AI Leadership Weekly · Issue #68 · Tuesday 27 January 2026 · 08:00 GMT

Good morning.

Microsoft’s new Maia accelerator addresses the cost and capacity of running models. Anthropic’s published constitution addresses how a model is intended to behave. OpenAI’s age-prediction announcement deals with the treatment of different users. These are three distinct layers of an AI service. None should be mistaken for a complete answer to the others, however reassuring the product presentation may look.

Anthropic publishes Claude’s constitution. Anthropic set out the principles intended to shape Claude’s behaviour.

ChatGPT adds age prediction. OpenAI described age prediction for consumer accounts, with additional protections for suspected under-18 users and a route to correct mistakes.

Microsoft introduces Maia 200. Microsoft unveiled an accelerator designed for AI inference, with initial deployment in its US Central datacentre region.

https://www.anthropic.com/news/claude-new-constitution

Anthropic · 22 January 2026

What happened. Anthropic published a new Claude constitution on 22 January. It describes the values and behaviour the company seeks to shape through training, while explicitly acknowledging that model outputs may not always meet those ideals. The document offers transparency into intent; it is not a guarantee that every response will comply with an organisation’s expectations.

Why it matters. Our take: Use a behavioural statement to sharpen a test, not as a replacement for one. Your service may need to handle an incomplete request, conflicting instructions or a request outside its authority. Decide what an acceptable response looks like in those situations. Keep local process rules distinct from a supplier’s general account of desired behaviour.

What to do. Ask the process owner to select three difficult interactions and write the response they would accept. Run the examples through the proposed setup and retain the outputs. Discuss both unnecessary refusals and unjustified confidence. Improve the workflow around the model where a behaviour cannot be relied upon.

https://openai.com/index/our-approach-to-age-prediction/

OpenAI · 20 January 2026

What happened. OpenAI announced on 20 January that it was rolling out age prediction on ChatGPT consumer plans to help identify likely under-18 accounts and apply additional safeguards. It described a way for adults incorrectly classified to confirm their age. The announcement concerns that consumer service, not a general business identity-verification system.

Why it matters. Our take: The transferable design question is what happens when an automated classification is wrong. Any workflow using a prediction to change someone’s experience needs an intelligible explanation and a way to resolve mistakes. Do not infer that a model-generated category is an established fact simply because a system can act on it.

What to do. Inspect one low-risk classification pilot in your business. Record the consequence of each category, who can challenge it and how a correction reaches the underlying record. Test the exception path deliberately. Do not introduce new personal-data collection just to imitate another product’s implementation.

https://blogs.microsoft.com/blog/2026/01/26/maia-200-the-ai-accelerator-built-for-inference/

Microsoft · 26 January 2026

What happened. Microsoft announced Maia 200 on 26 January, describing an inference accelerator deployed in its US Central region, with further regions planned. It claimed improved performance per dollar within its own infrastructure and outlined intended use across its AI services. That does not establish a corresponding reduction in the price paid by an individual customer.

Why it matters. Our take: Keep engineering economics and your commercial economics separate. A supplier may use greater efficiency to improve capacity, margins, speed or price. Your planning should rely on the service terms and performance you can obtain. Ask what changes for your workload rather than repeating a hardware comparison in a benefits presentation.

What to do. Review one material AI service cost with the commercial and delivery leads. Identify the actual charging unit, service constraints and renewal assumptions. Record any confirmed change separately from a future possibility. Revisit the business case when customer-facing terms or measured performance change, not merely when hardware is announced.

11 January · Claude expands its healthcare tools. Anthropic announced healthcare and life-sciences connectors and workflow tools. The launch included specialist administrative applications, not evidence of autonomous clinical judgement. Source

11 January · Google proposes a commerce protocol. Google introduced the Universal Commerce Protocol for agent-led shopping, alongside checkout plans and new tools for retailers. Source

8 January · Gmail adds Gemini features. Google announced new Gemini-powered Gmail features, with initial US-English access and different availability across free, paid and testing experiences. Source

7 January · OpenAI introduces ChatGPT Health. OpenAI announced a dedicated health-and-wellness experience with staged access. It was designed to support, not replace, care from qualified professionals. Source

More dated updates from the preceding 30 days.

16 January · OpenAI outlines advertising tests. OpenAI announced plans to test advertising for eligible adult US Free and Go users; the announcement preceded the test itself. Source

15 January · Anthropic examines how AI is used. The January Economic Index analysed sampled Claude interactions using measures such as task complexity and autonomy; it was not a workforce census. Source

14 January · Gemini adds Personal Intelligence. Google began an opt-in US beta connecting personal Google information. The initial offer did not extend to Workspace business accounts. Source

12 January · Claude introduces Cowork. Anthropic launched Cowork as a research preview for Max subscribers on macOS, allowing work with files in an authorised folder. Source

Minutes 0–4 · Choose one claim. Select an important promise about behaviour, accuracy or service efficiency. Rewrite it in terms of the person relying on the output. State the conditions under which it should hold and the consequence if it does not.

Minutes 4–8 · Specify the evidence. Decide whether the claim requires a representative test, a documented service commitment or operational measurement. Identify the owner of that evidence. Keep supplier-reported results clearly attributed and separate from your own observations.

Minutes 8–12 · Exercise the exception. Create a safe example that should cause the workflow to pause, correct itself or ask for help. Verify the route to a person. Check how the original record is preserved and how the decision can be explained afterwards.

Minutes 12–15 · Make the decision narrow. Approve only the scope supported by the evidence. Record remaining limitations and a review trigger. Do not turn a passed test on one task into unrestricted authority for every use of the same model or platform.

Build confidence through a series of specific, inspectable decisions. Clear expectations, appropriate evidence and a working correction route are stronger foundations than an assurance nobody can translate into practice.

Which important AI assurance are we accepting without knowing what evidence would prove it?

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

Explore Data Wave