What’s Next in AI: A 2026 Forecast for Enterprise Leaders

February 2024. Klarna’s CEO Sebastian Siemiatkowski takes the stage with a bold claim: his AI chatbot is doing the work of 700 customer service agents. The fintech world applauds. Efficiency! Innovation! The future!

Fast forward fourteen months. Same CEO, a very different tune, in comments to Bloomberg.

“As cost unfortunately seems to have been a too predominant evaluation factor when organizing this, what you end up having is lower quality.”

Klarna is now actively rehiring human agents, targeting students, rural workers, even passionate customers who want to work for them.

I keep thinking about this story. Not because it’s a failure. It isn’t, really. Klarna learned something expensive and had the guts to admit it publicly. That’s rare. What strikes me is how perfectly it captures where we are right now with AI. The gap between what we’re promised and what actually works has never felt wider.

I’ve been writing these annual AI forecasts for a few years now. (Here’s last year’s, if you want to see what held up.) Every January, same goal: cut through the noise and figure out what actually matters. This year feels different though, and not because of some magical “inflection point.” The industry claims that every year is an inflection point. This one feels different because the reckoning is finally here.

1. Agentic AI: Overhyped AND Underhyped (Yes, Both)

Every tech conference in 2025 declared it “The Year of the Agent.” Gartner predicts 40% of enterprise applications will include AI agents by end of 2026. But McKinsey’s 2025 State of AI survey tells a more sober story about agentic AI specifically: just 23% of organizations are scaling an agentic AI system, and another 39% are still experimenting. The 88% figure you’ve probably seen quoted everywhere is real, McKinsey’s own numbers confirm it, but it measures general AI use in at least one business function, not agentic AI specifically. Agentic AI is a small, fast-growing slice of that, not the whole picture.

Sounds great. Now let me tell you about July 2025.

Jason Lemkin, founder of SaaStr, was experimenting with Replit’s AI coding agent when, according to Fortune, the agent experienced hallucinations, faked reports, and created fake algorithms “to make it look like it was still working.” Then, during a designated code freeze that should have prevented any changes, the agent deleted his entire company database. Months of executive data, gone.

“This was a catastrophic failure on my part. I destroyed months of work in seconds.”

New York City’s MyCity chatbot, launched in October 2023 to help entrepreneurs, was caught in early 2024 advising businesses to break the law: fire employees for reporting harassment, keep customer tips. In Australia, Commonwealth Bank cut 45 customer-service roles for an AI voice bot in July 2025, then reversed the decision a month later after the union pushed back and the bank admitted its call-volume assessment was wrong.

The Real Problem Nobody Talks About

Most AI failures aren’t because the models aren’t smart enough. They fail because companies dump terabytes of messy data into a system and expect magic. They throw years of Slack messages into training sets and wonder why the AI doesn’t understand their “culture.”

Andrej Karpathy framed this shift back in 2023, in a post that is still widely cited:

“...a more complete picture is emerging of LLMs not as a chatbot, but the kernel process of a new Operating System.”

Karpathy’s point: the model becomes the kernel of something bigger. The orchestration, the guardrails, the memory, all of it still has to be built around that kernel. Most companies are running the kernel with no OS around it. No clear roles. No guardrails. No supervision layer.

Here’s the underhyped part: while splashy autonomous agents grab headlines and fail publicly, quiet agentic workflows are genuinely transforming back-office operations. Invoice reconciliation. Contract clause extraction. IT ticket routing. These aren’t exciting demos. They’re boring, bounded, and they work. The companies getting real value from agents aren’t the ones chasing autonomy. They’re the ones who found the right readiness foundation and picked problems where AI proposes and humans approve.

2. The DeepSeek Earthquake (And What Silicon Valley Missed)

If agentic AI challenged our assumptions about deployment, the next trend challenged our assumptions about who builds the models in the first place.

January 27, 2025. Nvidia’s stock drops 18% in a single day. The trigger? A Chinese company most Americans had never heard of.

