What to Expect from AI in 2026

Making predictions about AI is a bad habit.

What looks obvious today, could be obviously wrong a few weeks from now.

And yet, here I am – making AI predictions for the fourth year in a row.

The goal isn’t to be perfectly right, but to force reflection. To notice when something bigger changes direction.

So let's look at what worked (and what didn’t) in my 2025 predictions – and what I expect for AI in 2026.

Spoiler: AI is growing up. And grown-up AI should live up to different standards.

Looking Back: My 2025 Predictions

My predictions for 2025 said that AI agents wouldn't deliver the productivity gains everyone expected. Which is quite what happened. AI agents made a huge leap, but Microsoft still had to cut sales targets and we probably all still remember that infamous “95% of AI projects fail” study. By the end of 2025, production-deployed AI agent systems remain rare.

The other prediction I made was that 2025 will be the year where AI needs to show profit. And in fact, nearly every client conversation I had this year came back to the same question: Where's the ROI? AI costs are obvious. But AI value isn't. This became such a dominant theme that I dedicated a book on it. The shift from "can we build this?" to "should we build this?" happened.

As a bonus, I called out that xAI (Grok) should be under close watch as they’d rival the back-then leader OpenAI. By the end of 2025, xAI hasn’t become the unrivaled leader, but they’re playing in the champions league.

So overall, those points happened pretty much as expected.

Let’s see if I can keep that streak.

My AI Predictions for 2026

Prediction 1: AI Returns to the Office

I wrote earlier this year about bringing AI workers back to the office. In 2026 I expect that this shift will reach critical mass.

The current state of Enterprise AI means cloud dependency. Which is fine until you either process high-volume load, or sensitive documents you don’t want to share with another 3rd party. In these cases, sending data to external servers and paying consumption-based with zero cost predictability becomes a major bottleneck. “So what?” you could say? Well, it turns out that many of the AI-relevant enterprise workflows are exactly that: high-volume and critical data.

Why 2026 is different:

Three things changed at once.

1. Open-source models hit quality parity. OpenAI's OSS-120B. Alibaba's Qwen. Kimi K2. These deliver good-enough performance for practical business tasks.

2. Consumer hardware can run them. A high-end GPU in a workstation handles production workloads.

3. The software ecosystem matured. Non-specialists can deploy local AI stacks.

For high-volume workloads like document classification or data extraction - local deployment costs less than cloud APIs. One-time hardware investment vs. infinite consumption costs. And for regulated industries - healthcare. Legal. Finance - local processing becomes a major enabler.

For you this means:

• Identify workloads: high volume. Sensitive data.

• Start small: one workstation. One GPU.

• Plan hybrid: local for sensitive. Cloud for frontier.

This is what AI looks like when it’s treated as infrastructure, not magic.

Prediction 2: AI ROI Measurement Gets Serious

In 2025, people realized that AI had to prove its worth – otherwise it would just be another expensive IT line item. But how do you measure the impact of something so fuzzy as AI? That’s where I expect 2026 to give us a lot more structure.

Traditional IT ROI frameworks don’t work for AI because costs aren't front-loaded. They scale as you use the system. Forecasting isn't stable. You don't even know what happens next year. And AI outputs can feel random, with inherent variance.

When boards demand ROI using ERP-era frameworks, business units respond with fantasy math. Or abandon measurement entirely. Finance blocks investment. Everyone loses.

I see a practical consensus emerging. I call it the Two-Track AI ROI approach

Track 1: Productivity AI - ChatGPT. Copilot. Coding assistants. Measure adoption and satisfaction. Not ROI. Your sales team writes faster emails but sends the same number. That's fine. Just don't pretend it directly translates into more measurable profit.

Track 2: Engineered AI - Process automation. Custom applications. This needs hard financial discipline. Cost caps. Value targets. Accountability. Otherwise, this will quickly turn into a cost sink. On the other hand, that’s where the potential for true P&L impact lies.

It’s also the reason why my book is called The Profitable AI Advantage, not The Productive AI Advantage.

These two tracks are just the tip of the iceberg. The organizations that thrive won't spend the most on AI. They'll know which investments build capability and which must deliver real returns.

This is what happens when AI is finally forced onto the P&L.

Prediction 3: AI Stops Asking for Permission

The passive chatbot era is fading. So far, the post-ChatGPT AI dynamic was simple: You ask. It answers. You're the engine. It's the fuel.

In 2026, this paradigm is flipping. Announcements like Google's CC or ChatGPT’s Pulse are a glimpse into that future. AI doesn't wait for you anymore. It sends the notification. It suggests the follow-up. It evolves from a being a small brain you can ask, to being an arm that moves on its own.

Why this matters: The next big bet is agentic AI. Solutions that resolve issues without you ever seeing them. But you don't hand the keys to a machine until you trust it.

The proactive phase is the audition.

AI scans your inbox. Identifies a missed invoice. Earns trust to do it again. Then to draft a follow-up. Then to send it without asking. Every step done right earns permission for the next.

The failure mode is obvious. If AI spams you with "Good morning" notifications - people turn it off immediately.

That’s why AI winners in 2026 will be the ones that earn enough trust to disappear into the background (not the ones with the best chat interface).

Bonus: Watch Google

Last year I said watch xAI. This year: watch Google.

Google always had the money, data, and talent. What they lacked was the willpower and execution spirit to skip the bureaucracy and get moving.

The inflection point likely came when a guy in a boob t-shirt put his lunch aside to ask Sergey Brin why Bard sucked so much – and the Google cofounder couldn’t find a good answer. Today, Sergey is officially out of retirement, extremely hands-on, and calling Sundar whenever something needs to get out of the way.

As of today, Google leads the intelligence-efficiency frontier of modern LLMs. 2026 might be the year Google reminds everyone why they dominated AI research for a decade before ChatGPT made it cool.

Conclusion

You'll hear from me in a year on how these turned out.

The main theme for 2026 will be: AI is growing up. Local infrastructure instead of cloud dependency. Financial discipline instead of fantasy math. Earning trust instead of demanding attention.

If you want me in your corner in 2026 to help you turn AI from a cost center into a profit center, send me a DM with "2026" and I'll share what working together looks like.

See you next Friday,
Tobias