Enterprise AI Hit $177B: Agents Now Drive 80% of Model Tokens
Enterprise AI spending reached $176.8 billion by September 2026, up 5.4x in a single year, as autonomous agents transformed how businesses consume and pay for AI. That is the headline finding from Menlo Ventures’ 2026 State of Enterprise AI report, released November 3 and based on a survey of nearly 400 U.S. enterprise decision-makers overseeing more than $4 billion in AI spending.
The numbers carry a simple message for every enterprise AI and GTM leader: the agent economy has arrived, it is bigger than most projections expected, and the competitive landscape at the model layer is consolidating fast around two players.
The Scale of the Shift
Three years ago, enterprise generative AI deployment was a $1.7 billion experiment. It reached $8.5 billion in 2024, $33 billion in 2025, and $176.8 billion by September 2026, an acceleration of 5.4x that actually beat last year’s 3.9x pace. For every dollar the industry spent building AI models, enterprises spent $1.70 putting them to work.
The chart below shows the trajectory:
| Year | Enterprise AI Deployment Spend | Growth |
|---|---|---|
| 2023 | $1.7B | baseline |
| 2024 | $8.5B | 5.0x |
| 2025 | $33B | 3.9x |
| 2026 (to Sep) | $176.8B | 5.4x |
Source: Menlo Ventures, 2026 State of Enterprise AI
What is striking is that this growth is not driven primarily by new buyers. The share of enterprises paying for AI rose only modestly, from 55% to 65%. Growth is coming from existing adopters spending far more. The average enterprise is spending 2.5 times what it budgeted for AI this year.
That overshoot is the real signal. When organizations consistently exceed their AI budgets, it means AI is delivering ROI fast enough to justify pulling forward spend, not just experimenting. Agents are the mechanism driving that dynamic.
Agents Have Taken Over Model Consumption
The most operationally significant data point in the report: agents now generate 80% of all tokens on OpenRouter, roughly seven times human-driven usage. Since overtaking human traffic in February 2026, agent token consumption has grown 37x, compared with 5x growth for human-driven usage over the same period.
This is not a future trend. It is the present reality of how enterprises consume AI. The implications for how you price, build, and deploy AI products are significant:
- Latency and throughput matter more than conversational quality for most token volume.
- Cost per task, not cost per message, becomes the right unit of measurement.
- Reliability and retry logic are product requirements, not edge cases.
- The person approving the AI spend may never see the actual model output.
For enterprise leaders building AI-native workflows, the shift to agent-driven consumption means autonomous GTM systems and similar agentic operations are no longer leading-edge experiments. They are aligned with where the market is already spending.
Anthropic Leads, OpenAI Gains, Google Falls
At the model layer, the report shows concentration accelerating. Frontier labs (Anthropic, OpenAI, and Google) now command 69% of enterprise AI deployment dollars, up from 50% last year. Their spending grew 7.5x year over year, compared with 3.3x for everyone else.
Within enterprise LLM API spending specifically:
| Provider | 2025 Share | 2026 Share | Change |
|---|---|---|---|
| Anthropic | 40% | 55% | +15pp |
| OpenAI | 27% | 31% | +4pp |
| 21% | 9% | -12pp | |
| Other | 12% | 5% | -7pp |
Source: Menlo Ventures, 2026 State of Enterprise AI. Note: Menlo Ventures discloses it is an investor in Anthropic and OpenRouter; figures are estimates, not audited revenue.
Anthropic grew its lead on the strength of its Opus models, which excel at long-horizon agentic tasks and coding. Together, Anthropic and OpenAI now account for 86% of enterprise LLM API spending, up from 67% a year ago.
Google’s sharp drop from 21% to 9% is the clearest warning sign in the report. Despite Gemini 3 Pro reaching frontier performance in late 2025, Google has not matched the flagship releases from Anthropic and OpenAI since, and its API spend is now concentrated in cheaper Flash models. It illustrates how quickly market share can move when model capability diverges.
The near-frontier models (xAI’s Grok 4.7, Moonshot’s Kimi K3, DeepSeek’s V4, Z AI’s GLM-5.3) offer frontier-adjacent performance at lower prices yet together hold under 5% of LLM API spend. The gap between performance and spending share suggests enterprises are paying a premium for trust, reliability, and integration depth, not just raw capability scores.
Where the $77B Application Market Lives
Enterprise AI application spending reached roughly $77 billion in 2026, up from $19 billion in 2025. Two categories dominate:
- Coding: $34 billion (44% of app spend). Claude Code and OpenAI Codex alone account for $24 billion, or 70% of that total. AI-native companies capture 96% of coding spend, the highest AI-native share of any category.
- AI assistants and productivity: $25 billion (32%). Enterprise Claude, ChatGPT, and Gemini make up 62% of this segment.
Together, those two categories represent 76% of application spending. The report attributes this concentration to structure: coding and chatbots provide environments where agents can act, workflows they can learn, and results they can verify. Categories without that structure, where outcomes are subjective or processes are informal, have seen slower AI penetration.
Incumbents retain advantages in IT (91%), data science (88%), and HR (68%). New AI-native entrants are winning where the category is new. The strategic implication: enterprises looking to deploy agentic AI should start in structured workflows with clear success metrics, then expand outward.
The 55% Net New Spend Finding
One of the most discussed findings is that 55% of enterprise AI spending is coming from net new budgets, not existing software or labor allocations. Only 22% came from existing software budgets, and most of that stayed with incumbents converting licenses to AI-enhanced offerings (such as Microsoft 365 to Copilot, or Salesforce to Agentforce).
The report finds little evidence so far of enterprises replacing core systems like Salesforce CRM or Workday HRIS. AI is expanding the software economy, at least for now, rather than cannibalizing it.
For enterprise leaders, this matters for budget conversations. AI initiatives are more likely to gain approval when framed as new capability investment rather than as replacements for existing tools. The data supports that framing.
What This Means for Enterprise AI and GTM Teams
Three practical takeaways for enterprise AI builders and operators:
1. Agent-first architecture is not premature. With 80% of model tokens now agent-generated, designing your AI systems for agent-to-agent communication, structured outputs, and batch throughput is building for current demand, not future demand. Teams still building primarily for conversational interfaces are optimizing for 20% of the market.
2. Model selection is a strategic decision. Anthropic’s move from 40% to 55% of API spend in a single year shows how quickly capability advantages translate into market share. The enterprises choosing their primary model provider today are making a decision with compounding effects. Enterprise adoption readiness gaps often show up here: teams that have not standardized their model stack yet face higher integration costs as the market continues to consolidate.
3. Budget the overshoot. Enterprises are spending 2.5x what they budgeted. This is not a planning failure; it is a signal that AI is delivering enough value to justify pulling forward spend. Build that expected expansion into your planning rather than treating budget overruns as surprises.
The Four-Year Arc
Menlo Ventures has published this report annually since 2023. The arc is instructive: from $1.7 billion and mostly experimentation to $176.8 billion and mostly production agents in four years. The next phase, as the report notes, will likely see the application layer capture more value as agent categories mature and move beyond the labs’ first-party tools.
For enterprise teams building AI-native systems now, the window to build with the capabilities that led to this concentration is still open. Whether it stays open in the same form for another year is the right question to track.
If you are evaluating how to build agentic GTM or operational AI systems for your enterprise, talk to our team at Enera about where the architecture decisions matter most.
Sources: Menlo Ventures 2026 State of Enterprise AI | GlobeNewswire press release | AI Frontier Post analysis | SaaStr commentary