We hear some version of this story constantly: a team at a large enterprise runs an AI pilot. It works. Users love it. The business case looks clean. And then nothing happens for six months while legal, IT security, and procurement work through their processes. The team that ran the pilot moves on to other things, the vendor's champion changes roles, and by the time the deal comes up for renewal discussion, nobody can remember why it was exciting.
This is the dominant pattern in enterprise AI adoption right now. It's not a technology problem. The pilots work. It's a structural problem — and understanding it has direct implications for how AI companies should sell and what infrastructure products are needed to unblock it.
Why Pilots Work and Deals Don't
Pilots are run under conditions that enterprise procurement processes are not. The champion has executive air cover for the duration of the pilot. Data access is granted on an exception basis. Security review is deferred until the pilot produces results. The users in the pilot are enthusiasts who want the thing to work.
Post-pilot, every exception that was granted for the pilot needs to become a policy. Data access needs to go through data governance. Security review isn't deferred anymore — it's blocking. The users who need to use the system to justify the ROI aren't the enthusiasts from the pilot — they're the full team, including the skeptics and the people who weren't consulted about the original decision.
The delta between pilot conditions and production conditions is where most deals die. And it's a delta that the AI vendor usually didn't help the customer prepare for, because they were focused on making the pilot succeed.
The Four Blocking Factors
Based on conversations with procurement teams at large enterprises, security reviews, and post-mortems with our portfolio companies on deals that stalled, the blockers cluster into four areas:
Data governance. Most enterprise AI use cases require access to sensitive data — customer records, financial data, proprietary documents. Getting that access approved under normal data governance processes takes time and requires documentation that pilot processes don't require. AI vendors that have invested in data handling certifications, audit trail capabilities, and data residency options move through this faster. Those that haven't can be stuck in security review for months.
Security and compliance review. Enterprise security teams are rightfully cautious about AI systems. The questions they ask — about model access to data, about prompt injection vulnerabilities, about logging and auditability — are legitimate and often not answered in early sales materials. The AI companies that arrive at security review with complete documentation, clear architecture diagrams, and relevant certifications cut review time significantly. The ones that treat security as the customer's problem stall.
IT integration. Most enterprise AI use cases need to integrate with existing systems — CRMs, ERPs, document management platforms, identity providers. The work of building and maintaining those integrations often falls to the customer's IT team, which has a full backlog. The AI companies that provide robust integration infrastructure, maintained connectors, and clear integration documentation reduce the IT burden enough to unblock procurement.
Change management. This is the one nobody wants to talk about. Getting 500 people to change how they do their jobs because a new AI tool is now available is hard. It requires training, communication, incentive alignment, and sustained attention from management. AI vendors that treat this as the customer's problem to figure out get adoption that trails off. The ones that provide playbooks, training materials, and success frameworks alongside the product see much better long-term retention.
What This Means for Infrastructure Vendors
For companies selling into enterprises, the implication is clear: success is not defined by pilot completion. It's defined by production deployment and active use at scale. That requires investing in the post-pilot infrastructure — security documentation, integration support, change management resources — before the deal is signed, not after.
For infrastructure vendors, it creates a different opportunity. The blockers described above — data governance tooling, security audit capabilities, integration infrastructure, deployment playbooks — are products that enterprises need and that the application layer isn't well-positioned to provide. The vendors that help enterprises move from pilot to production faster are addressing a bottleneck that everyone in the ecosystem benefits from removing.
The AI companies with the lowest CAC in enterprise are the ones that help customers justify the deal to everyone who wasn't in the room when the pilot was approved. That's a different product skill than running a good demo.
Our Portfolio View
Several of our portfolio companies have learned this the hard way and adapted well. The ones that have built the best enterprise traction aren't necessarily the ones with the best models. They're the ones that made production deployment someone's job — a dedicated customer success engineering function that owns the customer relationship through the integration and rollout phase.
That function costs money. It's not a natural hire for a research-heavy founding team. But the economics are clear: a deal that stalls at pilot is worth nothing. A deal that goes to production is worth the ACV you projected. The investment in making production happen is almost always worth it.
Building enterprise AI infrastructure or working on the pilot-to-production problem? This is an area we invest in actively.
