Menlo Ventures State of GenAI in the Enterprise 2025: What Enterprise Buyers Are Actually Spending

Executive Summary

• Enterprise AI spend grew 3.2x in a single year — from $11.5B in 2024 to $37B in 2025 — making this the fastest-scaling software category on record. Budgets are no longer speculative.
• Buying beat building decisively: 76% of AI use cases in 2025 were purchased rather than built, up from 53% in 2024. The CTO argument for custom internal builds is weakening with each passing quarter.
• AI deals convert to production at 47% — nearly double traditional SaaS’s 25%. This data challenges the prevailing narrative that AI pilots rarely graduate; the failure rate is lower than expected, but the failure mode has shifted to post-production value capture.
• Coding dominated departmental spend at $4.0B out of $7.3B total departmental AI — and developers at top-quartile organizations now use AI tools daily at a 65% rate. No other function comes close to this adoption depth.
• True agentic deployment remains rare: only 16% of enterprise AI deployments qualify as genuine agents that plan, execute, and adapt. The remaining 84% are fixed-sequence workflows or routing systems marketed as agents.

What Enterprises Are Actually Spending

The $37B figure is real enterprise budget committed — not investor capital, not vendor revenue projections. Menlo estimates this from a combination of survey data (495 U.S. enterprise decision-makers, November 7–25, 2025) and proprietary revenue data on private companies. The methodology is more transparent than typical consulting survey estimates, though private company revenue estimates carry uncertainty.

The stack breaks down cleanly:

The 51/49 split between applications and infrastructure is notable. Infrastructure spending is foundational and sticky — once an organization’s data pipeline runs on a particular orchestration stack, ripping it out is a multi-quarter project. Application spending is where competitive differentiation happens and where the buying decisions most relevant to the 200–2,000 employee company occur.

The Buy-vs.-Build Data Has Settled

In 2024, the enterprise AI market was roughly split: 47% build, 53% buy. By 2025, 76% of use cases were purchased. This is a structural shift, not a swing back.

The explanation is not that enterprise developers suddenly stopped building — AI coding tools are growing 7.3x year-over-year. The explanation is that purpose-built AI applications have reached a quality threshold where the integration cost of building from scratch on foundation model APIs exceeds the cost of buying a vendor solution that already handles data plumbing, prompt engineering, evaluation, and monitoring.

For the mid-market CFO, this simplifies the budget conversation: the question is no longer “should we build our own AI system” but “which vendor’s AI system fits this workflow, at what price, with what exit rights.” The 47% production conversion rate for AI deals — nearly double traditional SaaS’s 25% — suggests that when buyers do select a vendor, the rate at which that selection turns into a production deployment is higher than any prior software category. The risk of buying has dropped. The risk of choosing the wrong vendor has risen.

Startups captured 63% of application revenue in 2025, reversing 2024’s 64/36 incumbent advantage. In specific departments the startup dominance is extreme:

This is a procurement signal: the established vendors (Salesforce, SAP, Workday, ServiceNow) are losing AI application share to companies that did not exist five years ago. Lock-in to a legacy vendor’s AI roadmap is not automatically protection — it may be exposure.

Coding: The First Category Where AI Has Won

Developer tools consumed $4.0B of enterprise AI spend in 2025, a 7.3x increase from 2024’s $550M. Half of all developers now use AI coding tools daily; in top-quartile engineering organizations, that figure reaches 65%.

This is not adoption theater. The Menlo data shows teams reporting 15%+ velocity gains — a figure directionally consistent with the 26% gain found in academic RCTs on AI-assisted programming (Peng et al., 2022), though Menlo’s figure is self-reported rather than RCT-measured and almost certainly reflects optimistic recall.

The competitive story inside developer tools is instructive for every category: Cursor overtook GitHub Copilot despite GitHub’s first-mover advantage, brand recognition, and Microsoft’s distribution muscle. The winning differentiators were repo-level context, multi-file editing, and workflow integration — not model quality alone. The lesson holds for enterprise AI purchasing generally: the incumbent with the largest installed base does not automatically win the AI layer.

Where the Agent Hype Meets the Production Reality

“Agentic AI” is the most overstated category in enterprise marketing. The Menlo data is direct: only 16% of enterprise AI deployments qualify as true agents — systems that plan, execute, and adapt based on observed results. The remaining 84% are fixed-sequence workflows or routing systems. Startups deploy genuine agents at a higher rate (27%), but the gap between marketing language and production reality is consistent across the market.

The $750M in agent spend (2025) is real, but it represents 10% of horizontal AI spend, not the majority. Copilots — fixed-context assistants that augment human decisions without autonomous action — still account for $7.2B, or 86% of horizontal AI.

For a CIO benchmarking their organization’s deployment sophistication, the implication is this: if the internal team claims to be running “agents,” probe the definition. Does the system select its own tools? Does it evaluate its outputs and retry? Does it decompose multi-step problems without a human-defined sequence? Most enterprise AI systems in production today answer “no” to at least two of these three questions.

The LLM Market Share Shift — With a Caveat

Menlo reports Anthropic at 40% of enterprise LLM API share, OpenAI at 27%, Google at 21%. Anthropic’s growth from 12% (2023) to 40% (2025) is striking.

The caveat is mandatory: Menlo Ventures is an investor in Anthropic. This creates a structural incentive to report findings that reflect well on Anthropic’s market position, even if the survey methodology itself is independent. The market share figures may be accurate — other signals (Claude Code’s benchmark dominance, developer community adoption) are consistent with Anthropic gaining share. But any analysis that uses these numbers should note the investor relationship.

