AI Leadership Weekly · Issue #78 · Tuesday 7 April 2026 · 08:00 BST
Good morning.
Google added both a new open-model family and more ways to create video. Mistral’s engineering account tackles the tools agents use behind the scenes. The common opportunity is to shorten the path from an idea to a usable first version. The practical challenge is keeping the next revision just as easy, rather than producing a one-off result that nobody can maintain.
IN 60 SECONDS
Google releases Gemma 4. Google introduced Gemma 4, adding open-model deployment options for organisations assessing on-device and self-managed AI.
Google Vids adds creation tools. Google expanded Vids with new generative video and music capabilities; feature limits and access depended on the account and plan.
Mistral designs tools for agents. Mistral described Spaces, a command-line tool designed to serve both people and agents working with development environments.
CEO / COO / CXO CHECKLIST
CEO: Pick an existing prototype rather than starting a new demonstration.
COO: Record its source inputs, operating environment and acceptance checks.
CXO: Ask somebody other than its creator to repeat the task without informal coaching.
TOP STORIES
1. Google releases Gemma 4
Google · 2 April 2026
What happened. Google released the initial Gemma 4 family on 2 April under Apache 2.0: two smaller effective-parameter variants, a mixture-of-experts model and a larger dense model. The announcement describes options for local hardware and more substantial infrastructure, with reasoning, tool-use and multimodal capabilities.
Why it matters. Our take: Model availability and operational readiness are separate questions. Running a model yourself can offer a different pattern of control, but someone still needs to manage capacity, updates, monitoring and failures. For an organisation considering local deployment, the decision should include the complete service, not only the model download. It may make sense for a bounded task with stable demand; it may be less attractive for an occasional, highly variable workload.
What to do. Compare one hosted and one self-managed option against the same job. Include the effort required to operate each safely. Keep the test narrow enough to expose practical differences rather than becoming a broad infrastructure programme.
2. Google Vids adds creation tools
Google · 2 April 2026
What happened. Google’s 2 April Vids update announced ten free video generations per month for personal Google accounts using Veo 3.1. It also described additional creation features, including premium capabilities for paid subscribers. The free allowance should not be read as making every feature free or approving personal accounts for company content.
Why it matters. Our take: Video is a possible alternative to a long internal instruction, particularly when a task is visual. The challenge is maintaining the content after the process changes. Give each video the same treatment as a written procedure: a named owner, source material, a review date and a clear audience. Otherwise a polished explanation can become a persuasive version of last month’s process.
What to do. Turn one short, already-approved instruction into a draft video. Have the process owner check every step and ask a new colleague to follow it. Compare comprehension, not production novelty.
3. Mistral designs tools for agents
Mistral AI · 31 March 2026
What happened. In a 31 March engineering article, Mistral described Spaces, an internal command-line tool used to set up and deploy projects. The team explained changes needed when coding agents became users, including explicit inputs instead of interactive prompts. This was a design account, not an announcement that every enterprise could buy a finished Spaces service.
Why it matters. Our take: The transferable lesson is to expose the actual business operation clearly. A process that depends on a person recognising a screen, choosing an undocumented option or supplying missing context is difficult to delegate reliably. Improving those interfaces can also help human colleagues. Start with explicit required information and clear outcomes, then choose whether a person or an agent supplies the inputs.
What to do. Find one automation that frequently stops for clarification. List the missing inputs and make them mandatory at the start, or provide a safe, documented fallback. Do not conceal uncertainty behind an automatic “yes”.
SIGNALS FROM THE LAST MONTH
17 March · OpenAI adds GPT-5.4 mini and nano. OpenAI launched smaller models for lower-cost workloads. Access routes and supported features differed between mini, nano, ChatGPT and Codex. Source
17 March · Mistral offers enterprise model training. Mistral introduced Forge, an enterprise offering for building models around organisational data rather than only customising prompts. Source
17 March · Personal Intelligence reaches more users. Google expanded Personal Intelligence for eligible US personal accounts. The announcement did not grant equivalent access to Workspace business accounts. Source
16 March · 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. Source
IN BRIEF
More dated updates from the preceding 30 days.
26 March · Search Live expands internationally. Google expanded Search Live, allowing more users to ask questions through voice and camera input while searching. Source
24 March · ChatGPT improves product comparison. OpenAI announced shopping updates in ChatGPT, including more visual product discovery and side-by-side comparison. Source
24 March · Claude Code previews auto mode. Anthropic introduced a research preview that automates some permission decisions in Claude Code while retaining checks for higher-risk actions. Source
23 March · Mistral expands its voice offer. Mistral announced a Voxtral update, extending its speech-focused model offering beyond text-only interactions. Source
THE 15-MINUTE PLAYBOOK
Minutes 0–4: Choose a completed prototype and describe the outcome in one sentence. Avoid describing the tool instead of the result.
Minutes 4–8: List every input its creator quietly supplied: terminology, source documents, examples, permissions and assumptions. Decide which should become part of the standard package.
Minutes 8–12: Write three acceptance checks and one condition that should stop the process. Include a content check, not just whether a file was produced.
Minutes 12–15: Give the package to a colleague and ask what is still unclear. Fix the instructions before expanding the technology. Record an owner and a review date so the package can evolve deliberately.
DATA WAVE MOMENT
The value of a successful experiment grows when another team can reuse it without recreating the original discovery work. Build from a real task, retain the useful learning and make the operating method understandable. A small repeatable capability is a better foundation than several impressive demonstrations nobody owns.
QUESTION FOR READERS
Which successful AI experiment would stop working the moment its creator went on holiday?
Brought to you by Data Wave — your AI & Data Team as a Subscription.
