message_id as the canonical per-message auction identifier.
Within AIP, AI Platforms are the protocol’s entry point - they surface intent and deliver outcomes, while all governance, selection, and settlement processes are handled by the Operator.
What is a platform?
A Platform is an application, interface, or agentic environment that:- Receives user intent AI Platforms capture user queries, commands, or questions expressed in natural language.
-
Emits a
PlatformRequest(AIP canonical request format) to an AIP Operator The Operator evaluates participation eligibility, derives a ContextRequest, then handles selection and mode determination. - Renders recommendations (recommend mode) AI Platforms display operator-returned recommendations using protocol-standard formats (weave, tail, or product card).
- Manages delegation consent (delegate mode) AI Platforms present delegation offers to users, collect consent, and relay that consent to the Operator for delegated session initiation.
- Tracks and forwards signed events AI Platforms report participation and interaction events using the operator SDK, ensuring serve-token integrity and attribution.
- Receives payouts for verified events Operators credit platforms for valid outcomes based on AIP settlement rules.
Examples of platforms
Examples of platforms include:- AI assistants and copilots
- Chat-based search interfaces
- Browsers with conversational or agentic UI
- Voice assistants
- Productivity applications with embedded AI
- Messaging applications with AI features
- Browser extensions or agent runtimes
- Augmented reality (AR) and virtual reality (VR) environments with conversational interfaces
- Autonomous multi-agent environments
Platform responsibilities
AI Platforms MUST
- Integrate with an operator through the canonical AIP APIs and event contracts
- Maintain a session identifier for each user conversation
- Display transparency labels for commercial participation
- Ensure accurate event propagation
- Provide
latency_budget_msin PlatformRequest to specify the selection timing window - Present delegation offers clearly and collect explicit user consent before handoff
- Make clear when the user is entering a brand-agent-managed delegated task flow
AI Platforms MAY
- Use operator-provided SDKs or backend libraries when convenient
- Add platform-specific UI or UX layers around recommendations
- Surface recommendations inside various modalities (text, voice, multimodal)
- Implement custom delegation consent UI
Selection interaction
The AIP selection model uses a time-bounded asynchronous selection window.How it works
- Platform sends a
PlatformRequestwithlatency_budget_msto the Operator - Operator evaluates participation eligibility
- Operator derives a
ContextRequestand filters brand agents that match the context and governance constraints - Operator distributes (broadcasts) the
ContextRequestto all matching brand agents - Operator computes the available response window from the platform’s
latency_budget_ms - Brand agents submit signed responses within the computed window (auction-style participation)
- Operator runs selection, determines interaction mode, and returns the
PlatformResponse - If no responses are received before the window expires, Operator returns a valid
no_matchresponse
AI Platform implications
- Always handle
no_matchresponses gracefully (render a fallback or continue the user workflow) - Record participation attempts in your telemetry even when no match is returned
- Specify
latency_budget_msin PlatformRequest to control selection timing
Delegation responsibilities
When the Operator returns a PlatformResponse with delegate mode:- Present the offer (exposure) - Show the user which Brand Agent is available and what it will do. Fire
exposure_shownwhen the offer is displayed. - Collect consent - The user must explicitly agree before any session is initiated.
- Confirm consent to the Operator - Send the user’s consent to the Operator. The Operator checks brand-agent availability, initiates the delegation session, and records
delegation_started. - Run the delegated session (multi-turn) - After the operator-mediated session start, the live task flow may continue directly between the Platform and Brand Agent. The Platform may host or relay that session, while the Operator remains the governance, audit, and settlement layer.
- Track outcomes and expiration - The Platform fires
exposure_shown, any applicableinteraction_started, and delegated-sessiondelegation_activityevents while the session remains active. The Operator recordsdelegation_started, resets inactivity state from verified activity, and recordsdelegation_expiredif the timeout elapses before the Brand Agent submitstask_completed.
- The Platform should make clear that the user is interacting with the selected Brand Agent
- Additional task information may be collected inside the delegated session itself
- The Platform should emit delegated-session activity or keepalive events while the user is still active in the session
- That new information is session-bound by default and is not automatically part of the original operator handoff scope
- No session is initiated
- The interaction falls back to recommend mode if content is available
exposure_shownmay still be tracked