AI Insights: Key Global Developments in November 2025

Staff Correspondent
3 days ago
7 min read

Welcome to the latest edition of our global AI update. This one captures the most significant shifts from late October and November, and a lot has been happening. India introduced its first comprehensive set of AI Governance Guidelines, designed to encourage innovation while promoting responsible adoption. Around the same time, companies like Cognizant began rolling out Claude to hundreds of thousands of employees, demonstrating how quickly enterprise AI is transitioning from trials to real-world workflows.

There has also been growing scrutiny on what frontier models actually cost to operate, and McKinsey’s new numbers confirm a familiar pattern. Most organisations are experimenting, but only a small group is successfully scaling AI in a meaningful way.

So as you dive in, here’s the bigger picture taking shape. Governance is becoming a necessity. Enterprise AI is finally shifting from pilots to production. The real advantage now comes from rethinking how work gets done, not just from choosing a new model.

Generative AI Model Updates

Anthropic- Extended Thinking & Tool Use Integration

Anthropic continued refining Claude's enterprise capabilities in early November:

Extended Thinking with Tool Use Integration (May 2025, Ongoing)

Claude 4 and Claude Sonnet 4 launched with extended thinking capabilities, enabling deeper cognitive processing for complex reasoning tasks.
Key capability: Extended thinking now works seamlessly during tool use phases, allowing Claude to reason through tool selection, execute external API calls, and process results- critical for agentic workflows.
Performance: 54% improvement reported in complex coding tasks via extended thinking mode, with dynamic token allocation from 1,024 to unlimited tokens based on task complexity.
Enterprise relevance: This hybrid reasoning system (instant mode for simple queries; extended mode for multi-step problems) is positioned as the foundation for agentic AI deployments.

Memory Feature Rollout

Claude's memory functionality expanded to Max and Pro tier users, enabling persistent context retention across sessions.
Implication: Organizations can now maintain session continuity for ongoing customer service, knowledge management, and iterative development workflows without having to rebuild context.

Microsoft 365 Integration

Native connectivity to Outlook and Teams data signals strategic positioning within enterprise knowledge ecosystems, reducing friction for adoption in established organizational workflows.

Google DeepMind - Gemini 2.5 Multimodal Leadership

Google's October announcements positioned Gemini for agentic workloads :

Gemini 2.5 Computer Use- Specialized model for UI interaction and task automation, outperforming alternatives on benchmarks for web navigation and form completion.
Research Applications: Partnerships with Yale (cancer therapy acceleration) and Commonwealth Fusion Systems (clean energy) demonstrate the application of AI to high-impact domains.

Regulatory & Governance Developments

India's Landmark AI Governance Guidelines

India's Ministry of Electronics and Information Technology (MeitY) released the India AI Governance Guidelines —a comprehensive, sector-agnostic framework that emphasizes innovation-enabling governance.

Key Structural Pillars:

Seven Guiding Principles (Sutras):

People-centricity, accountability, fairness, explainability, transparency, ethical deployment, and risk-based oversight.

Six Pillars of AI Governance:

Enablement (infrastructure, capacity building)
Regulation (policy & regulation, risk mitigation)
Oversight (accountability, institutions)

Institutional Framework:

AI Governance Group (AIGG): Inter-ministerial coordination across sectors
AI Safety Institute (AISI): Technical research, validation, and safety testing (online, under IndiaAI Mission)
Technology and Policy Expert Committee (TPEC): Policy development
Regulatory sandboxes: Safe innovation zones for pilot AI applications

Compliance Approach:

Sector-agnostic framework with sector-specific risk classifications (financial services, healthcare to follow in Q4 2025/Q1 2026)
Voluntary commitments backed by "techno-legal" solutions (watermarking, privacy-preserving architectures)
National AI Incident Database for harm tracking and transparency
Short-term action: Risk frameworks development; Medium-term: DPI integration with AI; Long-term: New legislation if needed

Strategic Significance for Global Operations:

This "hands-off" approach contrasts sharply with the EU's prescriptive risk-based classification. For multinational organizations:

EU Requirement: High-risk systems are tightly regulated, with compliance by August 2027
India Approach: Flexible governance with voluntary commitments and innovation priority
Implication: Dual-track compliance strategies are now necessary for global deployment

EU AI Act – GPAI Model Compliance (August 2, 2025 Milestone Reached)

The second phase of EU AI Act obligations for General-Purpose AI models became binding on August 2, 2025:

Immediate Obligations for GPAI Providers:

Publish detailed training data summaries and methodologies
Ensure strict compliance with EU copyright law
Maintain comprehensive technical documentation accessible to regulators
Assess and mitigate systemic risks for powerful models (>10²⁵ FLOPs)

Transitional Provisions:

Legacy models (placed on the market before August 2, 2025) have until August 2, 2027, to comply.
New systems must comply from the launch date.

