Callista AI Weekly (December 1 - December 7, 2025)
This week’s developments marked a decisive shift from experimentation to execution in artificial intelligence. Governments and global enterprises moved agentic AI from pilot pilots to production. Major vendors announced deeper integrations, more powerful models, and pre-built agents designed to take action within business workflows.
New AI Use Cases
Agentic AI in government: FDA moves from pilots to real workflows
The U.S. Food and Drug Administration (FDA) deployed agentic AI tools across the agency to help staff with complex, multi-step tasks - meeting management, review processes, inspections, and more. The system is optional, overseen carefully to ensure reliability, and already driving a grassroots innovation cycle: FDA employees are running an internal challenge to build new agentic AI solutions. The signal is unmistakable - public sector organizations are now using AI to modernize operations, not just test them. For businesses selling into regulated markets or working with public institutions, this legitimizes agentic AI for critical, audit-heavy processes, provided there is strong oversight and clear accountability.
Enterprise transformation: Swiss Re puts AI at the center
Swiss Re placed AI “at the center” of its new strategy, moving beyond coding and document acceleration to agentic systems that can reshape underwriting and claims workflows. In partnership with Palantir, Swiss Re is turning massive insurance data into an AI platform for faster, more consistent decisions. The firm plans to embed AI assistants across the business and pair technology with upskilling and governance to ensure responsible adoption. The takeaway for conservative, regulated industries: broad AI integration is now feasible when paired with robust data platforms, staff readiness, and clear guardrails.
Tackling legacy IT at scale: Air Canada’s AI-led modernization
Air Canada used AWS’s agentic AI service Transform to automate software modernization - converting thousands of legacy functions in days and reportedly cutting maintenance costs by 80%. This is a pointed reminder that high-impact AI use cases extend far beyond chatbots. Automating code refactoring and updates can accelerate modernization programs that would otherwise take months or years, freeing budgets and teams to focus on new capabilities.
Holiday retail and AI visibility: bots as traffic drivers—and kingmakers
Retailers like Walmart and Amazon rolled out shopping assistants (“Sparky” and “Rufus”) that guided customers through discovery, comparison, and purchase decisions. The payoff was dramatic: industry reports indicate AI-driven traffic to retail sites jumped over 800% year-over-year on Black Friday, with Salesforce estimating around $14 billion of global online sales influenced by AI or agent recommendations over the weekend. For marketers, a new discipline is emerging - generative AI optimization - where product data, content, and feeds are tailored for AI discovery and recommendation systems, much like SEO reshaped digital marketing a decade ago.
Physical AI in the office: one robot, many jobs
SoftBank and Yaskawa Electric unveiled a prototype “Physical AI” office robot coordinated by a cloud AI brain over multi-access edge computing. Rather than single-purpose automation, the robot can switch roles dynamically - from security rounds to deliveries - while integrating with building management systems and reacting to real-time sensor data. A Tokyo demo showed one machine performing multiple roles that usually require several bots or human intervention. For facility operators and hospitals under staffing pressure, flexible AI-driven robotics offers a path to safety and efficiency without the rigidity of narrow-purpose robots.
Major Vendor Updates
OpenAI and Accenture: scaling enterprise agents with standard toolkits
OpenAI and Accenture announced a broad collaboration to bring agentic AI “into the core of business.” Accenture will equip tens of thousands of consultants with ChatGPT Enterprise and jointly offer programs to help clients deploy AI agents in functions like customer service and supply chain. A notable element is OpenAI’s AgentKit, a toolkit to design and deploy custom agents. For enterprises, this partnership offers vetted playbooks and implementation muscle, accelerating adoption while mitigating execution risks that often derail large-scale AI programs.
Snowflake and Anthropic: $200M bet to embed Claude in enterprise data
Snowflake expanded its partnership with Anthropic to bring the Claude model family natively into Snowflake’s platform for 12,000+ customers. The goal is to enable AI agents that work directly over enterprise data, securely and under governance, writing SQL behind the scenes and returning answers in natural language. Early use cases include automated customer support and personalized investment reports that blend portfolios, market data, and compliance rules. The pitch is compelling: AI where the data already lives, with guardrails, to speed up analysis and reduce data movement risks.
