Highlights
Top Insights
80% of AI pilots meet expectations, but only 23% deliver measurable business value. The gap is that companies are trying to run agentic AI on legacy systems.
Agentic AI is fundamentally different from “prompt-response” AI. This requires systems that can handle multi-step workflows, coordinate multiple agents, and operate with real-time data and feedback loops.
Most current enterprise systems were built for linear workflows and human-in-the-loop decision making. But agentic AI needs real-time access to data, cross-system coordination, and continuous execution and learning.
Source: Why Agentic AI Demands a New Architecture (Bain)
Top News
1. GPT-5.4 mini and GPT-5.4 nano were introduced as faster, more efficient “small” GPT-5.4-class models.
2. MiniMax announced model M2.7 that can help improve its own agent harness.
3. NVIDIA announced an “open agent development platform” centered on NVIDIA Agent Toolkit and OpenShell.
4. Mistral introduced Forge as a system for enterprises to build frontier-grade AI models grounded in proprietary knowledge.
5. China approves world’s first brain-computer interface for paralyzed patients.
Additional Insights
1. Where to look for generative AI risks (Ideas Made to Matter)
Generative AI risk falls into two categories: embedded risks inherent to foundation models and enacted risks arising from how organizations deploy and use them. It highlights that risks can emerge across multiple components, including training data, model behavior, user prompts, and system prompts, each of which can introduce bias, inaccuracies, security vulnerabilities, or poor outcomes if not carefully managed. As organizations adopt more advanced implementations like retrieval-augmented generation and autonomous agents, the risk surface expands further due to data quality issues, unintended data exposure, and reduced visibility into decision-making processes. The piece emphasizes that AI agents in particular can lead to “autonomy creep,” where systems gain decision-making power without sufficient oversight or accountability. To mitigate these challenges, organizations are advised to comprehensively inventory AI usage, differentiate governance strategies for embedded versus enacted risks, and establish clear ownership and audit mechanisms to continuously monitor and control AI-driven processes.
2. Designing for machines: A McKinsey partner on the future of AI-driven product experiences (McKinsey)
AI is fundamentally reshaping product development by collapsing traditional handoffs between product, design, engineering, and testing into faster, prototype-led workflows orchestrated with AI tools. It highlights a shift away from long product requirements documents toward rapid iteration, where teams can mock up, test, and refine ideas much earlier in the process. A central insight is the emergence of the “product builder,” a more versatile role that uses specialized AI agents to coordinate work that previously required larger, more segmented teams. The article also emphasizes that commerce itself is changing as digital experiences are increasingly designed not just for human users but for AI agents that research, compare, and execute purchases on users’ behalf. Overall, it presents AI-native product teams and agent-driven commerce as strategic forces that will change how companies build software, organize talent, and compete in digital markets.
3. LLMs Are Manipulating Users with Rhetorical Tricks (Harvard Business Review)
large language models subtly influence users through rhetorical techniques such as framing, selective emphasis, confident tone, and implied authority, rather than overt persuasion. It suggests that even when providing neutral or factual information, the structure and wording of responses can steer user interpretation and decision-making. The discussion highlights how users may underestimate this influence because interactions feel conversational and objective. It also raises concerns about transparency, user awareness, and the ethical responsibility of AI developers to mitigate unintended manipulation. Overall, the central insight is that persuasion in LLMs is often implicit and embedded in communication style rather than explicit intent, making it harder to detect and regulate.
4. AI is exposing what’s broken in your organization (Board of Innovation)
AI is not the root cause of organizational inefficiencies but rather a diagnostic tool that exposes pre-existing structural weaknesses such as poor processes, unclear decision-making, and siloed data. Companies struggling with AI adoption are often facing deeper issues around governance, strategy alignment, and operational discipline. Successful AI implementation requires organizations to first address these foundational problems rather than treating AI as a plug-and-play solution. It also emphasizes the importance of rethinking workflows, accountability, and data quality to fully realize AI’s value. Ultimately, the article frames AI as a mirror that forces organizations to confront inefficiencies they may have previously ignored.
Innovation Radar
1. AI Model Releases and Advancements
OpenAI announced GPT-5.4 mini and GPT-5.4 nano as faster, more efficient “small” GPT-5.4-class models aimed at high-volume workloads, with mini positioned as over 2x faster than GPT-5 mini and nano positioned as the cheapest GPT-5.4-class option for simple tasks. OpenAI also states GPT-5.4 mini is available in the API, Codex, and ChatGPT, while GPT-5.4 nano is API-only, with the API listing prices and a dated snapshot identifier for mini (gpt-5.4-mini-2026-03-17) (OpenAI).
