2025 and the Next Chapter(s) of AI

In many ways, the current state of AI feels like living in the space between what we can imagine and what tools we have available to us, at work and in our personal lives, to make those dreams a reality.

But the gap is closing quickly.

Remember in 2023 when we thought only humans could write software, design games, draft marketing content, create a video ad, resolve a customer service issue, or summarize a set of documents? Pick nearly any consumer or enterprise scenario, and it’s a pretty safe bet that AI is already playing a role in making it more efficient, improving quality, or completely redefining how it gets done.

We’re only two years into the commercialization of generative AI, but it’s clear these technologies and capabilities will eventually form the frontend and possibly even the backend of nearly every application.

As a CTO, I share predictions every year, knowing that I might be wrong, understate (or overstate) many things, and will likely get lucky a few times and even nail a couple. I do this — not because I believe I’m correct — but as a necessary exercise to share what I’ve learned from the actual implementations I work on with our customers and teams. My hope is that some of these reflections will spur your own creativity, skepticism, and thoughtfulness around AI.

Last year, I shared a few predictions on how gen AI adoption would increase business utility and ultimately drive innovation. My thinking was that ​​organizations should focus on sustainable costs, broad access, and trust and security to get gen AI right in 2024. These were largely based on the fact that many companies last year were developing the foundations for scaled experimentation, rigorous evaluation, and the constant refinement of AI. Nothing stops a promising project faster than runaway costs, siloed efforts, and a lack of trust in what’s being built.

I’m happy to report that we saw a surge of companies moving their AI prototypes into production — a significant step, demonstrating the growing confidence in AI capabilities and their potential to deliver tangible value. I encourage you to check out this list of over 300 real-world examples of AI in action to inspire your own efforts.

As of January 2025, it’s still us humans writing the prompts, defining the reasoning flows, putting guardrails in place to manage agentic action, and supplying the policies and KPIs that will determine the success or failure of our AI projects. The more we can think and reason up front, the better we’ll design the requisite AI levers.

Against this backdrop, let’s get to what I see as the four key trends emerging that will shape how we collectively move forward this year.

1. Multimodal AI as the new standard

Multimodal AI, which integrates diverse data sources like images, video, code, and audio, alongside text, will become increasingly prevalent. This will enable organizations to provide more sophisticated and personalized customer experiences: Imagine searching for information using a combination of text, images, and voice commands. Or, interacting with AI-powered chatbots that can understand and respond to your visual cues, or accurately triage your health concerns based on shared audio, video, or images and immediately provide a personalized medical analysis. This is the power of multimodal AI and the new expectation for state-of-the-art models.

One of my favorite stories from last year was how the world’s largest advertising holding company, WPP, expanded its WPP Open operating system by leveraging the native multimodality of Gemini. These capabilities empower creatives to take an idea expressed via voice, an image, or a web link and generate social media ad copy that includes draft images and video clips, in minutes.

Another standout example is how Mercedes-Benz is implementing Automotive AI Agent into its MBUX Virtual Assistant to create a highly personalized multimodal experience for drivers and passengers alike. Riders will be able to use voice commands like, “Is there a good restaurant nearby?” "Does it have good reviews?" "Who is the chef?" and “Can you direct me there?” This is just the beginning of how multimodality and agentic capabilities can transform industries.

2. Agentic platforms for scale

AI agents emerged in 2024 as an abstraction for the grounding, reasoning, and augmentation tasks necessary to convert models into value. As organizations gain more experience with combining AI tools with their own intellectual property, data, and expertise, they will want a way to scale the experimentation and deployment of their AI agents. This will generally follow a pattern of discovery, connection, and automation, with agents acting as the critical bridge between the promise of AI in workflows and the realization of that value.

One of the first to put agentic platforms in action is Banco BV, one of Brazil’s largest private banks. Banco BV is using Google Agentspace — which brings together Gemini’s advanced reasoning, Google-quality search, and enterprise data — to enable its employees to discover, connect, and automate with AI agents across its broad set of data and critical systems, in both a secure and compliant manner.

At Deloitte, knowledge workers utilize Agentspace to bridge data sources quickly, fostering rapid experimentation and collaboration. In one case, NotebookLM, available as an out-of-the-box agent in Agentspace, even found a connection between topics across uploaded reports that Deloitte employees hadn’t caught themselves, which would have been difficult to spot under traditional silos of analysis.

3. Optimization of the AI stack

2025 will be the year of optimization. Companies will begin to shift their focus from simply experimenting with or implementing AI to optimizing its performance and maximizing its value. More than 70 percent of organizations are already seeing return on investment (ROI) from gen AI — and that number will only continue to rise as more companies move from production to optimization.

This increased focus on optimization reflects a deeper level of understanding of AI and a growing emphasis on extracting maximum value from these technologies. While optimization will continue at the hardware level, organizations will also move up the technology stack with emergent intelligence that selects the right model for a given user query across a number of attributes including cost, quality, and other important business value metrics. For instance, using a combination of our TPUs and GPUs, LG AI Research was able to reduce inference processing time for its multimodal model by more than 50% and operating costs by 72%.

For organizations to make the most of their AI investments in the future, they will need to invest in identifying the best AI models for their specific use cases, optimizing infrastructure for training and inference, and ensuring they have the ability to measure and optimize models for long-term relevance and effectiveness.

4. Silo busting

The rise of gen AI is helping break down the walls between departments and democratize access to AI tools. This new paradigm is empowering a wider range of users within organizations to participate in AI-driven innovation, fostering collaboration and accelerating the creation of novel customer experiences like never before.

As AI technologies and tools become more widely-adopted, they free up time previously dedicated to routine tasks, allowing individuals to focus on more creative and strategic endeavors. This increased capacity for creativity will most certainly drive innovation and lead to unexpected and unimagined breakthroughs.

To help enable that innovation and breakthrough, Workspace Business and Enterprise customers will now have the best of Google AI embedded directly into the tools they use every day. With Gemini for Google Workspace, teams around the world will be able to work faster and more efficiently right where they’re already spending their time — in Gmail, Docs, Sheets, Meet, Chat, Vids, and so much more.

Beyond 2025

This year will be a big one for AI and for us as humans; perhaps, the biggest to date. Aside from the instances I’ve already mentioned, I expect we’ll see AI used to address some of the world’s most pressing problems in ways we can’t even imagine.

Already, my colleagues at Google Deepmind shared how they are using AlphaFold to predict the structure and interactions of all of life’s molecules, which has the potential to transform our understanding of the biological world and drug discovery. Another incredible source of inspiration is the work our partners at the Asteroid Institute are doing with Google’s AI technologies to take what would be 130 years of research down to just three months, bringing space that much closer to our fingertips. In 2025, I’m also excited to see what AI will do for students around the world, with the opportunity to make education more personalized and accessible, bringing new possibilities to uplift an entire generation.

On a more personal note, this year will also mark my ten-year anniversary at Google. Back in 2015, it would have been impossible to predict the nature and extent of the AI disruption we are experiencing today. Still, the safest prediction with the highest ROI now, as it was back then, is to surround yourself with people who are curious, humble, and action-oriented. If you do that, you’ll always be able to navigate ambiguity and complexity successfully and satisfyingly — regardless of your industry, role, or the technology at hand.

Here’s to all our customers and Googlers building a future that we’ll sit in awe of a decade from now: Thanks for the last 10 years, and on to the next 10!