In March 2026, a Guardian essay about needing a living, breathing human over a metallic helpline echoed a broader truth: no matter how advanced AI becomes, people still want real human judgment in moments of friction. That tension threads through every major AI move this week, from multimodal embeddings to universal protocols that connect software and agents, reminding us that efficiency and empathy can coexist in enterprise AI.

At the heart of the technology shift is Gemini Embedding 2, Google’s latest embedding model that natively handles text, images, video, audio and documents in a single space. It promises lower latency and lower costs, enabling cross-modal search and retrieval without the usual transcription bottlenecks. In practice, it creates a unified semantic map where a poem and a landscape image can sit near each other when they share a concept, and where a user query can span a video clip, an audio phrase or a document inside the same vector space. The model uses a 3,072‑dimensional vector for each item and supports an 8,192 token context window, with Matryoshka Representation Learning enabling truncation to 768 or 1,536 dimensions for cost efficiency.

Enter Manufact and the Model Context Protocol, or MCP, with a bold claim: software products will be accessed by AI agents, not just people. The company is building open-source tools that let developers spin up MCP servers in six lines of code and a cloud platform that handles deployment and observability so teams can push an MCP-based app to production in under a minute. The MCP ecosystem is being described as the USB‑C for AI, offering standard connectors so agents can access any tool or data source through a single interface.

Beyond models, enterprise workflows are evolving. Anthropic has introduced Claude for Excel and PowerPoint with shared context across apps, enabling continuous sessions where a single prompt can pull data from a spreadsheet and render it in a pitch deck. Microsoft is pushing Copilot Cowork, intensifying the competition in the enterprise app arena. The RSAC Innovation Sandbox highlights a broader security-focused push, where new AI governance and SecOps tools are tested for real-world resilience.

As with every disruptive technology, there are limits and trade‑offs. Gemini Embedding 2, for instance, caps input by file—long assets must be fed in chunks rather than as a single file—and even within those limits, enterprises must re-embed catalogs to unlock cross-modal search. The pricing landscape is tiered too, with a free tier for experimentation and then per‑token charges for production workloads, with audio inputs priced higher due to native processing. Yet the potential is clear: a universal, multimodal map of information that can be queried with a single natural‑language request across formats.

Together with the wave of AI‑driven legal platforms, agent networks and publishing‑fraud protections, the industry seems to be moving toward a world where agents are the default interface for software. The week’s news—from a Guardian podcast debating boycotts of ChatGPT to debates about AI in legal and publishing domains—underscores the ongoing balance between human oversight and machine efficiency. This evolving landscape invites both caution and opportunity as enterprises rethink how they design, deploy and govern intelligent software in real-world workflows.

Sources

1. The AI assistant was offering me any help I needed. All I wanted was a living, breathing human
2. Google’s Gemini Embedding 2 arrives with native multimodal support to cut costs and speed up your enterprise data stack
3. Manufact raises USD6.3M as MCP becomes the ‘USB-C for AI’ powering ChatGPT and Claude apps
4. AI Legal Platform now valued at $5.5 Billion
5. Self-publish and be scammed: Jon’s tale of heartbreak highlights boom in fraudsters using AI to supercharge book swindles
6. Meta Acquires Moltbook, the AI Agent Social Network
7. ‘Happy (and safe) shooting!’: chatbots helped researchers plot deadly attacks
8. Amazon is determined to use AI for everything – even when it slows down work
9. Wednesday briefing: From missing billions to nonexistent datacentres, inside Britain’s AI drive
10. Anthropic gives Claude shared context across Microsoft Excel and PowerPoint, enabling reusable workflows in multiple applications
11. RSAC’s Innovation Sandbox is where cybersecurity’s next giants are born
12. Should we be boycotting ChatGPT? – podcast