sources; claims your sources do not support are rejected rather than shown.

Personal model fine-tuning

The assistant learns the workflows you approve and fine-tunes a personal model, trained entirely on your own machine.

Local adaptation of the agent's own instructions

Always on. The assistant's planning guidance can change only through a closed set of typed adjustments -- a preferred execution lane or a parallel planning strategy for a class of request, or the removal of a tool from a lane -- proposed by fixed rules over counts of outcomes, never over text. Each adjustment is shown to you and used only after you approve it, is measured live against the previous behaviour, and can be reversed at any time; most evaluations end with no change. It runs entirely on your own machine and produces no content, so there is nothing for it to mark as AI-generated.

Local image, video, speech, and 3D model generation

Creates images, video clips, synthesized speech, and 3D models reconstructed from a picture, entirely on your own machine. Every generated file carries an embedded machine-readable AI-generated marker in its format's standard metadata.

Answers in the form a step asks for

A step can constrain the on-device model to the exact form it needs -- a JSON schema, a pattern, or one entry from a fixed list of choices -- instead of asking for it in prose. An answer that does not meet the constraint is asked for again rather than shown. It runs on your own machine and changes nothing about what leaves it, which is nothing.

Audio the assistant hears

Audio clips you attach, paste, drop or record are heard by the on-device model alongside your message, when the loaded model takes audio. Clips stay on your machine, are sent to nothing else, and only their name and duration are kept in the conversation.

Body pose from an image

Estimates the body pose of the people in an image you choose -- keypoints, joint angles and posture -- entirely on your own machine. It does not identify anyone and infers no emotion or personal characteristic; the result is geometry only.

How AI-generated content is marked

When the agent autonomously composes and sends a communication to a third party, marks it as AI-generated. An email through the built-in mail tool carries two marks, over either supported mail protocol:

  • Human-readable line -- a short note in the message body (for example, “Sent by [your name]'s AI agent -- RunMyAI”).
  • Machine-readable header -- an X-RunMyAI-AI-Generated header, which is also read back on inbound mail so an incoming agent message is recognized as one.

An agent-to-agent message carries the machine-readable header. It carries no body line, because there is no human-readable body to carry it in: the message body is a signed, encoded envelope addressed to another agent, already labelled as an agent message at the protocol level, which no person reads as prose.

This marking is applied automatically and cannot be suppressed for autonomous agent sends. It is not applied to messages you write yourself and then send -- you are the author of those. Nor is it applied to the connection metadata the app exchanges to set up a call, which is machine signaling rather than a communication the agent composed.

Every file the AI generates also carries a machine-readable AI-generated marker embedded in the file itself, in the standard metadata mechanism of its format:

  • Images (PNG) -- an AIGenerated text chunk plus an XMP provenance packet.
  • Video (MP4) -- a marker box in the container.
  • Video (WebM) -- two Matroska tags, AIGenerated = true and Generator = RunMyAI, written by the video encoder and checked on every finished file.
  • Synthesized speech (WAV) -- a comment chunk in the container.
  • Generated PDFs (documents, invoices, reports) -- PDF document metadata plus an XMP provenance packet.
  • Exported drawings built from a reply (SVG) -- an XMP provenance packet in a <metadata> element after the root tag.
  • Quick action results exported as PDF or XLSX -- PDF document metadata plus an XMP provenance packet; an XLSX document property.
  • Certificate PDFs (they carry the model-written reasoning summary) -- PDF document metadata plus an XMP provenance packet.
  • Exported documents and spreadsheets (DOCX/XLSX/PDF) -- an AIGenerated custom document property (DOCX/XLSX) or document metadata (PDF), applied when the exported file carries agent-written content.

The XMP packets carry the industry-standard IPTC digital-source-type term for AI-generated content (trainedAlgorithmicMedia), so any tool that reads the common provenance vocabulary recognizes the file without knowing RunMyAI.

When you export a document or spreadsheet that contains content the agent wrote -- an agent-authored document, or agent edits you accepted -- the exported file carries the embedded AIGenerated marker listed above. A drawing or diagram the agent rendered from a reply, saved as an SVG file, carries the same marker as an XMP provenance packet. A quick action's result saved as PDF or XLSX, and a certificate PDF, are marked before they are saved, and the save is refused if the marker cannot be written; until 2026-10-09 those two saves applied no marker. A document you authored yourself is your own work and is never marked; where the agent only assisted your own standard editing without adding its own content, no marking obligation arises (EU AI Act Article 50(2), assistive-editing exception) and none is applied.

Text the AI generates inside the app -- chat answers, previews, and edit proposals -- is part of the AI interaction this page discloses and needs no separate mark. When AI-generated text leaves the app in a carrier with a standard metadata mechanism (email, documents, spreadsheets, PDFs), it is marked through that carrier as described above. Plain-text file formats (such as .txt and .md) define no metadata mechanism at all -- the file is nothing but the text itself, and altering the visible text to embed a mark would corrupt your content. Article 50(2) requires marking only as far as technically feasible, and no interoperable marking technique for raw text exists today, so plain-text files are not marked. If you publish AI-generated text to inform the public, the transparency duty for that publication rests with you under Article 50(4), which itself excepts text that has undergone your review and editorial control.

Important limits of AI output

AI output is probabilistic and may be wrong, incomplete, or outdated. Always verify before acting on it. The agent acts from your computer, using your internet connection and IP address; actions it takes that you authorized are attributable to you. See the full Terms for details.

Your choices

  • You set the autonomy level per product and per surface (default Manual).
  • The level you choose applies to every area; you can set any area to ask every time, and at Level 0 nothing is carried out.
  • Unless you switch it off, the agent asks before any action with an outside effect taken after it read external content, whatever the level.
  • Anonymous performance data is contributed automatically to the community ModelIndex, the same way everywhere; the report carries no per-device identifier, and in return you get model recommendations matched to your hardware.

Contact

Questions about AI use: legal@runmyai.app.