What’s in Store for AI in 2026

Sometimes the hustle and bustle of New York gets to me; its endless traffic, endless spores of people spilling on to the sidewalks, endless things to do and places to go. So the past few days, I’ve been spending time in New Hampshire, a much quieter part of the US. So much has happened in the world of AI and yet, as I look out my window to the quiet falling of snow onto classic New England homes, it feels like nothing much had changed. Most of us are still going to work next Monday, after an extended holiday, opening our laptops and checking through our emails.

And yet, so much has changed.

The way we work is faster. We now have a thinking partners available 24 hours a day. Prompts become designs that used to cost us thousands of dollars. We can hand off entire workflows to agents that keep running while we pour our coffee. We can do more with less.

Although the surface looks the same, the underneath has shifted entirely.

As 2025 and its big headlines comes to an end (what I wrote in a previous post), here are some predictions I have for the upcoming year. These large trends are what I think we will hear a lot more about in the coming year.

1.Physical AI becomes ubiquitous

From robots being deployed at China’s boundaries to K pop stars using them for back up dancers, 2026 might be the year we normalize the physical robot. The movies of Megan and Ex Machina we once thought as science fiction slowly becomes reality as physical AI actually becomes more useful. Sure, I am confident I will still load the dishwasher this coming year, (not my husband) or a robot of some sort, but as robots first find their place in factories, they slowly will seep into our homes too.

2. AI’s breakthroughs in science

Some of you might have seen me present this slide in my keynotes. As we move forward into 2026, what used to be a question mark year on Level 4 might slowly have a number become uncovered.

Chemistry saw an equally dramatic shift this year as AI systems began designing entirely new molecules that human chemists might never have imagined. In 2025, several AI-discovered compounds entered early clinical trials, marking a turning point where machines were no longer assisting chemists but actively expanding the boundaries of chemistry itself. In the area of Materials Science, AI models, trained on decades of experimental data identified new materials for batteries, semiconductors, and solar cells with properties previously thought incompatible.

Will the next Nobel Prize in Science go to an AI? Probably not, but as they become our thinking partners and aid scientists in scientific breakthroughs, how much credit would we give them when the next invention comes primarily from AI?

3. Building for AI, not the human

As AI agents start to crawl the web, shop on behalf of us, and work on behalf of us, we will realize that our current internet is not built for easy navigation for AI. If you look at the fumbles of Computer Use agents from Anthropic, Operator from OpenAI, it’s easy to see that the world wide web we created – from the buttons to Captcha’s to the cursor – were all made for humans to use. AI systems don’t have “eyes” like we do. They’re not exactly “reading” a page. All those things are archaic designs of the past that will need to change when the web has far more traffic from bots than humans.

AI-friendly software needs to have:

·       Clean, structured data that machines can read and act on

·       APIs and permissions that let AI take actions

·       Clear workflows so agents understand intent and context

·       Audit trails and controls so humans stay in charge

So how will we adjust the web? How will we identify them and make sure it’s safe? 

The next version of the internet might be far less visual and far more semantic. Structured data, APIs, machine readable intent layers, and standardized schemas will matter more than layout, color, or typography. Websites will increasingly have two faces: one for humans and one for machines. The human side will remain expressive and emotional. The machine side will be clean, explicit, and transactional.

Identity will also need to evolve. Today, we prove we are human by clicking boxes, identifying traffic lights, or dragging sliders. That logic breaks down in a world where bots are not inherently malicious and humans are no longer the primary users. Instead of proving humanity, the web will need to verify authorization. Is this agent allowed to act on behalf of a real person. What permissions does it have? What limits are enforced? Authentication will shift from human detection to agent credentialing, with cryptographic identity, signed intents, and auditable action trails.

Safety will follow the same path. Rather than blocking bots, systems will learn to negotiate with them. I wrote about the internet’s new toll booth in July 2025 here.

