Divya van Mahajan

16 Ways Enterprise AI is Changing

ai-daily-brief-podcast

16 Ways Enterprise AI Is Changing (2025) — Study Document

Overview

This episode of the AI Daily Brief podcast/video channel provides a detailed walkthrough and commentary on Andreessen Horowitz’s (a16z) annual enterprise AI report: “16 Changes to AI in the Enterprise — 2025 Edition.” The host, affiliated with both the AI Daily Brief and a firm called Superintelligent, cross-references a16z’s findings with first-hand observations from enterprise AI consulting and deployment work. The talk matters because it offers a data-driven snapshot of how large organizations are actually adopting, budgeting, and operationalizing AI in mid-2025 — moving well beyond the experimental phase into structural transformation.

Source video URL: Not available (internal/podcast recording dated 2025-06-17)

Prerequisites

  • Basic familiarity with Generative AI (GenAI) concepts: large language models (LLMs), prompting, fine-tuning, context windows
  • Understanding of enterprise software procurement models (SaaS, seat-based licensing, usage-based pricing)
  • Awareness of major AI model providers: OpenAI, Anthropic, Google (Gemini), Meta (Llama), Mistral
  • Familiarity with the concept of AI agents and agentic workflows (multi-step, autonomous AI task completion)
  • General knowledge of the build vs. buy decision framework in enterprise software
  • Awareness of the 2024 a16z enterprise AI report as prior context

Main Points

1. AI Budgets Are Larger Than Expected and Still Growing

  • Enterprise leaders expect an average 75% growth in AI spend over the next year.
  • Growth is use-case driven, not top-down budget mandates — new applications are creating demand for more spend.
  • A PwC study cited by the host found 88% of organizations increased AI budgets due to agentic AI, with ~75% increasing by 10% or more.

2. Spend Is Shifting to Permanent Budget Lines

  • In 2024, innovation budgets accounted for ~25% of GenAI spend; in 2025 that dropped to just 7%.
  • Reallocated central IT budget grew from 28% to 39%; business unit budgets rose from 21% to 27%.
  • This signals GenAI is graduating from experimental line items to core operational expenditure.
  • Agent pilots have a distinctly different character than prior GenAI pilots — they are treated as infrastructure investments, not “if” experiments.

3. Enterprises Are Behaving More Like Consumers — Using Multiple Models

  • Organizations are selecting models based on task-specific strengths, mirroring sophisticated consumer behavior.
  • Example differentiation: Claude for fine-grained code completion; Gemini for high-level systems architecture (partly due to larger context window).
  • 37% of enterprises now use five or more models, up significantly from prior years.

4. Market Consolidation Around a Few Leaders, With Nuanced Share

  • OpenAI retains overall market share leadership.
  • Google and Anthropic made significant gains over the past year.
  • 67% of OpenAI users have deployed non-frontier models in production; only 41% for Google, 27% for Anthropic — meaning Anthropic and Google users concentrate at the frontier.
  • Google gained disproportionately among large enterprises, driven by compliance trust and price-performance ratio.

5. Cost Is Now a Primary Buying Consideration

  • Closed-source, non-frontier models have dropped so dramatically in price that enterprises are increasingly choosing them over open-source alternatives, retaining ecosystem benefits.
  • In buying decisions, cost of ownership rose most as a consideration, while reasoning and accuracy/reliability considerations declined relative to prior year.
  • One enterprise quoted: “Gemini is cheap.”
  • The pricing pressure on closed-source models complicates the long-term open-source vs. closed-source competitive dynamic.

6. Fine-Tuning Is Viewed as Increasingly Unnecessary

  • Longer context windows and improved base model capabilities mean enterprises can achieve near-equivalent results by loading training data into context rather than fine-tuning.
  • Quote from enterprise: “Instead of parameter-efficient fine-tuning, you just dump it into a long context and get almost equivalent results.”
  • This has significant financial implications for startups positioned around enterprise fine-tuning services.

7. Reasoning Models Are Opening New Use Cases

  • 23% of enterprises are already using OpenAI’s O3 reasoning model in production.
  • 57% of respondents said reasoning models are accelerating their adoption.
  • Enterprises view reasoning models not just as incremental improvements but as enablers of previously impossible use cases.
  • Quote: “Reasoning models allow us to solve newer, more complex use cases.”

8. AI Procurement Is Maturing to Resemble Traditional Enterprise Software Buying

  • The buying process now includes checklists and price sensitivity typical of traditional software procurement.
  • Shift from innovation-driven evaluation (accuracy, reasoning benchmarks) to total cost of ownership as the dominant criterion.
  • Reflects the transition from pilot/exploration to ubiquitous organizational use.

9. Enterprises Are More Comfortable Hosting Directly With Model Providers

  • In 2024, most enterprises accessed models through existing cloud providers (AWS, Azure, GCP).
  • In 2025, companies increasingly go directly to model providers (OpenAI, Anthropic) to access the latest models faster.
  • A growing pain point: the gap in model quality between what employees use in personal/consumer accounts and what their employer provides through standard IT channels.
  • Example cited: Amazon reportedly considering replacing its internal AI code tool with Cursor due to employee demand.

10. Switching Costs Are Rising as Agentic Workflows Deepen

  • In 2024, enterprises deliberately designed applications to minimize switching costs and keep models interchangeable.
  • In 2025, agentic workflows with multiple interdependent steps make model substitution significantly harder — changing one model component can affect all downstream dependencies.
  • Agent platforms are racing to capture broader shares of enterprise agentic use cases to benefit from this lock-in dynamic.
  • The host predicts market forces will ultimately compel agent interoperability standards, despite current competitive fragmentation.

