Field Notes

ChatGPT ads, new Qwen models and practical AI workflows

The developments worth knowing, what they mean for your business, and concrete steps you can try.

Important AI news

ChatGPT ads: new tools for marketers, with different rollout stages

OpenAI announced new advertising features on 16 September. Sponsored Agents, which let someone open a business-sponsored conversation after clicking an ad, are being tested with selected US advertisers. An Ads Manager plugin supports campaign creation and analysis through prompts. HubSpot integration covers ads and lead follow-up. The Shopify app is available to US merchants, with expansion to other supported advertising markets scheduled for 23 September. These are separate rollouts, so the announcement does not mean every feature is available to every business.1

Why it matters: AI conversations are becoming another place where customers discover businesses. For a possible first campaign, our suggestion is to identify one offer and one useful outcome, such as a qualified enquiry. Check access and measurement before committing a budget.

Qwen-Image-2.1: transparent images and editing with multiple references

Qwen's new downloadable image model can create and edit images, generate transparent backgrounds, extract subjects from photographs and work with up to ten reference images. The official model card includes setup instructions and a demo link. Those capabilities could be useful for testing product cutouts, presentation assets or variations of a design. Running the model yourself requires technical setup.2

The business catch is the license: the research license permits non-commercial research and evaluation. Commercial use requires a separate license from Qwen. Downloadable weights do not mean unrestricted business use.3

Qwen adds audio/video analysis and a separate realtime model

Alibaba lists Qwen3.8-Omni-Flash as released on 17 September. It accepts text, images, audio and video and produces text. That makes summaries and analysis of recordings possible applications. The release table also lists a separate realtime variant on 21 September for its Singapore international service, supporting live audio/video interaction and text or audio responses. Access to these models requires integration into an application.4

The useful distinction: analysing a recording and holding a live conversation are different jobs. If you are planning a meeting assistant or customer-facing voice service, match the model variant to the job and test it on your own examples.

Three useful things for your working day

Turn meeting notes into a report and slides

A practical follow-up to last week's Claude news: Claude Docs and Claude Slides were announced on 16 September. You can create and edit documents and presentations inside a conversation, then export slides as PowerPoint or PDF. They are beta features on paid plans. The merger of Cowork and chat is rolling out first to Pro and Max over several weeks, so access may vary.5

Try this if the features are available in your account: give Claude a fictional project update and ask for a one-page report plus five slides. Specify the audience, decisions needed and unresolved questions. Check every number against your notes before sharing.

Get a second pair of eyes on a draft

Simon Willison describes using AI for spelling, grammar and fact-checking while keeping authorship himself. That suggests a useful editing pass for an offer, product description or customer email. AI feedback can also be wrong, so verify factual corrections yourself.6

A prompt to try: “Review this draft. List unclear statements, missing information and claims that need evidence. Explain each issue. Do not rewrite it yet.” You can then decide which changes improve your message.

Put AI into the step that needs judgment

In the published show notes for his Cognitive Revolution interview, Zapier CEO Wade Foster argues for using AI reasoning in selected steps and fixed code for the rest. This is an approach to workflow design, not evidence that any particular automation will work reliably.7

For a customer enquiry, our suggested split is: AI proposes a category and drafts an answer. Fixed rules assign the responsible team. A person checks the reply. Start by sketching that process on paper. No connector setup is needed to find the useful AI step.

Try this in five minutes

Hamel Husain's practical reminder on X: check AI classifications against human judgments, and avoid tuning only to the examples you already have. Here is a small exercise based on that idea.8

  1. Write down the expected categories for these fictional messages: “Please send me a copy of my invoice” → Billing. “The app crashes when I open it” → Technical support. “Can we arrange a demo for our team?” → Sales.
  2. Open your usual AI chat. Paste only the three messages and ask: “Classify each message as Billing, Technical support or Sales. Return the category and one short reason. If you cannot tell, say Unclear.”
  3. Compare the answers with your list. Count correct categories, then try one new message you have labelled yourself without showing the expected answer.

The result is a tiny, repeatable check instead of a vague impression. Three easy examples do not establish reliability. Before using this in a real process, add fresh, ambiguous cases and decide what needs human review.

One idea worth keeping

In “The Overhang”, Ethan Mollick describes choosing AI projects in areas he knew, spotting flaws and asking for another version. Our takeaway: start with a task whose result you can judge. For your next experiment, write down the required output and two acceptance criteria before prompting. Your knowledge of the business gives you something concrete to check.9

Sources

  1. Reimagining advertising with AI (openai.com)
  2. Qwen releases Qwen/Qwen-Image-2.1 on Hugging Face (huggingface.co)
  3. Qwen-Image-2.1 Research License: commercial use requires a separate license (huggingface.co)
  4. Qwen3.8-Omni-Flash: audio/video analysis and new realtime variant (alibabacloud.com)
  5. Claude Cowork and chat are now one Claude (claude.com)
  6. How To Write With An LLM (simonwillison.net)
  7. No Code Is Code: Zapier CEO Wade Foster on Headless Tools, Zapier MCP & Automation Bench (cognitiverevolution.ai)
  8. @hamelhusain: Can you use Jev for Evals? Yes! Remember that a LLM Judge is also classifier*.

Make sure to test your classifiers against human labels and don't overfit.

Hope this helps!