4. Regulation splits the AI world into speed zones

I mentioned the EU AI Act in the previous article where Europe has operated under an AI Lite version, while places like US has pushed full speed ahead, choosing the world wrought with misinformation and AI slop, where it’s hard to determine truly what is real or what is fake. Europe will seem more like the “old days” (it already has) as the US starts to seed AI in classrooms, replacing teachers, media, and work, often before norms are set. Deepfakes and AI mishaps will likely occur much more often in the US than Europe.

And then there’s China. China is interesting because it seems to be doing a bit of both. It’s both releasing new products quickly but also coming down with heavy regulation. Deployment is centralized, practical, and aligned with national goals. Once approved, systems scale fast and embed deeply into daily life. Deepfakes, synthetic voices, and generated media are regulated aggressively. Platforms are required to watermark AI content, register models, and ensure traceability. That does not eliminate misinformation, but it drastically reduces the kind of open slop and viral chaos seen in the US.

In car analogy, it’s kind of like Europe is driving a Volkswagen, the US is in a Ferrari and China is driving a fleet of autonomous trucks.

5. Open-Source and Specialized Models Surge in 2026

The era of adding more compute and data to build ever-larger foundation models is fading. Up until around 2025, the dominant strategy was pure scaling: making models massively larger by throwing more compute power, parameters and data at them. This followed "scaling laws" where bigger often meant dramatically better. But by late 2025, evidence mounted that this approach was hitting diminishing returns: huge investments in scale were yielding smaller and smaller performance gains. An insane amount of money will only make models slightly better. This opens the channel for open source and Small Language Models like Phi and Gemma where these lightweight variants run on laptops and edges.

6. Cybersecurity evolves to protect from within

Cybersecurity will be more important than ever as it becomes way too easy to clone a voice or recreate an image.

Traditionally, cybersecurity is a field where we’re protecting an organization from outsiders: scammers, phishing attacks, DDOS. But in the future, we might be protecting our organization from within: own delegated intelligence. An agent with access to email, payments, databases, or internal tools does not need to be malicious to cause harm. It only needs to be misaligned, confused, or subtly manipulated. A single injected instruction can cascade across workflows faster than a human could ever intervene.

Every action taken by an AI must be attributable, reversible, and bounded. You’ll hear about this term called audit trails a lot more. Enterprises demand to know not only what an agent did, but why it believed it was allowed to do so. They are given least privilege access which I wrote about here.

7. AI becomes the hotspot for commerce

In 2026, consumers will interact more with AI systems to find what they need. It won’t become the primary interface for commerce, but we can expect more traffic to shift over. Search, ads, and shopping gradually collapse into a single AI mediated experience. OpenAI has tested ads in 2025 with a full rollout in 2026. This reshapes retail, travel, financial products, and local services. Companies realize that being visible to AI agents matters more than being visible to humans. Answer Engine Optimization overtakes traditional SEO as the dominant discovery strategy.

As marketers, we used to learn about the marketing funnel. While that might hold true for human decision-makers, when the prospective customer is an AI agent, discovery, consideration, and purchase collapse into a single moment of delegation. The user expresses intent once, and the agent handles the rest. What to buy. Where to book. Which option fits preferences, constraints, and history. The brand is no longer chosen by a human scanning a page, but by an AI ranking relevance, reliability, and machine readable trust.

This forces a reordering of power. Retailers and service providers stop optimizing for clicks and start optimizing for comprehension. Pricing clarity, inventory accuracy, fulfillment reliability, and policy transparency become ranking signals. If an agent cannot confidently transact with you, it will quietly route demand elsewhere.

Conclusion: Humans move upstream to define intent, taste and judgement

Yes, AI still make mistakes. Scaling is reaching diminishing returns. The new versions of the models we are using seem less and less impressive than before. That’s something interesting about change. Sometimes it seems gradual. We’re like frogs in warm water. When the temperature gradually increases, it’s hard to feel the difference. As humans move upstream in the decision-making process, we will need to define intent, taste and judgement. Until then, let us quietly enjoy the last few days of this East Coast winter storm.

--

Hello! I’m Sharon Gai, a keynote speaker on AI and its effects on workers and the future of work. To book me to provide an update on the state of AI for your company in the format of a virtual session or in person session, click here.