When AI meets Product: March’25 AI Product Updates

Keeping up to date with new AI models, products, ethics, and trends

Welcome to the March edition of “When AI Meets Product — AI Product Updates”. This month has been packed with major launches and deep conversations about how AI is applied, scaled, and regulated. Here’s what stood out:

• GenAI Model Providers Updates — We saw major launches and announcements from leading model providers (Google, OpenAI, Mistral, Deepseek….) but also a clear trend towards launching more specialized agents and applied AI features from the same model providers side.
• AI Product & Agent Updates — Agents continue to dominate the roadmap: OpenAI is launching paid agents, Microsoft is building for clinical workflows, and no-code platforms like Cursor are booming… Meanwhile, voice is making a comeback as a key interface for AI, and there is big potential for more personalized experiences.
• AI Ethics — AI labeling, copyright, and research exploring the relationship between responsible AI and business.
• Other Resources —On the impact of work, collaboration with AI and my notes from deeplearning.ai’s AI Dev 25!

Let’s get started! 🚀

Model providers updates

March has been another busy month in the world of foundational models, with major launches and announcements from leading model providers. We’re seeing not only the release of new model versions but also a clear trend towards more specialized agents and applied AI features from the same model providers.

Model providers — new model versions

• Google released new versions of both its open-source and proprietary models. The open-source Gemma 3 has been described as “the most capable model you can run on a single GPU or TPU”. On the closed-source side, Gemini 2.5 was launched, labeled by Google as their “most intelligent” model yet. It has shown impressive performance on complex reasoning and coding tasks and has quickly taken the lead in several AI benchmarks.
• OpenAI made headlines with the release of GPT-4o image generation, a “natively multimodal model capable of producing highly accurate, photorealistic visuals”. Social media quickly filled with impressive AI-generated images and memes in Studio Ghibli style. In a more research-focused update, OpenAI and MIT Media Lab published a study on ChatGPT’s emotional impact, with findings such as emotional engagement being rare in real-world usage or voice mode having mixed effects on wellbeing.
• Microsoft introduced Phi-4-mini and Phi-4-multimodal, the latest in their Phi family of small language models.
• Mistral launched Mistral OCR, a new Optical Character Recognition API that goes beyond basic text extraction by understanding complex document elements like tables, equations, and embedded media. They also released Mistral Small 3.1, a compact yet powerful model aimed at competing with Gemma 3 and GPT-4o mini, offering improved text generation, multimodal capabilities, and a context window of up to 128k tokens.
• DeepSeek, China’s most promising LLM initiatives, launched DeepSeek V3. This new version brings substantial improvements in benchmark performance, front-end web development tasks, and Chinese language proficiency.
• Cohere introduced Aya, a new multilingual model that is part of a broader initiative to expand language coverage in AI.

With so many model updates, it’s a great time to check in on the current leaderboard. As of now, Gemini 2.5 leads most performance benchmarks, closely followed by Meta’s LLaMA 4, and GPT-4o.

Model providers — features and applications

• Anthropic introduced its own web search feature, integrating real-time internet access into Claude. This positions Claude to be more competitive with tools like ChatGPT’s browsing or Perplexity.
• AWS launched multi-agent collaboration for Bedrock agents, allowing different AI agents to work together to solve complex tasks, with a centralized mechanism for planning, orchestration, and user interaction.
• DeepSeek has been especially active in the open source space, publishing multiple new tools and models in what they’ve called their “Open Source Week”.
• MCP — The New Standard for Tool Use? Anthropic introduced MCP (Model Connection Protocol), an open standard for connecting AI models with external tools, APIs, memory systems, and databases. The launched happened some months ago, but lately it feels like everybody is talking about it. It’s designed to solve common challenges like limited context, prompt engineering, and custom tool integrations. With MCP, models can maintain context over time and interact with external systems more smoothly. Even OpenAI is getting involved, showing early signs of support for this shared standard — suggesting MCP might become a key piece of AI infrastructure going forward.

AI Products, applications & agents

AI is rapidly shifting from models to products — and in March, that meant a huge interest on agents, no-code tools, voice interfaces, and hyper-personalized user experiences.

Agents Are Everywhere — But Still Hard to Define

From OpenAI to startups and enterprise use cases, everyone is building agents. But what exactly is an agent? As TechCrunch notes, there’s still no shared definition — and every company seems to approach agent architecture differently. Also, as many companies move from models to applications and agents, the idea of benchmarking model performance on the application and agentic layer is gaining traction. The Smol Agents Leaderboard on Hugging Face is one of the first benchmarks following this new trend, which deserves full attention for Applied AI practitionaries.

