Mid-2025 AI Update: What’s Actually Working in Enterprise
As we cross the midpoint of 2025, the conversation around AI is shifting from potential to practice. While the race to build the next frontier model dominates headlines, the more critical story is one of diffusion—how this technology is actually being woven into the fabric of business. As I recently noted, China is accelerating this process through national strategy. The following list offers a playbook for leaders navigating this transition, outlining the key strategic, technical, and organizational patterns that are separating the leaders from the laggards. Use it to benchmark your progress and refine your company’s AI roadmap.
Readers keep Gradient Flow going. Subscribe free or paid to get new posts and help sustain this work.
Market Dynamics & Strategic Positioning
Model Commoditization
The performance gap between frontier models (GPT, Claude, Gemini) is narrowing rapidly, with competitive open models emerging within 3-6 months of any breakthrough. Foundation models are becoming interchangeable commodities rather than durable competitive advantages.
- Build model-agnostic architectures from day one to avoid vendor lock-in and leverage the best models for each use case.
Vertical Specialization Strategy
Breakout enterprise AI startups like Harvey (legal) and Sierra (customer service) demonstrate that deep domain expertise beats horizontal platforms. These companies win by mastering industry-specific workflows, terminology, and success metrics that generic solutions cannot address.
- For startups: Choose a specific vertical and become the definitive solution for that industry. Domain depth and “speaking the customer’s language” create defensible moats and higher willingness to pay.
Three-Tier Market Structure
The AI ecosystem is stratifying into foundation models (capital-intensive, low-margin), tools/infrastructure (higher margin but commoditizing), and applied AI solutions (highest margin, most defensible).
- For startups: Position in the applied AI layer, using foundation models as commoditized inputs to deliver specialized business outcomes where sustainable margins exist.
Myth of the Model-Only Company
Companies like OpenAI and Anthropic succeed not through models alone but as AI product companies. Their value lies in complete solutions: APIs, security, compliance, governance, and user-facing applications.
- For startups: Differentiate through complete solutions—workflow integration, UX, and business logic—not thin wrappers around public APIs.
The Data Foundation
Modern Data Platform as the Entry Ticket
GenAI’s appetite for unstructured data breaks traditional data warehouses built for neat rows and columns. The core bottleneck for most enterprises is not a lack of models, but a lack of pipelines to feed them with relevant, clean, proprietary data.
- For enterprises: Your AI progress is capped by your data infrastructure’s maturity. Prioritize building a multimodal data platform that can handle unstructured data before scaling your AI initiatives.
Data Quality Over Model Choice
The performance of any AI system, especially those using RAG, is limited by the quality of the data it can access. High-quality, domain-specific data is a more durable competitive advantage than access to any single foundation model. “Garbage in, garbage out” remains the iron law.
- For AI teams: Treat your data pipeline—ingestion, cleaning, and enrichment—as a core product. Your most valuable IP isn’t the model, but the high-quality data flow that feeds it.
Technical Architecture & Implementation
Complete AI Systems Over Pure Models
Reliable enterprise solutions require AI systems that orchestrate foundation models with traditional tools—calculators, APIs, databases, and custom code—to handle reasoning, computation, and data retrieval effectively.
- For AI teams: Design modular architectures where LLMs handle reasoning while specialized tools manage deterministic tasks. Think beyond single-model solutions.
Evaluation-Driven Development as Core IP
Your evaluation framework—comprising representative test cases, clear metrics, and production telemetry—becomes proprietary intellectual property that determines competitive advantage and guides optimization.
- For AI teams: Invest in evaluation architecture before building features. Design evals that define tradeoffs between intelligence, cost, and latency for your specific use case. Treat evals as first-class code.
Architectural Methods Over Fine-Tuning
Advanced prompt engineering, RAG, tool use, and prompt caching often outperform fine-tuning while being more accessible, less expensive, and less risky than “brain surgery on the model.”
- Follow an optimization hierarchy: exhaust prompt engineering and RAG first, turning to fine-tuning only when clear metrics justify the cost and complexity.
While the RAG-first hierarchy is the right starting point for most enterprise applications, it has a performance ceiling. This hierarchy shifts when building specialized agents for high-stakes domains—while architectural methods provide the foundation, post-training techniques become necessary to achieve the reliability and domain-specific reasoning that enterprise applications demand.
