AI Leadership Weekly · Issue #75 · Tuesday 17 March 2026 · 08:00 GMT

Good morning.

Google announced more Gemini features across familiar workplace applications. Mistral added an open model, while NVIDIA introduced tooling around personal agents. Together, the announcements offer three routes forward: use more of an existing suite, choose a model to operate yourself, or build around an agent runtime. The business question is which route removes a genuine obstacle without creating unnecessary work elsewhere.

Gemini expands across Workspace. Google announced new AI-assisted creation and analysis features in Docs, Sheets, Slides and Drive, with availability varying by product and plan.

Mistral releases Small 4. Mistral introduced Small 4 under Apache 2.0, combining capabilities within an open model that organisations could evaluate for their own deployments.

NVIDIA introduces NemoClaw. NVIDIA announced NemoClaw, adding runtime and security tooling around OpenClaw-based personal agents rather than replacing business-level controls.

https://blog.google/products-and-platforms/products/workspace/gemini-workspace-updates-march-2026/

Google · 10 March 2026

What happened. On 10 March, Google announced further Gemini capabilities across Docs, Sheets, Slides and Drive, including beta features drawing on selected files, email and web information. Access was tied to specified plans and rollout conditions. The announcement was not evidence that every employer’s Workspace subscription already included the same features or that all connected information was appropriate to use.

Why it matters. Our take: A connected preparation process could reduce repeated copying, but the recipient still needs to understand what supports the conclusion. Require a distinction between source facts, calculations and proposed narrative. A polished presentation should not make a draft assumption look like an approved business figure. Preserve that distinction as the output moves between applications.

What to do. Choose a recurring decision pack with approved inputs. Ask the team to trace two important numbers and one recommendation through its preparation. Test assistance on copies, then have the usual owner check the sources and calculations. Confirm the relevant plan and administrative settings before widening access.

https://mistral.ai/news/mistral-small-4/

Mistral AI · 16 March 2026

What happened. Mistral released Mistral Small 4 on 16 March under Apache 2.0. It described a model combining reasoning, coding and text-and-image capabilities, with configurable reasoning effort. The release adds an open model option; the Small name does not by itself establish that any ordinary laptop can run the service or meet a business workload’s requirements.

Why it matters. Our take: Translate the choice into a service brief: what data must stay where, how quickly must the work finish, and who will maintain the environment? Compare a managed deployment with an organisation-operated one using those requirements. Extra control is valuable when it solves a real constraint, not simply because the model weights are available.

What to do. Ask for a small, costed trial against an existing task. Include infrastructure, monitoring, updates and support alongside model access. Keep the inputs and acceptance checks the same as your current route. Require a practical explanation of what the organisation gains by taking on any additional operating responsibility.

https://nvidianews.nvidia.com/news/nvidia-announces-nemoclaw

NVIDIA · 16 March 2026

What happened. NVIDIA announced NemoClaw for the OpenClaw platform on 16 March. It described a stack using its OpenShell runtime to provide an isolated environment and policy-based privacy, network and security controls, with local and cloud model options. These were announced technical controls, not a guarantee that every configured agent would behave safely or produce correct business decisions.

Why it matters. Our take: Separate the boundary around the agent from the judgement about its work. Isolation and restricted access address part of the operating design; the owner must still define the task, acceptable output and escalation route. Do not let a reassuring platform description replace a demonstration of what the proposed configuration actually permits and prevents.

What to do. Ask the delivery team to demonstrate one allowed action, one blocked action and the evidence left for review, using a non-production environment. Then examine a plausible but incorrect result. Confirm who would catch it before use and who can stop the service if the operating behaviour differs from the agreed brief.

26 February · Nano Banana 2 launches. Google introduced Nano Banana 2, combining image-generation capabilities with a greater emphasis on speed and broader product access. Source

24 February · Cowork gains enterprise plugins. Anthropic introduced organisation-specific plugin distribution and customisation, allowing teams to package connected tools and reusable skills. Source

20 February · Claude previews code-security review. Anthropic announced a limited research preview of Claude Code Security, designed to identify vulnerabilities and propose fixes for human review. Source

19 February · Gemini 3.1 Pro enters preview. Google introduced Gemini 3.1 Pro in preview, targeting more complex reasoning tasks across its consumer and developer products. Source

More dated updates from the preceding 30 days.

9 March · Microsoft previews Copilot Cowork. Microsoft announced Copilot Cowork for longer, multi-step work, initially in limited research preview with broader Frontier access planned. Source

5 March · Anthropic studies labour-market exposure. Anthropic published an early analysis of AI exposure and labour-market outcomes, distinguishing observed usage from theoretical task coverage. Source

5 March · GPT-5.4 adds computer use. OpenAI launched GPT-5.4 across ChatGPT, the API and Codex, with native computer-use capabilities through developer and coding routes. Source

27 February · Amazon and OpenAI expand their relationship. The companies announced a strategic partnership including investment, compute and a planned stateful runtime on Amazon Bedrock. Source

Minutes 0–4 · Choose the final output. Select a real deliverable such as a weekly performance brief. Name its recipient and the decision it supports. Pick two important facts and one calculation that must remain traceable through the entire process.

Minutes 4–8 · Follow the information. Walk from the approved source to the finished output. Mark where a tool retrieves, transforms, interprets or copies information. Record changes of owner and any point where an assumption becomes indistinguishable from a source fact.

Minutes 8–12 · Introduce a safe exception. Use a test copy with a missing input or contradictory figure. Observe whether the workflow stops, flags uncertainty or invents a completion. Confirm what the next person sees and whether they know who must resolve the issue.

Minutes 12–15 · Accept the whole service. Review the final output with its intended recipient. Ask whether they can trace the selected claims and understand the unresolved points. Agree one improvement to the handover before adding another tool or expanding the permitted actions.

Connect capability around an outcome the business recognises. Keep the evidence, responsibilities and acceptance decision visible from beginning to end, so the benefit is a better service rather than a longer chain of impressive tools.

Who can explain—and stand behind—the result after all the tools have finished their part?

Brought to you by Data Wave — your AI & Data Team as a Subscription.

Explore Data Wave