There were many interesting news related to agents:

• OpenAI plans to commercialize agents, charging between $2,000–$20,000/month for tailored enterprise-grade use cases.
• Butterfly Effect, a Chinese startup, launched a general-purpose agent, Manus. In their demo, they showcase examples of the use of the agent of resume screening, property research, or stock analysis.
• Norm AI raised $48M to build compliance agents that interpret and automate regulatory tasks.
• Microsoft introduced a clinical workflow agent, Microsoft Dragon Copilot, to streamline healthcare operations, “the first unified voice AI assistant that enables clinicians to streamline clinical documentation, surface information and automate tasks”.

Also worth a look: AAA Framework for thinking about agent design and architecture.

No-Code AI Is Booming

No-code tools and the concept of vibe coding (“coding where you fully give in to the vibes, embrace exponentials, and forget that the code even exists”) is booming. Many of these tools already allow to build websites, back ends, databases and integration from scratch through natural language interactions. This allows builders to build prototypes much faster, but also non-developers to build their first digital products. I recently tried out Replit and was amazed by its capabilities. There are many more similar products though, such as Cursor or Lovable (Europe’s fastest growing startup!). If you’re interested in the topic, make sure to check Lenny’s interview to Lovable’s CEO.

Ethan Mollick notes, however, AI is especially strong for early-stage tasks (like prototyping), but less so in complex, multi-repo environments — raising questions about the “minimum viable knowledge” to work effectively with AI.

Voice is back

Amazon’s new generative Alexa is taking voice assistants to the next level, promising more natural and smart interactions (while free with Prime). In the meantime, Apple hasn’t launched yet any big generative updates on Siri.

a16z deep dive sees great potential for business and users on AI Voice agents, and considers the tech layer to be sufficiently mature to allow good applications (with low latency, good performance, cost efficiency…). Also Andrew Ng shared a similar point of view at his letter What I’ve Learned Building Voice Applications — Best practices for building apps based on AI’s evolving voice-in.

On personalization

We’re seeing more personalized user experiences driven by AI, especially in recommendations and search:

• This post is a great summary on the future of LLM based recommendations, as a means to generating “personal narratives — stories that resonate with users”.
• Google’s Gemini personalization update is another push toward adaptive experiences, aiming for Gemini to access Google apps and search history, to deliver “contextually relevant responses that are adapted to your individual interests”.
• In parallel Google is also expanding AI overviews and launching AI mode, and improving marketing personalization through AI. Both launches are a good example on how users are changing their browsing behavior to “zero-click” searches with great impact to brands and SEO.

Other Highlights

• DeepSeek is rapidly expanding in China, powering a growing range of consumer and enterprise products (automotive, smartphones, home appliances, healthcare and more).
• McKinsey’s latest report explores how are companies adapting their organizations and processes in favor of AI.
• Stripe’s CEO surprised the world by showing some data on how AI startups are growing even faster than SaaS — and why calling them “wrappers” underestimates their long-term value.

AI Ethics and Legislation Updates

As AI adoption accelerates, March has brought renewed attention to the legal, ethical, and human implications of these technologies.

• Spain is taking a bold step by introducing heavy fines for not labeling AI-generated content — marking one of the first concrete enforcement measures in the EU AI Act era. This regulation targets transparency and aims to curb misinformation risks.
• Thomson Reuters won a copyright fair use case against a competitor that trained models on its content. This ruling could set a precedent for how courts interpret fair use in AI training contexts.
• A recent Harvard Business Review study argues that responsible AI isn’t just good ethics — it’s good business. Responsible AI features — specifically privacy and auditability — serve as powerful product differentiators that can generate significant economic returns. However, only 15% of these same managers feel well-prepared to adopt RAI practices.
• “Why Thinking Hurts After Using AI” explores how heavy reliance on AI can impact negatively our critical thinking and reduce cognitive engagement over time.

Other resources

As AI matures, the conversation is shifting from capabilities to collaboration, creativity, and who gets to build.

• AI as a teammate: A recent paper explores how AI changes team dynamics — not just as a tool, but a collaborator.
• The geography of generative AI’s workforce impacts will likely differ from those of previous technologies.

Last but not least, I attended deeplearning.ai AI Dev 25 conference! It was really inspiring and a great learning and networking opportunity, you can check here my summary post about it!

Wrapping it up

That was it from “When AI Meets Product — AI Product Updates”. 2025 is starting really strong with new AI models, agentic use cases and impact on the way we work and develop tech products. Stay tuned!