Menlo: Enterprise AI Spend Hits $37 Billion

TL;DR: Menlo Ventures’ state of generative AI report puts enterprise AI spend at $37 billion in 2025, up from $1.7 billion in 2023 (Menlo Ventures, 2025). More than half of that, $19 billion, went to applications people actually use day to day. The takeaway: knowledge-work software backed by AI is now a real software category with real revenue, not a pilot-stage experiment.

For most of 2023 and 2024, the honest answer to “is enterprise AI real revenue or a demo?” was “mostly demo.” That has changed. The menlo ventures state of generative ai enterprise spend 37 billion figure is the clearest signal yet that buyers are paying production budgets for tools that read, write, and reason over their work.

The number jumped from $1.7 billion in 2023 to $11.5 billion in 2024 to $37 billion in 2025, a 3.2x increase year over year (Menlo Ventures, 2025). The report is based on a survey of 495 U.S. enterprise decision-makers combined with a bottom-up model of the market.

Why does the jump from $1.7B to $37B matter?

A 20x move in two years is not a hype curve flattening out. It is a new line item appearing on enterprise budgets and growing fast. Menlo Ventures calls generative AI the fastest-scaling software category in history, now capturing about 6% of the global SaaS market within three years of ChatGPT’s launch (Menlo Ventures, 2025).

The detail that matters most is where the money lands. The application layer (the user-facing software built on top of foundation models) took $19 billion, more than half of all 2025 spend (Menlo Ventures, 2025). Enterprises are not just renting raw model capacity. They are buying finished tools that do a job.

That is the difference between a research budget and a software category. By Menlo’s count, at least 10 products now generate over $1 billion in annual recurring revenue and 50 generate over $100 million (Menlo Ventures, 2025). Categories form when many vendors can sustain real revenue. That is what the data shows.

Where does the application-layer spend go?

Menlo Ventures splits the $19 billion application layer into three buckets (Menlo Ventures, 2025):

• Horizontal AI ($8.4 billion): general copilots and assistants used across functions, led by ChatGPT Enterprise, Claude for Work, and Microsoft Copilot.
• Departmental AI ($7.3 billion): tools built for a specific role, like coding or sales. This bucket grew 4.1x year over year, with coding alone at $4.0 billion.
• Vertical AI ($3.5 billion): industry-specific software, where healthcare leads at roughly $1.5 billion, about 43% of vertical spend.

Two of those three buckets, departmental and vertical, total $10.8 billion of spend aimed squarely at knowledge work: writing code, drafting documents, researching accounts, handling clinical paperwork. These are not infrastructure bets. They are tools that read and act on a company’s own information.

Buying is winning over building

In 2024, enterprises were split: 53% of AI solutions were purchased and 47% built in-house. In 2025, 76% of use cases were purchased rather than built (Menlo Ventures, 2025). Ready-made tools reach production faster and prove value sooner, so the build-it-yourself instinct is fading.

Adoption also moves differently than older software. Menlo found that AI deals convert to production at 47%, nearly double the 25% rate for traditional SaaS (Menlo Ventures, 2025). And 27% of AI application spend arrives through product-led growth, where an individual adopts a tool before procurement gets involved, roughly 4x the 7% rate in traditional software. Counting “shadow AI” on personal cards, that PLG share may approach 40%; one input here is an NBER finding that about 27% of ChatGPT usage is work-related (NBER, 2025).

The practical effect: tools spread bottom-up, often before central IT has connected them to anything. That is how you get a fast-growing category and a fragmentation problem at the same time.

The connective layer is still thin

Here is the gap inside the boom. Menlo reports that only 16% of enterprise deployments and 27% of startup deployments qualify as true agents that plan, act, observe, and adapt. Most production systems are still fixed-sequence or routing workflows wrapped around a single model call, with prompt design as the dominant technique and retrieval-augmented generation (RAG) the next most common (Menlo Ventures, 2025).

Spend backs that up. Of the $18 billion infrastructure layer, only about $1.5 billion went to the data and orchestration tools that manage storage, retrieval, and the connections between LLMs and enterprise systems (Menlo Ventures, 2025). The layer that gives an AI tool access to a company’s actual knowledge is the smallest slice of the stack.

That mismatch is the opening. Companies are buying many AI applications quickly, each landing in its own tool with its own slice of context. RAG (retrieval-augmented generation, where a model pulls in relevant documents before answering) only works as well as the connections feeding it. A unified semantic layer, a knowledge graph that links people, documents, projects, and the tools they live in, is what lets retrieval and agents reason across systems instead of inside one app. This is the problem SemanticOS works on: an operational brain that connects fragmented enterprise tools so people and AI agents can find and reason over institutional knowledge.

A concrete example

Vantage Health, a fictional regional health system, follows the report’s pattern closely. Its clinicians adopt an ambient scribe, a clear win in the $1.5 billion vertical-healthcare market. Its revenue-cycle team buys a separate AI tool for prior-authorization paperwork. Engineering runs Claude Code. Each tool delivers on day one, which is exactly why the deals converted fast.

Six months in, a care-coordination lead asks a simple question: what was the documented exception for a specific payer last quarter, and which clinician approved it? The answer exists. It sits across the scribe’s notes, the auth tool’s records, and an email thread none of those tools can see. Each AI app is smart inside its own walls and blind to the others.

A connective knowledge layer changes the shape of that question. When the scribe notes, the auth records, the clinician directory, and the email thread are linked as entities in one graph, the coordinator’s question becomes a single query that traverses all four sources. The agent retrieving the answer is only as good as the connections beneath it. That layer, not another standalone app, is what turns a pile of point tools into something that reasons across the whole organization.

Key takeaways

• Menlo Ventures puts 2025 enterprise generative AI spend at $37 billion, up from $1.7 billion in 2023, about 6% of the global SaaS market (Menlo Ventures, 2025).
• More than half ($19 billion) went to the application layer, with $10.8 billion of that aimed at knowledge-work roles and verticals. This is a real software category, not a pilot.
• Buying beat building: 76% of use cases were purchased in 2025, and AI deals converted at 47% versus 25% for traditional SaaS.
• The smallest part of the stack is the connective layer, the data and orchestration tools that link models to enterprise knowledge, at roughly $1.5 billion.
• That gap is the work ahead: a semantic layer or knowledge graph is what lets fast-growing AI tools reason across fragmented systems instead of inside one app.