DeepSeek released an AI model matching GPT-4’s capabilities. Open source. Free. And here’s the kicker: they claim they trained it for around $6 million. GPT-4 reportedly cost $100 million.

Now, I’ll be careful here: that $6 million figure is DeepSeek’s own claim, not independently verified. But even if it’s off by a factor of five, the implications are staggering.

The US had export controls. China was supposed to be years behind on advanced chips. The whole strategy was: control the hardware, control AI leadership. Turns out you can’t export-control ideas.

The Japan Parallel

This reminds me of the 1970s. American car companies building bigger, more powerful, more expensive vehicles. Japanese manufacturers focused on efficiency and reliability. Detroit dismissed them as “cheap imports.” By 1980, Japan was the world’s largest auto producer.

DeepSeek followed the classic disruption playbook. When you can’t beat incumbents at their own game, change the rules. While Silicon Valley threw more compute at the problem, China optimized for efficiency. Open source became their weapon.

OpenAI’s own CEO said as much, in remarks reported by CNBC in August 2025:

“It was clear that if we didn’t do it, the world was gonna head to be mostly built on Chinese open source models.”

Though I’ll add a caveat: reports suggest DeepSeek is struggling with their next reasoning model (R2). Chip restrictions may bite harder than expected. The “China will dominate AI” narrative might be premature.

3. Small Language Models: When Less Becomes More

DeepSeek proved you don’t need the biggest budget to build competitive AI. Small Language Models take that logic even further: you don’t always need the biggest model, either.

Quick question. When you ask ChatGPT about your company’s refund policy, do you need a model that also knows Shakespeare’s sonnets, quantum mechanics, and 1980s pop lyrics?

You’re renting a nuclear reactor to power a nightlight.

This is why Small Language Models are having a moment. We’re talking 1 to 10 billion parameters versus the trillion-plus in frontier models. Microsoft’s Phi-3 outperforms models twice its size on specific tasks. Google’s Gemma 3n runs multimodal AI (text, images, video, audio) directly on your phone.

Why This Matters Beyond Cost

SLMs run where big models can’t. Factory floors. Remote field sites. Your pocket. No internet required. Data stays local, which makes your CISO sleep better at night.

• A field technician photographs a broken part and gets diagnostic help without cell service.
• Warehouse workers update inventory by voice while their hands are full.
• Healthcare workers access patient information in rural clinics with spotty connectivity.

4. The Great AI Reckoning: Show Me the ROI

Smaller models, cheaper training, better edge deployment: none of it matters if the business case doesn’t hold up. And that brings us to the elephant in every boardroom.

Enterprises have poured hundreds of billions into AI. The returns? Mostly demos and pilots. S&P Global found that 42% of enterprises abandoned most of their AI initiatives in 2025, up from 17% just a year earlier. MIT put a sharper number on it: 95% of enterprise generative AI pilots deliver no measurable impact on the bottom line. “We’re still experimenting” only works as an excuse for so long. Boards are asking harder questions. CFOs want numbers, not narratives.

As Stanford HAI put it in its 2026 predictions:

“The question is no longer ‘Can AI do this?’ but ‘How well, at what cost, and for whom?’”

What Actually Separates Winners from Losers

If you’re trying to figure out which AI investments to prioritize, the AI Use Case Canvas provides the structured decision framework I use with my advisory clients. For a deeper dive into how to actually measure whether those investments are paying off, see my AI ROI measurement framework.

5. The EU AI Act Arrives: Compliance as Strategy

As CFOs tighten the screws on AI ROI, regulators are tightening them from the other side. Mark your calendar: December 2, 2027. That’s when the EU AI Act’s high-risk requirements for standalone (Annex III) systems take full effect, deferred from August 2026 under the Digital Omnibus. The Omnibus has cleared both the European Parliament and the Council as of June 2026, but publication in the EU Official Journal is still pending, so treat the new date as very likely rather than finalized.

