AI Leadership Weekly · Issue #64 · Tuesday 30 December 2025 · 08:00 GMT
Good morning.
Groq’s announcement leads this quieter holiday edition: a technology-licensing agreement with NVIDIA, not a sale of the whole company. We then look back at two December developments that deserve more than a passing mention: Mistral’s document-reading update and OpenAI’s work on browser-agent security. Those two items are labelled as lookbacks, rather than presented as fresh Christmas-week launches.
IN 60 SECONDS
Groq licenses technology to NVIDIA. Groq announced a non-exclusive inference-technology licensing agreement with NVIDIA.
Mistral upgrades document reading. Mistral introduced OCR 3, reporting better handling of scanned documents, handwriting and tables.
Atlas receives security hardening. OpenAI described further defences against prompt injection in Atlas, treating malicious webpage instructions as an ongoing security problem.
CEO / COO / CXO CHECKLIST
CEO: Read the actual supplier announcement before changing your plan.
COO: Assess extracted fields against the original document.
CXO: Keep browser-agent permissions narrower than an employee’s full access.
TOP STORIES
1. Groq licenses technology to NVIDIA
Groq · 24 December 2025
What happened. On 24 December, Groq announced a non-exclusive licence of its inference technology to NVIDIA and said some team members would join NVIDIA. Groq also said it would remain independent and that GroqCloud would continue operating. Its announcement did not disclose a transaction value. Do not replace those details with an unconfirmed acquisition headline.
Why it matters. Our take: For a customer, the immediate question is the continuity of the service being used, not the size of a rumoured deal. Separate ownership of technology, staffing changes and delivery commitments. Ask for an account-specific explanation where your operation depends on the supplier. Avoid inventing either reassurance or disruption beyond the evidence.
What to do. Review the dependency record for any affected service. Confirm the contracting entity, support route and recovery option with the supplier. Record an unanswered question as a question, not a risk already realised. Keep the current service assessment grounded in actual performance and documented commitments.
2. December lookback: mistral upgrades document reading
Mistral AI · 17 December 2025
What happened. Mistral introduced OCR 3 on 17 December, describing improved extraction from forms, scans, handwriting and complex tables, with API and Document AI access. The company’s comparative results are its own evaluations. Extracting a plausible text representation does not establish that every important field in a business record is correct.
Why it matters. Our take: Start with the fields whose errors cause rework or a poor decision. A misplaced decimal or a value attached to the wrong row deserves a different response from a cosmetic formatting error. Keep extraction, validation and posting separate, so uncertainty in reading the document does not silently become a confirmed transaction.
What to do. Choose a representative document set, including awkward scans and tables. Label the critical fields manually and compare them with the extracted output. Define an exception queue and retain links to source pages. Do not allow a pilot to post uncertain values into an operational record.
3. December lookback: atlas receives security hardening
OpenAI · 22 December 2025
What happened. OpenAI published an account on 22 December of its work to harden ChatGPT Atlas against prompt injection. It described automated attack discovery and iterative defences, while treating prompt injection as an open security challenge. The article does not claim that a browsing assistant can safely be trusted with unrestricted access.
Why it matters. Our take: External content should be treated as evidence to inspect, not as authority to change the assignment. Keep the business owner involved in deciding what an assistant may send, delete, purchase or modify. A security improvement is valuable, but it should not be used as a substitute for a narrow operating boundary.
What to do. Review one browser-based pilot with security and the process owner. Remove unnecessary permissions and require approval for consequential actions. Ask how the team will detect and recover from an unintended change. Use a safe test environment rather than experimenting against real customer records.
SIGNALS FROM THE LAST MONTH
16 December · ChatGPT updates image generation. OpenAI introduced a new ChatGPT Images experience and model, reporting improvements in editing and following detailed visual instructions. Source
11 December · GPT-5.2 launches. OpenAI released GPT-5.2, positioning the model for professional knowledge work, coding and longer, tool-assisted assignments. Source
9 December · A foundation for agent standards. The Linux Foundation launched the Agentic AI Foundation, bringing projects including MCP, goose and AGENTS.md into a shared organisational home. Source
9 December · Accenture partners with Anthropic. The companies announced a multiyear partnership combining Claude deployment, industry work and training for Accenture professionals. Source
IN BRIEF
More dated updates from the preceding 30 days.
18 December · Skills become an organisational resource. Anthropic expanded how organisations can distribute and manage Skills, bringing reusable instructions into a more centrally managed setting. Source
18 December · GPT-5.2-Codex reaches paid users. OpenAI introduced GPT-5.2-Codex for eligible paid ChatGPT users; broader API availability was still forthcoming in the launch announcement. Source
17 December · Coursera and Udemy plan to combine. Coursera and Udemy announced a proposed combination. Completion remained subject to the transaction’s conditions and approvals. Source
17 December · Gemini 3 Flash arrives. Google released Gemini 3 Flash, broadening access to its latest model generation with an emphasis on speed and cost. Source
THE 15-MINUTE PLAYBOOK
Prepare a three-decision January handover
Minutes 0–4 · Check continuity. Select one important AI supplier or service. Write down the evidence you have about continued operation and the questions still open. Identify the person who can obtain an authoritative answer and where the response will be recorded.
Minutes 4–8 · Choose the critical data. Pick one document-led process. Identify the few fields that determine a meaningful business action. Specify the source of truth and the conditions under which a person must verify the extraction before it proceeds.
Minutes 8–12 · Draw the action boundary. List what the pilot can read, suggest and change. Keep externally visible actions in a separate category. Review the boundary with the service owner and document how an incorrect change would be identified and reversed.
Minutes 12–15 · Make the next meeting useful. Give each unresolved question an owner, evidence request and decision date. Attach a real example to the meeting invitation. Cancel any demonstration that does not help answer one of those questions.
DATA WAVE MOMENT
Use a quiet planning moment to reduce ambiguity. Clear ownership, a small set of critical checks and an honest account of what remains unknown make a better starting point than a bigger collection of tools.
QUESTION FOR READERS
Which three unresolved decisions would most improve our ability to deliver useful AI in January?
Brought to you by Data Wave — your AI & Data Team as a Subscription.
