AI Leadership Weekly · Issue #61 · Tuesday 9 December 2025 · 08:00 GMT

Good morning.

Amazon and Mistral have added new options for organisations that need more than a standard chat interface. Their announcements take different approaches to model choice and customisation. OpenAI’s enterprise report adds a view of how its customers are using AI. Read together, the useful question is not which announcement sounds largest, but which part of your delivery problem it addresses.

Mistral 3 offers open models. Mistral introduced its third model generation, spanning smaller models and Mistral Large 3 under an Apache 2.0 licence.

Amazon expands Nova. Amazon announced new Nova models and Nova Forge, an offering for organisations seeking deeper customisation during model development.

OpenAI surveys enterprise adoption. OpenAI published enterprise usage and survey findings.

https://mistral.ai/news/mistral-3/

Mistral AI · 2 December 2025

What happened. On 2 December, Mistral announced the Mistral 3 family, including Mistral Large 3 and smaller Ministral 3 models. It released the family under Apache 2.0 and described deployment options across its own service and partner platforms. The announcement creates additional model choices; it does not supply a complete operating team.

Why it matters. Our take: An open model can give you more deployment flexibility, but ask who will maintain the service around it. Security configuration, updates, monitoring and support still need an owner in your chosen arrangement. The right comparison is a managed outcome against an internally operated outcome, not a licence price against a subscription price.

What to do. Ask for two costed options for the same bounded workload: a managed service and an organisation-controlled deployment. Include setup, testing, support and recovery. Identify the specific requirement that makes extra control valuable. Avoid paying for flexibility your use case will never need.

https://www.aboutamazon.com/news/aws/aws-agentic-ai-amazon-bedrock-nova-models

Amazon · 2 December 2025

What happened. At re:Invent, Amazon expanded its Nova portfolio with Nova 2 models and Nova Forge, a service for building customised variants using an organisation’s data during training. The announcements also distinguished different models and access stages. Availability of one Nova product should not be read as immediate access to the entire portfolio.

Why it matters. Our take: Customisation deserves a business reason. Start by asking whether the problem is missing knowledge, an unclear procedure or a genuine capability gap. A model trained on your data will not resolve contradictory operating instructions by management magic. Keep the simplest credible approach in the comparison until a test demonstrates its limit.

What to do. Take five recurring failures from an existing pilot. Classify the apparent causes before choosing a technical remedy. Ask the supplier what customisation would change, what test would prove it and who would maintain the training data. Price a smaller alternative alongside the proposed programme.

https://openai.com/index/the-state-of-enterprise-ai-2025-report/

OpenAI · 8 December 2025

What happened. OpenAI published its enterprise AI report on 8 December, using customer usage data and a survey of 9,000 workers across almost 100 enterprises. It reported growing use and benefits described by workers. These are observations from OpenAI’s customer base, including self-reports, rather than a controlled estimate of economy-wide productivity.

Why it matters. Our take: Use the report to improve your questions, not to fill a benefits spreadsheet with someone else’s average. A person saving time on preparation may spend it on better service, more work or additional review. Decide which outcome matters before counting the benefit. Keep experience measures and operational measures visible together.

What to do. For one team, pair a short user survey with a real workflow measure. Ask about usefulness and friction, then inspect turnaround time, accepted output and rework. Capture a baseline and a comparable sample. Explain differences in workload instead of attributing every improvement to AI.

24 November · Claude expands tool use. Anthropic introduced tool search, programmatic tool calling and tool-use examples for developers building agents with larger collections of tools. Source

19 November · Codex tackles longer assignments. OpenAI released GPT-5.1-Codex-Max, using context compaction to support software tasks that continue beyond a single context window. Source

18 November · Gemini 3 begins rolling out. Google introduced Gemini 3 Pro in preview; the more specialised Deep Think mode initially had more restricted testing access. Source

14 November · Claude previews structured outputs. Anthropic introduced a public beta for structured outputs on selected models, helping developers request responses in a defined data format. Source

More dated updates from the preceding 30 days.

1 December · DeepSeek updates its model line. DeepSeek released V3.2 and a separate Speciale variant. The latter’s temporary endpoint had different capabilities and availability conditions. Source

26 November · How long-running agents hand over. Anthropic described an engineering approach combining environment setup, progress records and incremental work to help coding agents continue across sessions. Source

25 November · MCP marks its first year. The Model Context Protocol published its November specification update, including additions for authorisation and longer-running work. Source

24 November · Opus 4.5 launches. Anthropic released Opus 4.5 with API launch pricing of $5 per million input tokens and $25 per million output tokens. Source

Minutes 0–4 · Name the constraint. Write down the work that is too slow, too costly or too inconsistent. Identify its recipient and volume. State what happens today when the task cannot be completed correctly, rather than starting with a preferred product.

Minutes 4–8 · Compare two routes. Describe the simplest approved approach and the more ambitious alternative. Include data preparation, integration, training and ongoing support. Mark unknown costs as questions to resolve, not as zeros that make one option look cheaper.

Minutes 8–12 · Choose the evidence. Select one outcome measure and one quality safeguard. Record the existing baseline, who will collect results and the cases to compare. Add a user-experience question that might explain why the numbers change.

Minutes 12–15 · Limit the commitment. Fund the next useful learning step, with a named decision date. State what would justify expansion, further testing or stopping. Preserve a record of the option you rejected and the evidence that could reopen it.

Put the business case close to the people doing the work. The aim is a decision your team can explain and revisit—not an impressive estimate that nobody owns once the pilot begins.

Which part of our proposed AI spend directly addresses a measured business constraint?

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

Explore Data Wave