AWS: models, frontier agents, hardware, and on-prem “AI factories”
Amazon’s AWS showcased a full-stack push at re:Invent:
New AI models: Four additions to the Nova family emphasize reasoning, multimodal understanding, conversation, coding, and agentic tasks. Nova Forge introduces “open training” so companies can fine-tune with their own data. Early users like Reddit and Hertz reported noteworthy gains, including 90% reliability automating web UI tasks.
Autonomous “frontier” agents: AWS unveiled pre-built agents for software engineering (code-named Kiro), security (a virtual analyst), and DevOps (routine maintenance), citing tests where agents sped development and caught subtle logic bugs missed by conventional tools. Companies including Commonwealth Bank of Australia and SmugMug tested these agents.
AI infrastructure: Trainium3 UltraServers on 3nm boast 4x the performance of the prior generation, with early adopters like Anthropic cutting training time and costs. AWS also introduced “AI Factories” to install AWS AI hardware and services in customer data centers, important for tight data sovereignty and latency requirements in Europe and government.
Customization tools: Bedrock and SageMaker now offer simpler fine-tuning, including Reinforcement Fine Tuning. Salesforce reportedly boosted task accuracy by 70% using these methods. The message: specialized models and agents are becoming easier for non-research teams to deploy.
Google’s Gemini in action: action-taking advisors in Ads and Analytics
Google is embedding Gemini-powered agents directly into core marketing tools. Ads Advisor can analyze campaigns and take actions with user approval - generating keywords and copy, adjusting bids, and fixing policy issues like disapprovals. Analytics Advisor converses to explain traffic anomalies and suggests next steps after automatic root-cause analysis. For marketing teams, these are practical co-workers inside existing tools—accelerating performance improvements while requiring governance to review and validate AI-initiated changes.
Nvidia’s “Physical AI” for autonomy and robotics
Nvidia announced Alpamayo-R1, an open vision-language-action model for self-driving research, building on its reasoning-focused Cosmos lineage. With an open GitHub release and the “Cosmos Cookbook” for training, Nvidia is nudging the community toward more human-like judgment in robotics and vehicles. For automotive, manufacturing, and logistics, open models can speed innovation by lowering barriers to experimentation, especially where proprietary data and domain adaptation are essential.
China’s AI momentum: models, apps, and open releases
Baidu’s ERNIE 5.0: A multimodal model for text, images, and audio with improved reasoning and factual accuracy - positioned to compete with global leaders. Baidu also introduced domestic AI chips, hedging against trade restrictions.
Alibaba’s Qwen app: A major upgrade to its consumer AI assistant (beta) signals a push beyond enterprise cloud into mainstream users - capable of multi-step content creation like research reports and slideshows from a single prompt.
DeepSeek open-source frontier: On Dec 1, Hangzhou-based DeepSeek released two massive 685B-parameter models under an MIT license, claiming parity with or better performance than GPT-5 on several benchmarks. Technical contributions include a sparse attention mechanism for ultra-long contexts up to 128k tokens at a fraction of compute cost, plus improved tool-use continuity. For businesses, this points to rapidly expanding options outside Big Tech, potentially lower costs and reduced lock-in, alongside support and trust questions that open models must address.
AI Governance Developments
Europe’s recalibration and investment in compute
The European Commission proposed delaying enforcement of the AI Act’s strict “high-risk” system rules to late 2027, extending timelines by approximately 16 months to reduce immediate burden. The package also aims to simplify requirements for smaller firms and clarify overlaps with other laws. Simultaneously, the EU and the European Investment Bank are backing “AI Gigafactories”,large-scale compute centers for AI, through InvestAI, a €20 billion program targeting up to five major computation hubs across Europe. The dual message: a phased approach to regulation paired with significant public support for infrastructure, so innovation is not throttled even as compliance expectations mature.
