AI Leadership Weekly · Issue #57 · Tuesday 11 November 2025 · 08:00 GMT
Good morning.
Three developments this week are worth the attention of teams moving beyond a chat window. Google added managed file retrieval to the Gemini API, Anthropic described a different way to connect agents to tools, and Moonshot published Kimi K2 Thinking. They offer different kinds of choice: how information is found, how tools are used and where a model can run.
IN 60 SECONDS
Gemini adds File Search. Google introduced File Search in public preview, offering managed retrieval over uploaded files through the Gemini API.
A different way to connect tools. Anthropic published an engineering approach using code execution with MCP-connected tools to reduce the information passed through model context.
Kimi K2 Thinking is released. Moonshot AI published Kimi K2 Thinking and its model card, adding an open-weight option for reasoning and tool-using workloads.
CEO / COO / CXO CHECKLIST
CEO: Identify the approved source of truth and who keeps it current.
COO: Give an agent only the actions needed for its assigned task.
CXO: Compare the full cost of operating a service, including human checking.
TOP STORIES
1. Gemini adds File Search
Google · 6 November 2025
What happened. Google’s 6 November release notes announced the File Search API in public preview, enabling developers to ground responses in their own data. This was a developer capability to test, not a finished knowledge-management service for every organisation.
Why it matters. Our take: The business decision sits upstream of the answer: which material belongs in the collection? Consider a customer-service assistant. An old procedure and its approved replacement may both look relevant, yet only one should govern the response. Decide how the trial identifies current authority rather than asking the model to settle an undocumented disagreement.
What to do. Choose a small, approved collection and ten questions with known answers. Include an expired document, a conflicting version and a question the collection cannot answer. Ask the reviewer to check the evidence as well as the wording. A useful result may be “not enough information”, accompanied by the right escalation.
2. A different way to connect tools
Anthropic · 4 November 2025
What happened. On 4 November, Anthropic described an engineering pattern in which agents write code to interact with tools connected through the Model Context Protocol, or MCP. The approach can keep intermediate processing outside the model’s conversation. The article also identifies additional requirements for secure execution, resource limits and monitoring.
Why it matters. Our take: A practical way to review an agent proposal is to ask for its permitted actions in plain English. “Prepare a customer summary” and “overwrite the customer record” are different authorities. Do not let a convenient technical connection silently turn one into the other. The business owner should recognise every consequential step in the design.
What to do. Walk through one workflow with its builder. Label each step as reading, calculating, proposing or changing. Give the trial an explicit boundary before changes leave its test environment. Agree what happens when a tool fails halfway through and who receives the unfinished work.
3. Kimi K2 Thinking is released
Moonshot AI · 7 November 2025
What happened. Moonshot’s Kimi K2 Thinking model documentation, updated on 6–7 November, describes a model designed for reasoning alongside tool use, with weights available for deployment. Its performance comparisons are supplier-reported; they are not evidence of results in your own business process.
Why it matters. Our take: Owning more of the deployment can be a deliberate choice. It should answer a concrete requirement, such as a defined hosting arrangement or control over model changes. It should not be approved simply because “open” sounds inexpensive. Someone still needs responsibility for the service, its dependencies and its users when the trial becomes operational.
What to do. Ask for a one-page comparison of a managed service and a self-managed option for the same task. Require named owners for availability, updates, access and incident handling. Count the people and infrastructure needed alongside usage charges. Keep the evaluation on non-sensitive examples until the proposed data route is approved.
SIGNALS FROM THE LAST MONTH
23 October · Company knowledge in ChatGPT. OpenAI introduced connected company knowledge for Business, Enterprise and Edu, with source citations and existing access permissions. Source
21 October · ChatGPT gets a browser. OpenAI launched Atlas on macOS, with browser-agent functionality in preview for eligible paid plans rather than universal access. Source
16 October · Claude gains reusable Skills. Anthropic introduced Skills: packages of instructions and resources that Claude can use for recurring tasks, including document work. Source
16 October · Copilot expands on Windows. Microsoft announced more voice and screen-aware Copilot experiences. The experimental Actions capability was still headed towards Windows Insiders. Source
IN BRIEF
More dated updates from the preceding 30 days.
3 November · AWS supplies OpenAI compute. OpenAI announced a seven-year, $38 billion AWS agreement. Compute delivery was separate from distribution of proprietary models through Bedrock. Source
28 October · GitHub introduces Agent HQ. GitHub announced a common environment for directing coding agents, with broader third-party agent integrations planned over subsequent months. Source
28 October · Microsoft and OpenAI revise terms. The partners announced a revised agreement, including additional Azure purchasing commitments and changes to compute sourcing, not unrestricted API distribution. Source
27 October · Claude enters Excel in preview. Anthropic announced an Excel research preview and expanded financial-services capabilities. Access was limited rather than generally available. Source
THE 15-MINUTE PLAYBOOK
Draw one workflow before adding another component
Minutes 0–4: Name the result. Choose an output someone actually needs: a resolved query, a reviewed briefing or an updated record. Identify the person who can say whether it is acceptable.
Minutes 4–8: Trace the information. Write down where each important input comes from, who owns it and how a reviewer can confirm it is current. Mark unresolved ownership rather than hiding it inside the proposed AI step.
Minutes 8–12: Draw the permission boundary. Separate what the system may read, what it may suggest and what it may change. Put an explicit decision point before any material external action. Include a route for missing evidence.
Minutes 12–15: Define the first test. Agree the sample cases, success measure and person responsible for collecting the results. Compare elapsed time and correction effort with today’s method. Set a stop condition as well as a target.
Your output is a one-page experiment with a clear owner—not a shopping list of disconnected features.
DATA WAVE MOMENT
Build from within the work. Start with the people, information and decisions that matter, then connect the technology around them. The first useful step is a well-defined problem that your team can test together.
QUESTION FOR READERS
Which part of our proposed AI workflow still has no clear owner?
Brought to you by Data Wave — your AI & Data Team as a Subscription.