Extraterritorial Reach:

Any GPAI model accessible in EU markets. Including indirectly- falls under compliance obligations. For regulated industries (fintech, healthcare), this creates immediate dependencies on third-party compliance validation.

Enterprise AI Adoption & Industry Watch

Cognizant Deploys Claude at Scale

Scale: Cognizant announced the deployment of Claude models and agentic tooling to up to 350,000 employees globally across corporate functions, engineering, and delivery teams.

Integration Roadmap:

Software Engineering Productivity: Claude + Claude Code integrated with Cognizant Flowsource Platform for coding, testing, documentation, and DevOps acceleration via MCP-based tool access,
Legacy Modernization: Combining Cognizant's modernization frameworks with Claude's code understanding for large-scale codebase refactoring without architectural overhauls,
Agentic Deployment: Cognizant Neuro AI Multi-Agent Orchestration + Anthropic Agent SDK for domain-specific, multi-agent systems with human-in-the-loop controls,
Industry Solutions: Vertical solutions beginning with financial services, embedding agentic workflows into regulated environments with governance controls.

Strategic Implication:

This partnership represents the convergence model: enterprise consulting, frontier AI, and existing platforms combined to achieve scaled, responsible AI deployment. It signals a market maturation from point solutions to integrated, end-to-end transformation.

McKinsey Global AI Survey- November 2025

Key Finding: Persistent Pilot-to-Scale Gap

88% of organizations report regular AI use in at least one function (up from 78% year-over-year)
BUT only ~1/3 have scaled AI beyond pilots; ~2/3 remain in experimentation phase
Enterprise-level EBIT impact limited: Only 39% report measurable EBIT impact; most attribute <5% to AI

AI High Performers (6% of respondents) Share Three Traits:

Ambitious transformation mindset: Focus on growth and innovation, not just efficiency
Workflow redesign: Redesigned processes coupled with AI adoption
Strategic investment: >20% of digital budgets allocated to AI

AI Agents Traction:

62% of organizations are experimenting with or deploying AI agents
23% actively scaling agentic AI; 39% in pilot phases
Recognition that agents require organizational restructuring, not just tool integration

Critical Insight: The 95% pilot-to-scale failure rate persists because organizations prioritize cost reduction over business transformation. High performers instead redesign workflows around AI's unique capabilities, fundamentally changing how work is executed.

OpenAI's Financial Trajectory Under Scrutiny (November 6-12, 2025)

Financial Disclosure:

2025 Operating Losses: ~$9 billion on ~$13 billion revenue (~70% burn rate)
Q3 2025 Net Loss: ~$12 billion (internal Microsoft disclosures)
Projected 2028 Losses: ~$74 billion (75% of revenue), driven by compute infrastructure spending

Strategic Positioning:

OpenAI signed $1.4 trillion in compute commitments over 8 years with cloud and chip partners
Spending ~$100 billion on backup data-center capacity alone
Projected cumulative cash burn through 2029: ~$115 billion
Expected profitability: 2029–2030

Competitive Comparison:

Anthropic: Projects 33% burn rate in 2026, down to 9% by 2027; expects break-even by 2028
OpenAI: Expects 57% burn in 2026–2027, not dropping below current levels until 2029+

Market Implication:

The stark divergence in financial trajectories reflects different strategic bets- OpenAI betting on compute dominance and explosive growth; Anthropic pursuing capital-efficient scaling. For enterprises evaluating vendors, this signals different cost structures and sustainability profiles.