Intellectual property intensifies: NYT vs. Perplexity
The New York Times sued Perplexity on Dec 5, alleging its AI search tool scraped and reproduced Times content without permission and falsely attributed AI text to the NYT. Following similar publisher concerns, this case highlights unresolved norms around training data, content reproduction, and licensing. For businesses, the compliance signal is clear: ensure that AI systems - whether in-house or vendor-provided - respect IP rights, use appropriately licensed data, and avoid output that mimics proprietary content verbatim.
Safety practices under scrutiny
A new AI Safety Index by the Future of Life Institute criticized major AI developers’ safety efforts, noting gaps between public principles and concrete controls, and the lack of robust plans for potentially superintelligent models. Companies are also lobbying against stringent rules while public concern rises. For adopters, expect mounting pressure, from regulators, customers, and partners, to demonstrate bias mitigation, error handling, and safe operation in AI deployments.
United States: the Genesis Mission
Beyond regulation, the U.S. signaled a drive to leverage AI for national competitiveness. An Executive Order in late November launched the “Genesis Mission,” a government initiative to integrate federal datasets and supercomputers into an AI platform for scientific discovery across energy, health, and materials. Agencies like the Department of Energy began implementing it this week, with opportunities for private-sector collaboration and funding. The message to industry: while safety debates continue, governments are also investing to accelerate AI’s positive impact.
Breakthrough Research
Long-term memory for AI agents: Titans + MIRAS
Google researchers introduced the Titans architecture and MIRAS framework to give AI a form of long-term working memory. Today’s Transformer models struggle with prolonged tasks because they can only attend to limited context and don’t update themselves during sessions. Titans combines Transformer-like quick reasoning with a memory module that updates in real time as new information arrives - “learning on the fly” and retaining salient details over extended sequences. Early results include dramatic speed-ups for long-context processing (demonstrated up to 128,000 tokens). For enterprises, this points to agents that maintain continuity across multi-day projects, large document reviews, or persistent customer engagements, reducing rework and improving quality where context is king.
Open-source frontier with sparse attention
As noted earlier, DeepSeek’s newly released 685B-parameter models pair performance claims rivaling top-tier systems with a sparse attention mechanism to process extremely long inputs at a fraction of the cost. By indexing and focusing on relevant portions of text, the architecture sidesteps quadratic scaling challenges and reportedly makes book-length analysis 70% cheaper without losing accuracy. The models also maintain chain-of-thought across tool calls—addressing a persistent challenge in agent frameworks. For industry, the consequence is democratization: frontier-like capability may be accessible without proprietary vendors, accelerating domain-specific fine-tuning and reducing cost. The flip side is governance: powerful open models in many hands increase the importance of responsible deployment.
AI in science: from genomics to creativity
Academic studies this week showcased AI’s reach across research disciplines. At the American Society of Hematology, researchers used AI to uncover how subtle DNA structure failures can trigger certain blood cancers—an illustration of pattern-finding power in genomics. Nature cautioned that while AI speeds up research—drafting papers and generating hypotheses, over-reliance can introduce biases. Another study reported by TechXplore found that AI collaboration can increase human creativity in design tasks, offering non-obvious options that spur novel outcomes. For R&D-heavy sectors, from pharma to advanced manufacturing, these signals suggest AI as a versatile co-investigator, amplifying discovery while requiring careful oversight for quality and ethics.
Alignment and human preferences
A NeurIPS 2025 best-paper study highlighted miscalibration risks in reward models trained from human feedback: systems can optimize for proxy signals (e.g., clicks) while missing what diverse users actually value, a form of “reward hacking.” For customer-facing AI - chatbots, recommenders, decision aids - this underscores the need for representative feedback, ongoing audits, and calibration checks to ensure the AI’s incentives align with long-term satisfaction and brand trust.