Using AI for credit scoring? Hiring decisions? Healthcare? You’ll need pre-market assessments, comprehensive documentation, human oversight mechanisms. Penalties for getting the high-risk obligations wrong: up to €15 million or 3% of global revenue. (The steeper €35 million, or 7% tier, is reserved for prohibited practices, like social scoring or manipulative AI systems: a different and more severe category. Mixing the two up is an easy, expensive mistake.)

Vision Compliance’s 2026 EU AI Act Readiness Report found 78% of the enterprises it assessed had not taken meaningful steps toward compliance: 83% had no formal inventory of the AI systems they use, 74% lacked a designated internal owner for AI compliance, and 61% had no process yet for generating the technical documentation high-risk systems will require.

Most enterprises aren’t ready. That’s an opportunity as much as a threat. Companies building compliance into their AI systems now will have structural advantages. The ones who wait will spend years playing catch-up.

Governance becomes the new competitive moat. Companies with transparent, auditable AI will win enterprise deals over competitors who can’t explain how their systems work. Like GDPR, the EU AI Act will become a de facto global standard. If you haven’t started, the Minimum Viable AI Governance (MVG) framework is the fastest path to a defensible baseline.

6. AI for Science: Where the Real Magic Is Happening

I’ve spent five sections on what’s hard and what’s broken. Let me end with what gives me genuine hope.

While everyone debates chatbots, something genuinely transformative is happening in research labs. Quietly. Without the hype.

Google DeepMind’s AI co-scientist proposed drug candidates for liver fibrosis that were validated through actual lab experiments. In a separate case, it predicted an antimicrobial resistance mechanism that matched experimental results before those experiments were even published. Hypothesis development that used to take years, compressed to days.

As Peter Lee, president of Microsoft Research, put it:

“AI will generate hypotheses, use tools and apps that control scientific experiments, and collaborate with both human and AI research colleagues.”

Where I Might Be Completely Wrong

Any honest forecaster owes you transparency about uncertainty. So here’s where my predictions could fall apart:

• AGI could surprise us. Stanford experts predict no AGI in 2026. I agree. But AI capabilities have consistently surprised everyone. A breakthrough in reasoning or memory could change everything overnight.
• The failure rate might be temporary. Maybe most AI projects fail because we’re early, not because of structural problems. Perhaps 2026 is when playbooks mature and success rates surge.
• China’s lead might stall. DeepSeek is impressive, but they’re reportedly struggling with R2. Chip restrictions may hurt more than expected.
• Energy constraints could bite. AI’s power consumption is growing faster than its efficiency. At some point, physics wins. Unless nuclear or renewable breakthroughs change the equation entirely.

Honestly? We’re all navigating without a map. Anyone claiming certainty is selling something.

The Bottom Line

I started with Klarna for a reason. It’s not a cautionary tale. It’s a success story. They tried something bold, learned it didn’t work as expected, and had the courage to publicly admit it and change course. That’s exactly what the AI industry needs more of.

Three Challenges for You in 2026

1. Stop chasing the hype cycle. When everyone zigs toward the latest demo, zag toward fundamentals.
2. Question everything. When vendors promise transformation, ask for case studies.
3. Invest in your people. The most important AI capability isn’t the model. It’s human judgment.

The AI revolution is real. It’s just slower, messier, and more human than the headlines suggest. And if that disappoints you, you were probably building on hype, not strategy. The rest of us? We’ve got work to do.

AI Strategy Frameworks to Act on These Trends

Ready to move from forecast to action? These frameworks on AskAjay.ai connect directly to the themes above. For governance readiness ahead of the EU AI Act, start with the Minimum Viable AI Governance framework and the Governance Playbook. To evaluate specific AI use cases against these trends, use the AI Use Case Canvas.

For the agentic AI and ROI themes, the 5-Pillar AI Readiness Assessment provides the diagnostic to evaluate organizational preparedness. And for founders navigating this landscape, the Founder’s Playbook for Responsible AI offers the ten principles that separate sustainable ventures from those chasing hype.