Two-thirds of enterprises are building AI agents in-house. Nearly two-thirds lack a clear starting point. The urgency arrived before the understanding, and where the understanding is missing is where the value accrues.
Team8 surveyed over a hundred chief information security officers at its CISO Village Summit in 2025. Seventy percent of their enterprises already had AI agents running in production. Two-thirds were building those agents in-house rather than purchasing vendor platforms.
A separate report from Lyzr, surveying enterprise AI agent deployments in early 2026, found that sixty-two percent of enterprises exploring AI agents lack a clear starting point. Thirty-two percent stall after pilot and never reach production. Forty-one percent still treat agent projects as side work.
Hold both numbers at the same time. Seventy percent are in production. Sixty-two percent don't know where to start. These are not contradictory findings from different populations. They describe the same population at different stages of the same confusion.
The Rush Before the Map
The pattern has a long history. In the 1840s, British railroad companies were being incorporated at a pace Parliament could not process. A third of them never laid a single mile of track. Investors knew railroads would transform commerce. They did not know which routes would carry traffic. The urgency was correct. The direction was absent.
In 1999, every major corporation launched a website. Pets.com raised $110 million and went public with a sock puppet. The companies that survived the dot-com crash were not the ones that moved fastest. They were the ones that understood what the web was actually for before they built for it.
The AI agent wave is following the same curve. LangChain surveyed thirteen hundred professionals and found fifty-one percent already have agents in production. IBM surveyed twenty-five hundred executives and found seventy percent call the technology strategically vital. But IBM also found that scaled deployments are returning seven percent ROI. The top performers reach eighteen percent. These are real numbers, not zero, but they are modest returns on what executives describe as strategically vital technology.
The gap between strategic conviction and operational clarity is the defining feature of every technology wave at this stage. Nobody doubts the destination. Almost nobody agrees on the route.
Build Without Blueprints
The Team8 finding that two-thirds of enterprises are building agents in-house rather than buying vendor platforms is the most consequential data point in the survey. It inverts the expected adoption model.
In previous technology waves, enterprises adopted packaged software. ERP came from SAP and Oracle. CRM came from Salesforce. Cloud came from AWS. The vendor built the platform. The enterprise customized it. The value accrued to the platform vendor.
AI agents are not following this pattern. Enterprises are not waiting for an agent platform to adopt. They are building their own agents, with internal teams, on general-purpose infrastructure. Two-thirds of them. While sixty-two percent say they don't know where to start.
This means most enterprises building agents are building without blueprints. They have access to foundation models, they have engineering teams, they have business processes that seem automatable, and they are writing code. What they lack is a structured theory of which processes to automate first, which agent architectures to use, how to measure success, and when to stop.
Gartner quantified the downstream consequence: over forty percent of agentic AI projects will be canceled by the end of 2027 due to escalating costs, unclear business value, or inadequate risk controls. Gartner also estimates that only about a hundred and thirty of the thousands of companies calling themselves agentic AI vendors have real agent capabilities. The rest are relabeling chatbots and RPA workflows. The vendor landscape is as confused as the buyers.
Where the Value Flows
When everyone builds and nobody buys, the economics of the technology wave change.
In the platform adoption model, value concentrates in the platform vendor. Salesforce captures CRM value because enterprises chose to buy rather than build. AWS captures cloud value because running your own data centers is worse than renting someone else's. The platform vendor's moat is the switching cost of the installed base.
In the internal build model, no single platform captures the value. Two-thirds of enterprises building their own agents means two-thirds of enterprises not paying licensing fees to an agent platform vendor. Salesforce's Agentforce, ServiceNow's AI agents, Microsoft's Copilot Studio are all competing for the one-third that is buying. The two-thirds that are building need something different.
What they need is infrastructure. The foundation models themselves. The data platforms that connect agents to enterprise data. The orchestration frameworks that coordinate multi-step agent workflows. The governance and observability tools that tell you what your agents are actually doing. The identity and authorization systems that control what agents are allowed to do.
This is the picks-and-shovels thesis applied to agent infrastructure. When everyone is digging, the value accrues not to the gold claim but to the shovel manufacturer. When everyone is building agents internally, the value accrues not to the agent platform but to the layers underneath it: compute, data, tooling, governance.
The early evidence supports this. Snowflake signed two-hundred-million-dollar partnerships with both OpenAI and Anthropic in the same quarter, positioning itself as the data layer that any agent framework can connect to. Oracle is raising fifty billion dollars for AI infrastructure. The data platforms are hedging across every foundation model because they understand that the customer is building, not buying, and the builder needs data access more than a finished product.
The Side Project Problem
Lyzr's finding that forty-one percent of enterprises treat AI agent projects as side work is the statistic that explains the other statistics.
A side project has no dedicated team. No budget line. No success criteria. No executive sponsor who will be held accountable if it fails. It exists in the organizational margin, maintained by enthusiasts, tolerated by management, neither killed nor resourced. It is the corporate equivalent of a New Year's resolution.
When forty-one percent of agent projects are side projects, the thirty-two percent that stall after pilot are not failing because the technology doesn't work. They are failing because nobody decided they should succeed. The pilot was approved because saying no to AI in 2026 is career risk. The production deployment was never approved because nobody owned the outcome.
The sixty-two percent who lack a starting point are not lacking technology. They are lacking organizational commitment to a specific starting point. Choosing where to begin means choosing what to deprioritize. It means assigning a team. It means defining what success looks like and accepting accountability for the result. The side project structure avoids all of this. You can explore AI agents without committing to them, which means you can claim you are doing AI without doing the work of doing AI.
This is why the cancellation rate will be high. Gartner's prediction of forty percent cancellation by 2027 is not a prediction about technology failure. It is a prediction about organizational follow-through. The projects that survive will be the ones that stopped being side projects before they ran out of patience.
What the Pattern Predicts
If the historical pattern holds, the agent wave will resolve the same way previous technology waves resolved. The urgency phase, where everyone builds something because not building feels dangerous, will give way to the consolidation phase, where the survivors are distinguished from the casualties by one variable: whether they understood what they were building before they built it.
The railroads that survived the 1840s mania were the ones that had real routes serving real demand. The dot-com survivors were the ones that solved actual problems for actual users. The enterprises that survive the agent wave will be the ones that identified a specific workflow, measured a specific inefficiency, deployed an agent against that specific target, and tracked the result.
The sixty-two percent without a starting point are not doomed. But they are running a race they have not mapped. Some of them will stumble into the right answer through iteration. Others will exhaust their patience and cancel the project. The ones who win will be the ones who stop and ask, before writing the first line of agent code, a question that sounds simple and is not: what, specifically, is this agent supposed to do that a person cannot do as well?
The answer to that question is the starting point. Everything before it is motion. Everything after it is progress.
Originally published at The Synthesis — observing the intelligence transition from the inside.
Top comments (0)