IDHL Labs AI Edit: September 2026

It's been a rollercoaster of a month for AI.

Some of the people building the most powerful models are publicly questioning whether things are moving too quickly. Researchers have warned about advanced AI becoming difficult to control, while Anthropic’s own CEO has called for development to slow. 

Concerns around AI risk aren’t new, but it's harder to brush aside when the warnings are coming from inside the industry.

And yet, while that debate plays out, AI is finding its way into more of the ordinary things we do online. This month, we've seen it built further into how we advertise, edit video, shop, and use business software.

One conversation is asking where the tech will eventually take us, while the other is working out where AI fits into the tools and platforms people use every day. And that's where most of this month's updates sit.

1. ChatGPT ads can start a conversation

OpenAI is testing Sponsored Agents, a new ad format that lets someone start a conversation with a brand's AI agent without leaving ChatGPT. Wayfair is among the brands testing it.

Rather than sending a user to a landing page that tries to answer their questions, they can now just ask them from within ChatGPT. For brands, it puts more of the customer relationship somewhere other than their own site.

2. Amazon gives advertisers a route into ChatGPT

Amazon Ads has partnered with OpenAI so select US advertisers can extend campaigns into ChatGPT through Amazon DSP.

OpenAI still controls how the ads appear, but advertisers can access ChatGPT through a platform they're already using. It makes a new advertising channel much easier to buy without creating a separate setup for it.

3. Adobe puts AI generation inside the edit

Adobe Premiere's new Generative Media Tool lets editors generate video and sound effects from prompts inside their timeline, then add them as editable clips.

Putting generation directly into Premiere removes the need to jump between an AI tool and the edit. It keeps AI inside a workflow, rather than sending them somewhere else to generate the asset first.

4. Microsoft lets AI decide what your keywords won’t

Microsoft has made AI Max available across Search campaigns. It can find relevant searches beyond an advertiser's keyword list, create messaging variations, and choose the landing page it thinks best matches someone's intent.

Advertisers still get reporting and controls, but Microsoft is making more of the decisions between the original query and the ad someone eventually sees. For search teams, it places more emphasis on the inputs, boundaries, and judgement around the system compared to manually mapping every route.

5. Google shortens the journey from ad to conversation

Google has made multimodal video creation generally available in Demand Gen's Asset Studio, letting advertisers storyboard and create different video formats.

It's also testing Demand Gen ads on YouTube that let viewers start a conversation with a brand through messaging apps, giving customers a more direct route from the ad to actually speaking to the business.

Both updates give advertisers more options at either end of the campaign: creating the ad and giving people somewhere to go when they respond to it.

6. Anthropic wants retailers to build their own shopping agents

Anthropic has released blueprints that businesses can use to build Claude-powered commerce agents across retail, travel, and telecoms.

A lot of the AI shopping conversation has focused on people asking a general assistant what to buy, but Anthropic is giving retailers another route: building agents around their own products and customer experience.

Brands should think about both sides of AI shopping: how their products show up in assistants like Claude, and what kind of AI shopping experience they offer customers themselves.

7. Meta launches its own AI agent

Meta has launched Muse in the US, a personal AI agent designed to take on tasks rather than just answer questions. It can browse the web, make purchases, book appointments, and work across connected apps, with some tasks continuing in the background.

For a personal agent to be genuinely useful, people need to be comfortable giving it access to quite a lot of their digital lives. That makes trust particularly important for Meta, a company that has faced scrutiny over its handling of user data in the past.

8. Salesforce is letting Claude become the interface

Salesforce and Anthropic have announced Claudeforce, starting with 37 sales skills that let Claude work with live Salesforce data, update pipelines, and take governed actions without the user leaving Claude.

Software companies have spent years trying to get people into their interface and keep them there. Claudeforce accepts that someone might choose to work in Claude instead, while Salesforce provides the data, permissions, and actions underneath it.

In the end...

For years, businesses have put effort into their digital presence: their website, landing pages, ads...and critically, what happens when someone clicks on them.

Today, more of those interactions could start and end somewhere else, with an assistant deciding what information to surface, which systems to call on, and what happens next.

So, while everyone tries to work out when AI is going to destroy us all, there's a more immediate question for businesses: what happens to your customer experience when you no longer own the interface?