AI Update Q3/2026

Every quarter, I review AI developments from the past twelve weeks and ask myself: What has actually changed enough to matter outside of the AI bubble?

So this week, I went through my notes to prepare my Q3 AI Briefing – a quarterly webinar I host for clients – and boiled it down to seven shifts that I found most relevant.

I’ll share the full analysis in the workshop, but today I’d like to highlight the three main ones for you.

Because although we saw many models, it wasn't the models themselves that caught my attention.

Let's dive in!

One of the major shifts for me this quarter has been that AI increasingly stopped feeling like a chatbot that waits for the next prompt.

After Claude Cowork, OpenAI launched ChatGPT Work in July. The main difference between “Chat” and “Work” mode is that you shift from talking to an LLM to get answers to talking to an LLM to delegate tasks using tools. The model works inside a larger agent environment that does the orchestration and reports the results back to you. I wrote about this recently in Leaving the Chatbot Era Behind.

Meta just pushed in the same direction – with apparently big success. Muse launched earlier this month and reached around 2.8 million downloads in its first 12 days. It’s a “personal AI Agent” that comes dressed in a plushy-naughty costume but is essentially a personal VM in the cloud that can just do things for you like sending emails, planning travel, or shopping across supported services. Imagine OpenClaw, but without the hassle.

“Meta is back” some people say, but let’s see how many will stick to Muse after the initial launch hype. The vision for Muse is to be accessible anywhere – via Meta’s AI glasses, dedicated hardware, and even its own email address.

Microsoft just announced the “Muse for Business” with Copilot Autopilot (and once more winning the “worst product name” award).

So all the big AI labs are trying to do the same:

Shifting the unit of interaction from questions to tasks.

Instead of asking an AI and then staring at a chat window to get an answer, the workflow increasingly looks like this:

Make no mistake – the chat interface isn’t going away. I probably chat with AI more than ever. Increasingly, however, chat also becomes the interface for asynchronous task delegation.

Q3 was also when McKinsey’s highly anticipated “State of AI” report got another update. And one number that caught my eye was this:

That’s a pretty big deal.

For decades, the economics of software were relatively straightforward: Building custom software used to be pretty expensive. So unless you wanted something highly specific or strategically very important, you just bought the tool. SaaS for the win.

Then AI started to mess with that calculation and triggered the SaaSpocalypse earlier this year:

I wrote about how enterprises can take advantage of it in Cheap Codification a few weeks ago. The interesting thing really isn’t that everyone and their mother can now build apps (I expect the vast majority will still never build their own apps). But what really changed the game is that the cost of turning business rules into working software is collapsing.

One of the key reasons that business rules were never spelled out as code was simply that it was too complicated to do. Now you can literally talk to ChatGPT and it will spit out n8n workflows (or PowerAutomate, or Zapier, or whatever you want.)

The other day I had someone from a financial back office team tell me how they used Copilot to write PowerShell scripts because PowerShell turned out to be the only application that didn’t need further approval.

Brave new world.

Interestingly, most of these solutions don’t even need AI anymore.

The PowerShell script was renaming hundreds of files in a specific way that was otherwise done manually.

And this is not just my interpretation. Forrester recently argued that the traditional “build vs. buy” question itself is becoming outdated, because AI makes it much cheaper to customize, compose, or generate software.

But reality is messy and this doesn’t apply to every software category equally – I’ll talk more about this in my upcoming briefing. With regards to complex enterprise software, I still believe the basic principle from my AI Make or Buy article holds:

Buy the commodity technology. Build what makes your process different.

What changed is the price at which building becomes economically sensible.

So if you’re evaluating another $50K+/year SaaS tool for a fairly specific internal problem, there’s a new question worth asking:

What would it cost us to just build the part we actually need?

The third shift is less exciting, but probably one of the most immediate – especially if your business is operating in Europe.

On August 2, another major part of the EU AI Act went into actual enforcement.

One of the most “obvious” effects is that – depending on the system – providers and deployers now have to disclose when users are interacting with AI, label certain AI-generated or manipulated content, and make some synthetic content machine-readable as AI-generated. The EU even designed some chic logos along with guidelines!

But back to the EU AI Act.

Companies here are increasingly starting to wonder:

These are, in fact, great questions even without regulation.

It turns out, a lot of the basics of bringing an AI use case into production – and keeping it there – will also get you surprisingly far when it comes to compliance.

So these were three shifts out of seven that I noticed in Q3.

If you’re interested in the full briefing, which also includes current model updates, open models, operational risks and why cheaper intelligence doesn’t necessarily mean cheaper AI – then check out my upcoming webinar here.

See you around,
Tobias