Business Models & Economic Impact
Outcome-Based Pricing Revolution
The shift from seat-based to outcome-based pricing (charging only for successful results like tickets resolved or contracts reviewed) represents a fundamental disruption that aligns vendor incentives with customer value.
- For startups: If your product delivers measurable outcomes, price based on those outcomes. This model is compelling to customers but requires deep accountability and robust measurement.
Labor Budget Capture
AI’s true economic impact comes from capturing budgets previously allocated to human labor, not just displacing software spend. This expands addressable markets by orders of magnitude beyond traditional software categories.
- For startups: Size opportunities by total process costs including human labor. Your TAM isn’t today’s software market—it’s the entire labor budget for the function you’re automating.
Core vs. Context Strategic Framework
Geoffrey Moore’s framework is central to AI strategy: Core capabilities create competitive differentiation; Context functions are necessary but non-differentiating (like basic HR systems).
- For enterprises: Buy “context” AI solutions and focus internal development on “core” AI capabilities that create unique competitive advantages. Building your own HR system wastes energy; building proprietary AI for wealth management or drug discovery creates durable advantage.
Enterprise Adoption & Go-to-Market
Production-Ready Solutions Over Demos
While enterprises show universal interest in AI, 42% abandon pilots due to reliability and governance concerns. The gap between experimentation and production deployment remains the primary challenge.
- Focus on solving real problems reliably with governed, production-ready solutions. Impressive demos don’t convince potential users—trustworthy, secure solutions do.
Problem-Focused Selling
Successful enterprise AI sales emphasize business outcomes and customer value in their terminology, not technical capabilities or model performance metrics.
- For startups: Research customer businesses deeply, understand specific pain points, and articulate value in customer language. Lead with business impact and prove it through focused proofs of value.
Data Governance as Core Feature
Access controls, guardrails, data classification, and audit trails aren’t afterthoughts—they’re core features that determine enterprise adoption success and are often the primary deployment blocker.
- For AI teams: Address identity management and data classification before deployment. Productize audit trails and red-team reports as part of your core feature set to accelerate security team alignment.
Use Cases & Agentic Workflows
Rise of Autonomous Agents
The market is evolving from simple AI tasks to multi-step, autonomous “agentic workflows” that complete entire business processes end-to-end — in areas like sales, coding, customer support, and document processing.
- For startups: Identify high-value, multi-step business processes and automate them completely. Move from selling “tools” to selling “outcomes” with fundamentally better economics.
Three-Phase Enterprise Evolution
Enterprises typically move through three phases: pilots, selective rollout, and broad adoption. Many stall at the pilot stage due to organizational friction rather than technical limitations.
- For AI teams: Ship solutions that help customers break through adoption bottlenecks. Budget time for policy alignment, enablement, and measurement alongside technical development.
Organizational & Workforce Transformation
AI-Native Workforce Expectations
A new generation of employees accustomed to ChatGPT and similar tools expects AI-enhanced work environments. Companies unable to provide AI-native experiences face recruitment and retention challenges.
- For enterprises: Deliver consumer-grade UX with enterprise-grade controls. Leverage users’ existing AI fluency with familiar conversational interfaces.
Workforce Flattening and Role Evolution
AI tools enable individual contributors to handle broader responsibilities, blurring traditional role boundaries and potentially flattening organizational hierarchies.
- For AI teams: Design applications that enable role expansion and cross-skilling rather than simple productivity improvements. Consider impacts on organizational structure and accountability.
Change Management Complexity
While engineering teams adopt AI quickly, legal, security, and procurement operate on quarterly cadences, creating deployment bottlenecks despite CEO enthusiasm for AI initiatives.
- For enterprises: Treat policy alignment, enablement, and change management as first-class requirements. Understand that organizational readiness often gates deployment more than technical capabilities.
Trends
Open Source Acceleration
Open-weights models now trail proprietary breakthroughs by only months, fundamentally changing competitive dynamics and forcing vendors toward more open approaches.
- Standardize on open models for most workloads, reserving proprietary models only for specialized edge cases requiring highest quality or lowest latency.
Zero-Friction Infrastructure
Current AI adoption builds on two decades of cloud, SaaS, SSO, and API infrastructure, creating an environment where adding AI is just another integration rather than a fundamental rebuild.
- For startups: Leverage existing infrastructure rather than rebuilding foundations. Compete on domain fit and workflow intimacy, not plumbing. Differentiation has moved up the stack.
