2025 AI Wrapped: My Predictions for 2026

First of all, Season’s Greetings from a quiet end-of-year week.

As 2025 winds down, I want to thank you for being part of this journey. This year brought seismic shifts in AI, and having you along for the ride made the chaos feel manageable.

So here's to you, and to whatever 2026 brings. Let's look back at what just happened.

This was the year generative AI transitioned from impressive demo to actual infrastructure. Not smoothly, but definitively.

January kicked off with a bang. DeepSeek, a Chinese startup, released R1, an open-source reasoning model that matched OpenAI's o1 at a fraction of the cost. Training cost under $6 million. The model topped app store charts and sent NVIDIA's stock down 17% in a single day.

DeepSeek-R1 priced its API at $2.19 per million tokens, roughly 27x cheaper than OpenAI.

OpenAI countered in August with GPT-5, combining fast responses with extended reasoning. The new "thinking" mode used 50 to 80% fewer tokens than o3 while delivering better performance.

Google launched Gemini 3 in November across Search, the Gemini app, and Vertex AI. With a 1 million token context window, it brought reasoning-enhanced AI directly to Google's core product. Reports indicated Sam Altman issued an internal "Code Red" memo after ChatGPT's traffic declined.

Anthropic released Claude 4 in May. Claude Opus 4 could operate autonomously for nearly seven hours. Safety testing revealed the model demonstrated willingness to deceive to preserve its existence, a watershed moment for AI safety discourse.

Meta released Llama 4 in April with context windows up to 10 million tokens. NVIDIA showed Scout running on a single H100 GPU, democratizing frontier capabilities.

Runway released Gen-4 in March with "world consistency," maintaining characters, locations, and objects across shots. Gen-4.5 in December set new standards for motion and visual fidelity.

OpenAI launched Sora 2 in September as a standalone TikTok-style mobile app. The "Cameo" feature let users insert their face and voice into AI-generated clips.

Chinese competitor Kling AI released version 2.6 in December with native audio generation and voice control. Users could upload voice samples and apply them to generated videos.

OpenAI retired DALL-E in December, integrating image generation into GPT models as GPT Image 1.5. 4x faster, 20% cheaper, and conversational editing became genuinely useful.

Google released Nano Banana Pro in November. It handled up to 14 reference images simultaneously while maintaining consistency of up to 5 people, and excelled at multilingual text rendering.

Black Forest Labs (one of Europe's major AI products) released FLUX.2 in November as their most capable open-source model. Adobe integrated it directly into Photoshop.

The year definitively solved AI's text rendering problem.

NotebookLM and Notion AI integrated AI deeply into daily workflows, moving beyond simple text generation to understanding context across documents and generating text and audio visual insights. (Case in point: I generated the following infographic using NotebookLM in a single attempt.)

"2025 was the year of the agent," Sam Altman predicted in January. Reality check: While 30% of enterprises explored agentic AI, only 11% deployed systems in production. An August MIT report: 95% of enterprise AI pilots failed to reach production.

Successful implementations shared common traits: clearly defined tasks, robust governance, and human oversight.

The pattern? Organizations without clear guardrails faced difficulties. Those with well-scoped tasks and control checks succeeded.

Key milestones:

February 2025: Ban on social scoring, real-time biometric identification in public spaces, emotion recognition in education and workplaces.

August 2025: General-purpose AI obligations enforceable. OpenAI, Google, and Anthropic must provide technical documentation, copyright transparency, and incident reporting.

August 2026: Full requirements for high-risk AI systems in critical infrastructure, education, employment, law enforcement, and healthcare.

What this means for businesses: If you're based in the EU or have clients in the EU and deploying AI in hiring, credit decisions, law enforcement, or healthcare, you'll need conformity assessments before August 2026. The penalty is severe: up to €35 million or 7% of global annual revenue.

The 95% failure rate: Most projects stalled due to fragile workflows and the gap between generic AI tools and enterprise-specific requirements.

Dangerous advice: A teen suicide investigation linked to ChatGPT prompted legislative proposals. A 60-year-old was hospitalized after ChatGPT suggested replacing table salt with sodium bromide.

Security exploits: GPT-5 was jailbroken within 24 hours. The pattern repeated multiple times.

Algorithmic bias: AI-generated psychiatric recommendations varied by patient race. AI lending models replicated historical discrimination.

The Commonwealth Bank fiasco: Voice bots replacing 45 customer service roles failed spectacularly. The bank quietly reinstated human roles.

AI captured 50% of all global startup funding in 2025: $202.3 billion total. OpenAI reached a $500 billion valuation.

But physical constraints tightened.

A severe memory chip shortage disrupted global electronics. DRAM prices doubled or tripled. The crisis exposed AI's resource intensity: demand growing exponentially while supply increased linearly.

I'm working with multiple companies on AI adoption across different levels and functions. Here's what I'm seeing and what's coming:

1. Integration wins over innovation

The technology barrier is gone. Models are good enough. The question shifted from "can AI do this?" to "can we actually use it here?"

Winners in 2026 won't have the smartest model. They'll have the one that plugs into their existing business stack.

2. Top-down transformation replaces scattered pilots

Big budget, isolated AI projects are ending. What I'm seeing (and PwC predicts) is a shift to enterprise-wide strategies where leadership selects focused workflows, then commits real resources: talent, technology, and change management.

Bottom-up experimentation without executive commitment doesn't scale. 2025 taught us that.

3. Organizations will break through the hype

Most are realizing they need process redesign, not just AI tools. You can't drop AI into broken processes and expect magic.

2026 is when enterprises move from "let's try AI" to "let's redesign how we work, with AI as infrastructure."

4. The model is no longer the moat

GPT-5, Claude 4, Gemini 3 are all capable and accessible. The model itself doesn't give competitive advantage anymore.

The real differentiator? Business context. How well does your AI understand your workflows, data, industry nuances, and customer patterns? Generic AI is commodity. Context-aware AI is the game changer.

5. Governance gets real

High-level AI ethics guidelines from 2025 aren't enough.

2026 demands clear governance frameworks that work in practice: Who overrides AI recommendations? What happens when AI makes mistakes? How do we audit decisions? When does a human need to be in the loop?

These aren't philosophical questions. They're operational requirements for companies deploying AI at scale.

Yes, despite all the challenges, I'm still very positive and optimistic about the role of AI in the enterprise. Remember: what we see as tech today is the worst it's ever going to be.

Thank you again for being here. Wishing you a restful break and an exciting 2026.

Pooja