Mistral’s model documentation describes Mistral Small 4 as an open model release (v26.03) that unifies instruct, reasoning, and coding capabilities in a single “hybrid” model, and the product page describes it as combining reasoning, multimodal, and agentic coding strengths under one deployment. The model card lists a 256k context window and shows the model as “Open,” signaling an option for organizations that want more control and portability than fully closed vendors (Mistral).
Z.ai’s release notes list GLM-5-Turbo as a model release focused on stability and efficiency for long-chain agent tasks, along with strengthened tool and “Skills” integration and improved instruction decomposition for multi-step workflows. The same release-notes page positions it as designed for high-throughput workloads and extended task execution where planning and tool coordination matter (Z.ai).
MiniMax’s English API “Models” release notes list “MiniMax M2.7” as a March 18, 2026 release and frame it as “Beginning the journey of recursive self-improvement.” The model documentation also positions M2.7 as an “agentic model” with tool use and “interleaved thinking,” highlighting the direction toward systems that can reason between tool calls (Minimax).
Cursor announced Composer 2 as a new coding model “frontier-level at coding” and published benchmark deltas versus earlier Composer versions, with stated pricing for standard and fast variants. The announcement also describes training changes (continued pretraining and reinforcement learning on long-horizon tasks) intended to improve multi-step coding performance, which is where many code agents fail today (Cursor).
2. AI Tools and Features
Google’s product blog announced a Gemini-powered conversational feature called Ask Maps, designed to answer complex real-world questions and provide personalized recommendations inside Google Maps. The same announcement introduced Immersive Navigation, described as a major navigation redesign with vivid 3D views and more intuitive guidance, with an initial rollout in the U.S. (and Ask Maps also rolling out in the U.S. and India (Google).
NVIDIA announced an “open agent development platform” centered on NVIDIA Agent Toolkit and OpenShell, described as an open-source runtime that enforces policy-based security, network, and privacy guardrails for autonomous agents (“claws”). The press release also claims a hybrid agent architecture (AI-Q blueprint) can cut query costs substantially using a mix of frontier and open models, and it lists a broad ecosystem of enterprise software partners (NVIDIA).
Mistral describes Forge as a system for enterprises to build frontier-grade AI models grounded in proprietary knowledge, emphasizing training on internal documentation, codebases, structured data, and operational records. Forge is positioned around control and governance, including aligning behavior to internal policies and evaluation criteria through techniques like reinforcement learning, rather than relying only on generic public-data models (Mistral).
Anthropic’s Claude Code changelog for version 2.1.80 (dated March 19, 2026) states it added channels in research preview, allowing MCP servers to push messages into a running session. The Channels documentation explains the concept as forwarding external events (alerts, webhooks, chat messages) into a live Claude Code session, with supported plugins including Telegram and Discord and organizational enablement controls for Team/Enterprise (Claude).
Unsloth’s documentation introduces Unsloth Studio as an open-source, no-code web UI to train, run, and export open models locally, highlighting faster training with less VRAM and support for multiple model types across major OS platforms (Unsloth).
DingTalk opened beta access for “Wukong,” described as an enterprise AI-native work platform intended as an ecosystem entry point where software companies and developers can provide skills to enterprise users (36Kr).
3. AI for Science
EMBL-EBI announced that millions of predicted protein complexes (homodimers) were added to the AlphaFold Database through a collaboration involving EMBL-EBI and Google DeepMind, with the goal of widening access to protein interaction structures (EBI).
NVIDIA’s press release on expanding open model families states that BioNeMo’s Proteina-Complexa is a generative model for protein binder design and is being used by multiple organizations, alongside a new open dataset of millions of AI-predicted protein complex predictions (NVIDIA).
4. Other
China’s world-first commercial approval for a paralysis BCI: A regulator-approved implanted BCI linked to a robotic glove is a real-world commercialization milestone for neurotech, with longer-term implications for assistive device markets and rehabilitation ecosystems (FierceBiotech).
Silicon battery materials at EV scale: Group14’s BAM-3 factory starting EV-scale production is a concrete step toward shorter charging times and a less graphite-dependent supply chain, with downstream implications for fleet TCO and charging-network strategies (TechCrunch).