Conclusion
Across government and industry, agentic AI is moving from concept to cornerstone. The FDA’s deployment, Swiss Re’s enterprise-wide shift, and Air Canada’s swift modernization each demonstrate tangible outcomes: leaner processes, faster decisions, lower costs. Major vendors are racing to meet enterprise demand, embedding agents into core tools, offering fine-tuning shortcuts, and even delivering AI hardware on-prem for sovereignty and latency. Meanwhile, governance is evolving: Europe’s delayed timelines coupled with big investments in compute; Switzerland’s clear guidance on public data and sovereignty; and legal and safety bodies demanding transparency and control. Research breakthroughs promise longer-term gains, agents that remember, reason over vast contexts, and align more closely with human preferences.
Ready to explore how Agentic AI can transform your organization? Visit us at https://www.callista.ch/agentic-ai to discover how we can guide your journey into this exciting new era of AI-powered productivity.
Sources
FDA News Release – “FDA Expands Artificial Intelligence Capabilities with Agentic AI Deployment” (Dec 1, 2025). U.S. Food & Drug Administration.
Accenture Press Release – “OpenAI and Accenture Accelerate Enterprise Reinvention with Advanced AI” (Dec 1, 2025). Accenture Newsroom.
Reinsurance News – “Swiss Re puts Palantir-powered AI at heart of new strategy” by Kane Wells (Dec 5, 2025).
AWS News Blog – “Frontier agents, Trainium chips, and Amazon Nova: key announcements from AWS re:Invent 2025” (Last updated Dec 4, 2025).
Anthropic News – “Snowflake and Anthropic announce $200 million partnership to bring agentic AI to global enterprises” (Dec 3, 2025). Anthropic News blog.
Google Ads & Commerce Blog – “Google’s AI advisors: agentic tools to drive impact and insights” by Dan Taylor (Nov 12, 2025; updated early Dec 2025). [Announcement of Ads Advisor and Analytics Advisor].
TechCrunch – “Nvidia announces new open AI models and tools for autonomous driving research” by Rebecca Szkutak (Dec 1, 2025).
Reuters – “China’s Baidu unveils new AI processors, supercomputing products” by Liam Mo and Brenda Goh (Nov 13, 2025). [Includes note on new ERNIE model].
Reuters – “Alibaba unveils major consumer AI upgrade with new Qwen chatbot” by Liam Mo and Eduardo Baptista (Nov 18, 2025).
VentureBeat – “DeepSeek just dropped two insanely powerful AI models that rival GPT-5 and they’re totally free” by Michael Nuñez (Dec 1, 2025).
SoftBank Press Release – “SoftBank Corp. and Yaskawa Electric Begin Collaboration on ‘Physical AI’ Utilizing AI-RAN” (Dec 1, 2025).
Salesforce News via MarketingProfs – AI-driven retail stats during Black Friday 2025 (as reported Dec 5, 2025 in MarketingProfs AI Update).
SDxCentral – “Switzerland restricts US cloud access in the public sector” by Yoana Cholteeva (Dec 2, 2025).
News Admin (Swiss Government) Press Release – “Switzerland is embracing AI but increasingly exposed to fake news” – Federal Statistical Office survey results (Dec 5, 2025).
Axios – “NYT sues Perplexity for copyright infringement” by Sara Fischer (Dec 5, 2025).
Future of Life Institute AI Safety Index 2025 – coverage summarized in Reuters piece “Study finds AI companies’ safety practices lag global standards” (Dec 2025).
Hogan Lovells summary – “President Trump launches Genesis Mission to unleash AI innovation & discovery” (analysis of Nov 24, 2025 Executive Order).
Google Research Blog – “Titans + MIRAS: Helping AI have long-term memory” by Ali Behrouz et al. (Dec 4, 2025).
NeurIPS 2025 – Best Paper Award announcement (Dec 2025) and TheNeuron.ai summary “The Best Papers at NeurIPS 2025, Explained” (Dec 2025) – [for AI alignment paper reference].
TechXplore – “Can AI make us more creative? Study reveals surprising benefits of human-AI collaboration.” (Dec 2025).
