OpenAI is testing Sponsored Agents with selected advertisers in the United States. The feature lets a person choose a separate, clearly labelled business-sponsored conversation after seeing an ad. South African businesses should prepare accurate information and safe handovers, but should not assume the test is available locally or that advertising buys influence over independent answers.
The useful preparation is largely ordinary business work: clear product or service details, supported claims, current policies and an identifiable person responsible for exceptions. A conversational interface makes those foundations more important, not less.
Reviewed 20 September 2026. This article analyses OpenAI's 16 September announcement and current help material. It is not a hands-on Sponsored Agents test, a promise of local availability or an offer of a D4J integration with this test.
What did OpenAI announce?
The official announcement describes Sponsored Agents as an optional conversation a user can start after encountering an ad. It says the sponsored conversation is distinct from ChatGPT's independent answers and separate from the original conversation.
The same announcement also covers advertising-management tools, creative assistance, optional text customisation and integrations with HubSpot and Shopify. These are separate capabilities with their own availability conditions. They should not be compressed into a claim that any business can now launch a fully autonomous advertising agent everywhere.
At announcement, Sponsored Agents were being tested with selected US advertisers. The Shopify app was described as available to US-based merchants, with international availability planned for 23 September in markets where ChatGPT Ads is available. That future date is not evidence that South Africa is included.
OpenAI's ads help page also distinguishes ads from independent answers. Buying advertising is not buying an organic recommendation or a guaranteed citation.
What should a South African business verify first?
Check the current advertiser-country and account eligibility before planning a launch. Do not infer access from the ability to use ChatGPT, from an overseas case study or from a software subscription.
Establish the exact feature being discussed. Access to an advertising account does not necessarily include access to Sponsored Agents. An integration available through one CRM does not establish the same integration for another CRM.
Then check the operational constraints: billing, business verification, relevant policies, supported destinations and who may approve campaign changes. Keep exploration separate from live spend. A useful feasibility check can end with “not available to this account yet” without becoming a failed project.
Prepare a trustworthy business information pack
Start with the information a prospective customer needs to decide whether to proceed:
- What the product or service does and does not include.
- Current prices, or an honest explanation of how a quote is determined.
- Service areas, eligibility and delivery conditions.
- Supported comparisons and genuine evidence.
- Returns, cancellation, booking or support policies where relevant.
- The next step and the person or team responsible for it.
Give each item an owner and a review date. Keep unsupported claims out of the source material. If the website promises something the sales team cannot fulfil, a conversational agent can repeat that inconsistency more efficiently.
Do not upload a private customer database merely to make answers sound personalised. Decide which information is appropriate for the purpose, what permissions apply and what should never enter the system.
Our AI automation readiness guide provides a broader checklist for information, access and human responsibility.
Decide what the agent must not do
A sales conversation can quickly become consequential. A customer may ask for a binding discount, a guarantee, professional advice or a change to an existing order. Decide which answers can be supplied from approved information and which require a person.
For a hypothetical installation service, an agent might explain published service areas and collect a request for assessment. It should not invent a fixed installation price where a site inspection is required. The appropriate next step is an honest handover, not a confident estimate presented as an agreed quote.
Define the boundary for booking, payment, cancellation and account changes separately. Do not assume that the ability to discuss a service gives the system authority to transact on behalf of the business.
Test the difficult questions, not only the easy ones
If access becomes available and a pilot is appropriate, prepare a review set before launch. Include ambiguous requests, outdated information, unsupported locations, conflicting instructions and questions whose answer is genuinely unknown.
A useful result can be a correct refusal to promise something, an explicit clarification or a handover. Do not score every referral to a person as a failure. The business objective is a reliable customer experience, not maximum autonomous completion.
Review factual accuracy, disclosure, privacy, escalation and whether the next step works. Preserve examples of failures without exposing customer details. Assign an owner to correct the information or configuration and retest it.
How would you judge commercial value?
Keep the stages separate: an ad impression, a chosen sponsored conversation, a website visit, an enquiry and a commercial outcome. Do not treat a longer conversation as proof of better lead quality.
Use the reporting the product actually provides, explain its attribution limits and reconcile meaningful outcomes with your own records where appropriate. Define a spending boundary and a stopping condition before testing. Vendor examples are not a forecast for your account.
The same discipline applies to ordinary automated follow-up. A system that answers more messages but creates unsupported commitments can be commercially worse than a slower, well-controlled process.
The sensible action now
For most businesses without confirmed access, prepare the information and boundaries rather than rebuilding the CRM around a limited test. Those improvements also help the website, human sales team and existing support channels.
See our AI automation scope for bounded workflow planning and CRM setup for enquiry handover. Use the existing Discuss project form to describe the customer questions and operational constraints. Product availability must be confirmed before any specific deployment is promised.