The open-source story is cleaner: enterprise adoption of open-source LLMs fell from 19% to 11% in one year. This contradicts the narrative that open-weight models (Llama, Mistral, DeepSeek) would capture enterprise workloads at scale. The primary failure mode is not performance — it is support, compliance attestation, and the total cost of running and maintaining proprietary infrastructure. Chinese open-source models represent 1% of enterprise LLM API usage despite strong developer-community traction.

Key Data Points

*Menlo Ventures is an investor in Anthropic. Interpret LLM market share figures with this disclosure in mind.

Cross-Reference Against the Existing Corpus

Confirms:

• The buy-vs.-build tilt toward purchasing is consistent with Futurum’s finding (1H 2026, n=830) that 65.9% of enterprises are on integrated platforms and 41% are actively reducing application count.
• The 47% AI-to-production conversion rate is higher than the prevailing Accenture narrative (“90% of pilots never advance”), but the two figures measure different things: Menlo measures deals that reach production, Accenture measures pilots that reach enterprise scale. Both can be true.
• Agent marketing versus agent reality: the 16% true-agent figure is consistent with OutSystems’ finding (96% claiming agent deployment, 94% reporting sprawl) — most organizations conflate workflow automation with agency.

Contradicts or complicates:

• McKinsey (November 2025, n=1,993) finds only 6% of companies are AI high performers capturing >5% EBIT. Menlo’s 47% production conversion rate does not contradict this — converting to production is not the same as capturing measurable EBIT impact. The two datasets are measuring different stages of the same journey.
• Writer/Workplace Intelligence found only 29% see significant ROI. Again, consistent: production deployment is earlier in the value chain than significant ROI. Most of the 47% that reach production have not yet captured the full value from what they deployed.
• The 3.2x spend growth in one year is the strongest evidence against the “AI fatigue” narrative circulating in late 2025. Budget is increasing, not contracting, even as individual project ROI remains contested.

What This Means for Your Organization

The spend data answers the “is this real” question definitively. Enterprise AI is not speculative budget — it is $37B in committed enterprise spend growing at 3.2x annually. The question for any CIO or CFO is not whether to participate but how to deploy capital where the 47% production-conversion advantage applies and avoid the categories where the market is still sorting itself out.

Three decisions the data sharpens:

First, developer tools should be a near-certain investment for any organization with an engineering function. At 50% daily adoption among developers and 15%+ self-reported velocity gains, this is the most de-risked category in enterprise AI. The question is which tool — and the data suggests that first-mover advantage does not protect incumbents (Copilot vs. Cursor), so vendor selection should prioritize workflow fit over brand recognition.

Second, the 76% buy-vs.-build shift validates a vendor-first procurement posture for most AI use cases, but the 91% startup share in finance and operations means the vendor is unlikely to be the organization’s existing ERP or CRM provider. Budget conversations that assume Salesforce Agentforce or SAP Business AI will handle all AI needs are misaligned with where enterprise spend is actually going.

Third, the 16% true-agent figure is a useful governance checkpoint. Before approving budget for any “agentic AI” initiative, a CFO can ask three questions: Does this system select its own tools? Does it evaluate its own outputs? Does it operate without a human-defined sequence? If the answer to all three is no, the initiative is a workflow automation — which may still be valuable, but should be scoped and priced accordingly.

If the Menlo data raised specific questions about how these spend patterns apply to your organization’s AI budget cycle, that is a conversation worth having directly — brandon@brandonsneider.com.

Sources

Primary:

• Menlo Ventures, “2025: The State of Generative AI in the Enterprise,” n=495 U.S. enterprise decision-makers (C-suite, VPs of Engineering and Product, technical leaders responsible for AI purchasing), survey conducted November 7–25, 2025, published December 9, 2025. URL: https://menlovc.com/2025-the-state-of-generative-ai-in-the-enterprise/
  • Credibility: MEDIUM-HIGH — Independent VC-sponsored survey with disclosed methodology; buyer-panel respondents rather than vendor customers; transparent sample size and survey window. Deducted from HIGH due to (1) Menlo is an investor in Anthropic, whose market share figures are prominently featured; (2) private company revenue estimates cannot be independently verified; (3) U.S.-only scope; (4) survey participants are “companies actively using AI tools,” excluding non-adopters. Temporal tier: TIER 1 (November 2025 fieldwork, December 2025 publication).

Press coverage used for triangulation:

• GlobeNewswire press release, December 9, 2025. https://www.globenewswire.com/news-release/2025/12/09/3202258/0/en/Menlo-Ventures-2025-State-of-Generative-AI-Report-Enterprise-Investment-Hit-37B-in-2025-Tripling-in-One-Year.html
• Yahoo Finance, same date.
• NoJitter summary. https://www.nojitter.com/ai-automation/menlo-ventures-estimates-19-billion-in-gen-ai-spend-during-2025

Cross-reference corpus:

• McKinsey State of AI, November 2025 (n=1,993) — see research/01-ai-native-landscape/mckinsey-state-of-ai-november-2025.md
• Futurum Group 1H 2026 AI ROI Survey (n=830) — see research/01-ai-native-landscape/futurum-enterprise-ai-roi-1h-2026.md
• BCG AI at Work 2025 — see research/01-ai-native-landscape/bcg-ai-at-work-2025.md
• OutSystems AI agent deployment data — referenced in wiki/assistive-to-agentic-shift.md

Brandon Sneider | brandon@brandonsneider.com April 2026