Scope and evidence
Sponsored Agents: published specifications and open questions
Checked 4 October 2026. This is a readiness assessment, not a working integration or an OpenAI accreditation. The GPT Agency's proposed CEM is Conversation Engine Management; earlier material used CEM for Conversation Engine Marketing.
What is established
OpenAI describes a clearly labelled business-sponsored conversation opened from an ad, distinct from the original ChatGPT conversation, with a possible onward link to the business. Its 16 September announcement describes selected US advertisers; the help page calls availability a limited alpha. OpenAI announcement, availability help.
The general help page says early-access requests are not being accepted. A separate restricted HubSpot interest form nevertheless invites businesses to register for possible future testing, with eligibility and spend requirements confirmed at onboarding. This is an exception to the broad earlier statement that there is no interest route; neither page establishes our access. No application has been submitted.
OpenAI's Ad Tools Terms describe advertiser responsibility for agent configuration, actions and outputs, and require appropriate rights to supplied materials. A commercial Pre concept would require a separate rights and brand-authorisation review; our illustrative persona is not an authorised historical-person replica or a Nike deployment.
What the public material does not establish
| Specification needed | Status in reviewed sources | Our consequence |
|---|---|---|
| Import format, instructions API and file/context limits | Not established | Build a portable internal persona; do not call it a validated Sponsored Agent package |
| Model selection, voice, languages and audio streaming | Not established | Our Realtime web demo is a different system |
| Supported tools, MCP and live-data connectors | Not established | Prepare capability mappings; test each host separately |
| Character and campaign settings writable by our CEM | Not established | Export reviewed changes; no automatic OpenAI publication |
| Outbound handoff event IDs, receipts and export | Not established | Persona CPC billing is conditional on an authorised measurement route |
| Agent-specific transcript access, retention and private export | Not established | No promise of exporting ChatGPT conversations into Nike's AWS |
| Agent operating charges and agency revenue share | Not established | No claim that platform fees fund our work or give us a cut |
| Approval, version rollback and service limits | Not established | Treat release integration as a discovery milestone |
The existing Ads API partner guide covers account access, brand review and conversion setup. It does not, by itself, document persona import or a Sponsored Agent lifecycle. An ordinary OpenAI model API key is not evidence of Ads account access.
Proposed deployment method
Keep a versioned constitution, character, sourced knowledge, interests, approved destinations and measurement policy in our own format. Compile only host-supported fields through an adapter. Every adapter has a capability report: supported, unsupported, unverified. Preserve evidence references and record any lost behaviour. Run the Gauntlet in the destination environment before claiming compatibility.
For Pre, prepare a rights-cleared brand character, approved catalogue and sporting source pack, commercial destinations and scope limits. Distinguish a product-discovery agent from a private member companion: health-related disclosures must not become ad targeting, leads or sales persuasion. A shared soul can have different permissions in different environments.
Measurement and privacy
The platform's entry click is separate from our proposed outward handoff CPC. We need an authorised event or destination receipt that can reliably distinguish those clicks, reject duplicates and invalid traffic, and reconcile disputes. Redirecting links or adding tracking is subject to the host's capabilities and approval; no such integration is installed.
OpenAI describes advertiser reporting as aggregated rather than access to private conversations in its ads buying announcement. Public material reviewed does not settle agent-specific log export. Brand-hosted website/app conversations are a separate data flow whose collection, access and retention we can design with the client. A private cloud is an access boundary, not an anonymisation technique.
Discovery questions before a Sponsored Agent commitment
- Which advertisers, countries and partners can participate, and through which route?
- What exact build inputs, tool permissions, voice capabilities and update interfaces are supported?
- Can our approved character be reproduced, tested and versioned? What must change?
- Which reporting events identify an outward handoff, and may we invoice against them?
- What conversation data may the brand receive, for what purposes, and with what controls?
- Who pays inference and media costs, and what limits, approvals and rollback facilities exist?
We can sell preparation and specialist development today. A promise of seamless Sponsored Agent deployment must wait for access, written specifications and a successful end-to-end test.