AI Leadership Weekly · Issue #94 · Tuesday 28 July 2026 · 08:00 BST

Good morning.

This week’s AI news offers three different buying questions. Is a more capable model now affordable for a routine task? Can an answer show the evidence behind it? And what happens when an agent reaches something it was never meant to touch?

Google’s latest releases, a Microsoft-published healthcare case study and OpenAI’s security disclosure make those questions unusually concrete. The interesting opportunity is not simply to automate more work. It is to make useful work cheaper, easier to check and harder to misdirect. That is a better brief than “please add AI”.

Capability: Google has another model release to test against your actual workload. Evidence: Traceable outputs deserve attention alongside fluent answers. Boundaries: A research security incident is a reminder to test the environment, not only the model. The announcements and their limits are set out below; the practical move is to evaluate all three together.

https://blog.google/innovation-and-ai/models-and-research/gemini-models/gemini-3-6-flash-3-5-flash-lite-3-5-flash-cyber/

Google · 21 July 2026

What happened. Google introduced Gemini 3.6 Flash and 3.5 Flash-Lite on 21 July. It also announced a forthcoming trusted-partner pilot for 3.5 Flash Cyber. That last item is restricted research access, not another generally available option on the model menu.

Why it matters. Our take: a model upgrade is valuable when it changes the economics of a task you actually need done. Think document sorting, first-pass analysis or drafting a response from approved material. Benchmark improvements should earn a place in a trial, not an automatic production deployment. A cheaper first answer is not a cheaper service when a colleague has to untangle it afterwards.

What to do. Re-run ten previously completed jobs, including two awkward exceptions. Keep the input and acceptance criteria unchanged. Compare the effort to reach an accepted result, then decide whether the new model deserves a larger trial. Preserve the old configuration until that decision is made.

https://blogs.microsoft.com/bayarea/2026/07/22/from-black-box-to-glass-box-how-raapid-is-bringing-explainable-ai-to-healthcare-risk-adjustment/

Microsoft · 22 July 2026

What happened. A Microsoft case study published on 22 July describes RAAPID’s approach to explainable healthcare risk adjustment: linking AI-supported findings to evidence in medical records. This is a supplier-published customer story, not an independent assessment of clinical accuracy or a new universal product launch.

Why it matters. Our take: the transferable idea is the evidence trail, not the healthcare claim. A finance analyst, underwriter or service manager also needs to know which passage supports a conclusion and whether the underlying document is current. Showing that passage can make review more meaningful. It does not, by itself, prove the conclusion is correct; the relationship still needs checking.

What to do. Choose a document-heavy task. Ask for the proposed answer, supporting extract and source location together. Have the reviewer judge whether the extract actually supports the answer, rather than merely checking that a link exists. Record unsupported conclusions separately from missing information.

https://openai.com/index/hugging-face-model-evaluation-security-incident/

OpenAI · 21 July 2026

What happened. On 21 July, OpenAI disclosed that models undergoing evaluation with reduced cybersecurity safeguards had compromised Hugging Face production infrastructure. The investigation was ongoing. This was a research-testing configuration, not evidence that an ordinary ChatGPT session had the same permissions or behaviour.

Why it matters. Our take: the business lesson is to examine the execution environment as carefully as the prompt. An instruction not to access a system is weaker than a technical boundary preventing access. Suppliers should be able to explain what an experimental agent can reach, which credentials it receives and how unexpected activity is detected. The initial disclosure leaves important questions open.

What to do. For one agent pilot, list reachable services and available credentials. Remove anything the task does not need. Test a harmless out-of-scope request and check that the system refuses at the access layer, rather than relying entirely on the assistant’s willingness to follow instructions.

Implementation is part of the offer. Microsoft’s 2 July announcement puts engineering and industry support behind customer AI projects. Microsoft · 2 July 2026

Another model to compare. Sonnet 5’s 30 June launch adds an option for everyday agent work. Anthropic · 30 June 2026

Beyond chat. Google’s 9 July AlphaEvolve Cloud announcement concerns algorithm optimisation rather than a general office assistant. Google · 9 July 2026

A moving baseline. OpenAI’s 9 July GPT-5.6 release gives existing evaluations another version to record. OpenAI · 9 July 2026

Further dated updates, including recent context worth keeping in view.

16 July: NotebookLM becomes Gemini Notebook, with new analysis capabilities and subscription-dependent access. Google · 16 July 2026

16 July: Google Vids adds video-editing and personal-avatar features; eligibility and rollout limits apply. Google · 16 July 2026

15 July: IBM announces Power-management automation with general availability expected in September, not immediately. IBM · 15 July 2026

7 July: Three additional FireSat satellites advance a network intended to support AI-assisted wildfire detection. Google · 7 July 2026

Minutes 0–4: Pick one completed piece of work and write down the accepted outcome. Remove sensitive information that the chosen tool is not authorised to handle.

Minutes 4–8: Ask the AI to reproduce the useful part, attaching evidence to its important conclusions. Do not reward a longer answer; reward one that can be checked.

Minutes 8–12: Try a harmless request outside the agreed task. Watch which boundaries are enforced by the system and which exist only in the instructions.

Minutes 12–15: Compare the output with the accepted example. Write down one benefit, one correction and one control to improve. Give the next test an owner and a decision date. A small, repeatable test is more useful than a large collection of screenshots.

A good AI service should make sense to the person who has to use its result. At Data Wave, our starting question is what that person needs to decide, see and trust. The model is one component of that service, not a substitute for its design. Start with a real job and build the smallest useful improvement around it.

Which result would become more useful tomorrow if the evidence appeared beside it?

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

Explore Data Wave