Menlo Ventures: agents account for 80% of enterprise AI use as market grows to $177 billion

Until recently, the word “agent” was a trendy addition to startup pitch decks, while AI sat under “experiments” in corporate budgets. If Menlo Ventures’ latest report is to be believed, that has changed: AI has become a significant expense, and chatbots are no longer doing most of the work.

What the report claims

The State of Enterprise AI report puts corporate AI spending at $177 billion. A year earlier, in 2025, the figure was $33 billion. Menlo partner Deedy Das told Bloomberg Tech that agents now account for 80% of usage. Businesses, he said, are moving away from chatbots toward systems that carry out complex tasks on their own. He also noted that Anthropic and OpenAI dominate corporate API spending.

An important caveat: all we have is a brief recap of the broadcast, without the report’s methodology. It is unclear what that 80% measures: tokens, requests, spending, or deployments. This is no minor detail. An agent makes many model calls for a single task, so its share of token usage will almost inevitably be enormous, even if chatbot deployments still outnumber agent deployments.

A simple example of a chatbot versus an agent

A chatbot follows a “one question, one answer” pattern. An employee pastes in a customer’s email and gets a draft reply. The model is called once; the human reviews the response and sends it.

An agent is given the whole task: “handle this return request.” It opens the CRM, checks the order history, reviews the return policy, writes to the customer, and updates the status on its own. This takes dozens of model calls and several external tools. The arithmetic is straightforward: user numbers may stay the same while the API bill multiplies. This shift probably explains part of the growth from $33 billion → $177 billion, rather than simply more people opening a chat.

What practitioners should do

For developers. Measure the cost per task, not per request. An agent that takes an average of 30 steps needs limits on both step count and token budget; otherwise, a single workflow stuck in a loop can burn through a month’s quota. Log every step and maintain a rollback option: when a system writes to the CRM on its own, a mistake ends up in actual customer records.

For product managers. The value now lies in a process completed from start to finish, rather than an attractive chat window. The question for any feature is: what task does it handle end to end, and how will we measure whether it completed it correctly?

For founders. The market is concentrating around two model providers, and that creates a risk: they set the prices, limits, and terms. Build an abstraction layer over the model so switching providers does not require rewriting the product. The niche for “yet another chatbot built on an API” looks increasingly difficult to defend.

How to read these figures

The report comes from a venture capital firm that invests in this market itself, so it has an interest in portraying rapid growth. Spending estimates in studies like these are typically based on surveys and expert assumptions rather than companies’ billing records. The trend is probably real: agents do consume more resources and are moving into production. But the figures themselves are best treated as a guide rather than precise statistics. Until the methodology is published, $177 billion and 80% should be cited with the qualifier “according to Menlo’s estimates.”