AI Risk, Validation & Research

SAGE Framework – Multi-Turn Safety Evaluation (November 2025)

(Source- ACL Anthology)

A new automated safety evaluation framework addressing critical gaps in real-world deployment testing emerged this month. SAGE (Safety AI Generic Evaluation) employs adversarial agents with diverse personality profiles (Big Five model) for context-aware, multi-turn harm evaluation.

Key Research Findings:

Harm increases measurably with conversation length, contradicting assumptions of static safety
Model behavior varies significantly across user archetypes and scenarios
Policy sensitivity: Tightening child-focused policies substantially increased measured defects across diverse applications

For Risk Practitioners:

Application-level safety evaluation must move beyond single-turn benchmarks to dynamic, policy-aware testing that reflects the objective deployment complexity. Enterprise deployments should adopt comparable frameworks to ensure safety and durability in production.

Bias Research & Fairness Mitigation

(Source- Stanford Report)

Stanford Report identified pervasive biases against older women in generative AI outputs, including systematic resume downgrading- highlighting inconsistent mitigation even in state-of-the-art models.

Systematic Fairness Review: Categorized interventions into three layers:

Preprocessing: Data augmentation, rebalancing, resampling
In-Processing: Fairness constraints during training, adversarial debiasing
Post-Processing: Output adjustment to reduce discrimination

Implementation Reality: Effective bias mitigation typically requires approaches across all three categories, yet implementations remain fragmented by domain and organizational context.

Compliance & Risk Considerations For Executives

Critical Action Items (Q4 2025 - Q1 2026)

For EU-Based Operations:

Audit AI systems for GPAI Act compliance (deadline: August 2, 2027, for legacy models)
Document training data sourcing and copyright compliance mechanisms
Establish systemic risk assessment protocols for high-FLOPs models
Prepare for potential penalties: Article 99-100 enforcement live since August 2025

For India-Exposed Operations:

Monitor AI Safety Institute rollout and incident classification frameworks
Assess eligibility for regulatory sandboxes
Map AI applications against sector-specific risk guidance (financial services, healthcare guidance expected Q4 2025/Q1 2026)

For All Organizations:

Implement dynamic safety evaluation: Application-level frameworks capturing multi-turn, context-aware risks
Establish bias audit infrastructure: Systematic testing across demographic and application dimensions
Build incident reporting systems: Internal equivalents to proposed national databases
Align governance with business units: Shift AI oversight from IT-only to business-unit accountability, paired with compliance frameworks.

Looking Ahead

Model Releases & Enhancements

Continued frontier model refinements targeting agentic capabilities (reasoning, memory, tool use orchestration)
Open-source competitive pressures are yielding cost-optimized variants across 7B-40B parameter ranges.
Multimodal video generation capabilities (Sora 2 integration) are likely to influence enterprise marketing and content automation use cases

Regulatory Evolution

India's AI Incident Database launch and initial sector-specific risk classifications (target: Q1 2026)
EU codes of practice finalization for GPAI compliance (enforcement agency coordination ongoing)
Potential U.S. executive order updates aligned with European trends

Enterprise Adoption Inflection

Higher failure rates are expected for implementations lacking workflow redesign.
Board-level accountability for AI ROI will drive harder vendor evaluation and internal capability building.
Agentic AI for IT service desks, knowledge management, and back-office automation is gaining traction.

Research & Validation Priorities

Application-level safety testing frameworks becoming standard practice (not optional)
Bias evaluation shifting from binary fairness metrics to context-aware policy alignment.
Federated and edge-based AI models are gaining adoption in regulated sectors.

To wrap things up

November 2025 marks a pivotal moment for enterprise AI: governance frameworks are now mandatory, frontier models serve agentic workflows, not just chatbots, and scaling remains stubbornly difficult. Organizations should expect compliance costs to rise, safety evaluation to become procedural, and competitive advantage to depend more on workflow redesign than model choice.

The convergence of Indian flexibility, European strictness, and corporate pragmatism suggests a middle path: governance that enables scaled AI while embedding safety and transparency into technical architecture itself.

Stay informed with our regulatory updates and join us next month for the latest developments in risk management and compliance!

For any feedback or requests for coverage in future issues (e.g. additional countries or topics), please contact us at info@riskinfo.ai. We hope you found this newsletter insightful.

Best regards,

The RiskInfo.ai Team