11. External Benchmarks Rising as Evaluation Proxies (Likely Temporary)

  • Use of external benchmarks as the primary model evaluation method increased; internal and project-specific benchmarks declined slightly.
  • The host flags this as potentially temporary: public benchmarks are increasingly saturated at the top and lack discriminatory power.
  • Prediction: by the next annual survey, internal benchmarks will return as a major evaluation criterion, driven by advances in model evaluation (“evals”) methodology visible in developer communities like the AI Engineer Summit.

12. Build vs. Buy Is Shifting Toward Buying — With Important Caveats

  • In 2024, enterprises built by default because quality vertical AI applications didn’t exist.
  • In 2025, a marked shift toward buying third-party applications as the AI app ecosystem matures.
  • However, the host argues a bifurcation is emerging:
    • Non-regulated industries / common use cases (e.g., CPG customer service): off-the-shelf-ish solutions with light customization will suffice.
    • Regulated, high-value industries (finance, healthcare): custom-built agentic solutions will remain the default.
  • Even “buying” still requires data integration and customization — the classic off-the-shelf model doesn’t fully translate to AI agents.

13. Outcome-Based Pricing Is Promising but Struggling

  • Traditional per-seat pricing is widely seen as ill-suited for AI.
  • Early outcome-based pricing experiments are nascent and problematic.
  • Only 15% of CIOs prefer outcome-based models; 39% prefer usage-based, 21% still prefer seat-based.
  • Biggest challenges with outcome-based pricing:
    • Lack of clear measurable outcomes: cited by 47% of respondents
    • Unpredictable and unscalable costs: cited by 36%

14. Software Development Has Become a Default Enterprise AI Use Case

  • The percentage of enterprises with software development as an in-production AI use case jumped from under 40% to over 70% in one year.
  • Driven by: high-quality off-the-shelf tools, improved model capabilities, broad industry relevance, and clear ROI.
  • Other growing use cases: enterprise search, data analysis, data labeling.
  • Customer service saw a slight decline in in-production deployment.

15. Employee Consumer Behavior Is Driving Enterprise Purchasing Decisions

  • Much of the growth in enterprise AI app adoption is driven by prosumer/consumer market familiarity.
  • Example: Many CIOs chose to purchase enterprise ChatGPT because “employees loved ChatGPT” — brand recognition drove procurement.
  • Bottom-up employee demand (e.g., for Cursor) is increasingly shaping enterprise AI tool selection.

16. AI-Native Companies Are Beginning to Outpace Incumbents on Quality and Speed

  • Early AI adoption favored incumbents due to: established trust, existing distribution, and ability to fund large capital requirements (e.g., Microsoft → OpenAI, Google → Anthropic).
  • This dynamic is shifting: enterprises increasingly prioritize access to the best, most current models as fast as possible.
  • The coding assistant space illustrates this vividly: agentic-native tools like Cursor have dramatically outpaced first-generation tools like GitHub Copilot.
  • Enterprises preferring AI-native companies cite faster pace of innovation as the overwhelming reason.

Key Concepts

  • Agentic AI / Agentic Workflows: AI systems that autonomously execute multi-step tasks, often involving tool use and sequential decision-making, as distinct from single-prompt interactions.
  • Reasoning Models: LLMs with enhanced deliberative or chain-of-thought capabilities (e.g., OpenAI O3) that enable more complex, multi-step problem solving.
  • Fine-Tuning: A training process that adapts a pre-trained model to a specific task or domain using additional data; increasingly being replaced by long-context prompting.
  • Context Window: The maximum amount of text/tokens an LLM can process in a single interaction; larger windows reduce the need for fine-tuning.
  • Outcome-Based Pricing: A pricing model where payment is tied to measurable business results rather than usage volume or seats.
  • Usage-Based Pricing: A pricing model where cost scales with the volume of API calls or tokens consumed.
  • Prosumer Market: Consumers who use a product professionally or at a high level, bridging consumer and enterprise use — their preferences often influence enterprise procurement.
  • Switching Costs: The friction, effort, and risk associated with replacing one AI model or platform with another in an existing workflow.
  • External Benchmarks: Standardized, publicly available tests used to compare AI model performance (e.g., MMLU, HumanEval); increasingly seen as insufficient for enterprise differentiation.
  • Evals (Evaluations): Custom or systematic methods for assessing AI model performance on specific tasks or workflows relevant to a particular enterprise use case.
  • Build vs. Buy: The strategic decision of whether to develop custom AI solutions in-house or purchase third-party applications.
  • Vertical Agents: AI agents designed for specific industries (e.g., healthcare, finance) as opposed to general-purpose tools.
  • Functional Agents: AI agents designed for specific business functions (e.g., HR, legal, customer service) regardless of industry.

Summary

The a16z 2025 Enterprise AI report, as analyzed through the lens of the AI Daily Brief, paints a picture of enterprise AI transitioning from experimentation to structural integration. Budgets are growing rapidly and moving into permanent IT and business unit lines, driven primarily by the emergence of agentic workflows and expanding use cases rather than top-down mandates. Enterprises are becoming more sophisticated consumers of AI — using multiple models based on task-specific strengths, demanding direct access to frontier models, and evaluating purchases with cost-of-ownership discipline typical of mature software procurement. Key shifts from 2024 include the decline of fine-tuning as a required practice, the rise of reasoning models opening genuinely new capabilities, increasing switching costs as agents deepen organizational integration, and a notable move toward buying third-party AI applications as the vertical ecosystem matures — though heavily regulated industries will likely continue building custom solutions. Pricing models remain unsettled, with outcome-based approaches aspirational but practically immature. The overarching narrative is one of acceleration: enterprises are adapting faster than expected, behaving more like sophisticated consumers, and — for the first time — beginning to favor AI-native companies over incumbents in key categories like software development, signaling that speed of innovation is displacing established trust as the primary competitive differentiator.