AI Leadership Weekly · Issue #76 · Tuesday 24 March 2026 · 08:00 GMT
Good morning.
This week’s announcements pull in different directions. OpenAI added smaller models for repeatable work, Google expanded personal-account features, and Mistral introduced an enterprise model-training offer. That variety is helpful: not every business problem needs a larger model or a bespoke one. The right comparison starts with the task, the information it needs and the way the result will be used.
IN 60 SECONDS
OpenAI adds GPT-5.4 mini and nano. OpenAI launched smaller models for lower-cost workloads.
Personal Intelligence reaches more users. Google expanded Personal Intelligence for eligible US personal accounts.
Mistral offers enterprise model training. Mistral introduced Forge, an enterprise offering for building models around organisational data rather than only customising prompts.
CEO / COO / CXO CHECKLIST
CEO: Separate routine extraction, preparation and judgement in one high-volume process.
COO: Name the permitted information sources for each step, including what must stay out.
CXO: Compare the cost of an accepted result, not just the price of a model response.
TOP STORIES
1. OpenAI adds GPT-5.4 mini and nano
OpenAI · 17 March 2026
What happened. OpenAI introduced GPT-5.4 mini and nano on 17 March, targeting faster, lower-cost workloads. Mini was announced across the API, Codex and ChatGPT, while nano was API-only. The launch describes uses including classification, extraction and supporting larger agents. These are vendor capability claims, not a measured saving for your organisation.
Why it matters. Our take: You do not need one model to do everything. Imagine a supplier enquiry: a lightweight step identifies its subject, another retrieves the relevant record, and a more capable step handles an unusual contractual question. But a weak first step can misroute the whole case. Your comparison therefore needs to include missed exceptions, repeated attempts and review time. Cheap individual responses can still produce an expensive service.
What to do. Take twenty completed cases. Label which stages were mechanical and which needed judgement. Test one cheaper route against the existing approach, retaining a clear escalation path and comparing the final accepted output.
2. Personal Intelligence reaches more users
Google · 17 March 2026
What happened. Google expanded Personal Intelligence on 17 March, including broader US access in AI Mode in Search and a rollout to more personal users. The announcement explicitly excludes Google Workspace business, enterprise and education accounts. Connecting personal sources is a user choice, not an automatic enterprise deployment.
Why it matters. Our take: An impressive personal demonstration can create the wrong expectation in a business meeting. Colleagues may assume the same capability exists under their organisation’s controls. Before designing around it, check the account type, region and permitted sources. For work use, distinguish information that helps the individual from information the whole process is entitled to use. A customer’s circumstances should not become casually available because they appeared in somebody’s inbox.
What to do. Add three columns to your AI trial register: account type, approved data sources and access owner. Mark any personal-account experiment clearly so it cannot be mistaken for an approved business service.
3. Mistral offers enterprise model training
Mistral AI · 17 March 2026
What happened. Mistral introduced Forge on 17 March as a way for organisations to train and adapt models using proprietary knowledge. Its approach spans training, refinement and evaluation rather than simply adding documents to a prompt. Mistral positions this around enterprise terminology, processes and control of the resulting models.
Why it matters. Our take: Custom training is a possible response to a persistent capability gap, not the default answer to a messy knowledge base. First ask whether the system lacks facts, instructions or learned behaviour. Missing current facts may call for better retrieval. Unclear instructions may call for a better process. Training deserves consideration when repeatable errors remain after those basics are fixed and there is enough reliable material to test an improvement.
What to do. Pick one recurring failure. Write down why better source material or clearer instructions would not solve it. Without that explanation, keep custom training on the options list rather than the delivery plan.
SIGNALS FROM THE LAST MONTH
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
IN BRIEF
More dated updates from the preceding 30 days.
23 March · Mistral expands its voice offer. Mistral announced a Voxtral update, extending its speech-focused model offering beyond text-only interactions. Source
16 March · NVIDIA introduces NemoClaw. NVIDIA announced NemoClaw, adding runtime and security tooling around OpenClaw-based personal agents rather than replacing business-level controls. 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
10 March · 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. Source
THE 15-MINUTE PLAYBOOK
Minutes 0–4: Choose a repeated request and draw its five or six steps. Mark the point at which a wrong answer becomes consequential.
Minutes 4–8: Assign each step to a simple rule, a lightweight model, deeper reasoning or a person. This is a proposed design, not permission to automate it immediately.
Minutes 8–12: Specify what happens when information is missing or two sources disagree. Make the fallback visible rather than letting the system guess.
Minutes 12–15: Select ten ordinary cases and ten awkward ones for a comparison. Name the reviewer and define what would justify expanding the trial.
DATA WAVE MOMENT
The opportunity is not to spread one assistant across every desk. It is to match the right capability to each part of the work, then connect those parts into a service people can trust. Start with one process where the volume and exceptions are understood. Make the result visible before adding more technology.
QUESTION FOR READERS
Where are we paying for sophisticated reasoning to compensate for a badly defined process?
Brought to you by Data Wave — your AI & Data Team as a Subscription.
