Speed Bumps in Enterprise AI Adoption

Enterprise AI adoption fails more often than it should. Not because the technology is inadequate — current models are heavily underutilized — but because familiar enterprise adoption patterns are challenged by AI in specific, non-obvious ways. What follows are speed bumps we encounter frequently, each with an AI-specific influence that makes it particularly important to navigate deliberately.

#1 — Locating Yourself on the Adoption Spectrum

At one end, AI is a powerful operational tool — it improves existing workflows, reduces costs, and increases throughput without changing the fundamental structure of the business. At the other end, AI represents a strategic pivot that reaches the core of what the organization does, how it competes, and what its economics look like in five years. Both ends are defined by business outcomes — revenue, margin, and survivability — not by the technology itself. The challenge is internalizing where on this spectrum your enterprise falls. Without that clarity, you risk overcommitting to a transformation the business doesn’t need, or underscoping a response to what is actually a competitive survival issue.

What makes this especially hard with AI is that the technology offers broad, foundational capabilities with no inherent specificity to any given enterprise’s workflows. Understanding what AI can do is a necessary starting point, but the genuine difficulty is going from that base of understanding to what it means for your organization — which workflows, which economics, which competitive dynamics. That assessment is high-stakes and cross-functional by nature, and it cannot be done well by either the business side or the technical side alone.

A discipline borrowed from venture capital can help here: before committing capital, bring cross-functional leadership together to jointly answer the hard questions — what specific business constraint or opportunity does this address, why is AI the critical enabler, what are the unit economics, and what evidence would tell us it isn’t working? The enterprises that clear this speed bump are the ones where business leaders can defend the technology thesis and technical leaders can defend the P&L impact. That cross-functional commitment to shared evaluation is what turns a position on the spectrum into a genuine organizational conviction.

#2 — Building a Contextualized Understanding of AI

Two dominant voices drive the public discourse around AI. Frontier labs focus on AGI and expanding capability horizons — aspirational by mission. On the other side, economists and industry skeptics point to high failure rates and unmet expectations. Both perspectives are valuable and the debate between them is healthy. But neither provides the contextualized understanding that an enterprise leader needs: what can AI do for my organization, given my specific workflows, constraints, and competitive position?

One approach that has served us well is studying where AI has demonstrated clear success and asking why. Coding is a great candidate — not just because AI works well there, but because the domain properties that enabled its success are instructive. Abundant, well-structured training data with clear patterns at varying complexity. Outputs that can be validated — code works or it doesn’t. That validation feeding a learning loop where the system improves through verified results. These aren’t properties of the AI models; they’re properties of the domain that allowed the models to be effective. The assessment question shifts: does my enterprise domain have analogous properties — structured patterns, representational data, a path to validation?

At its core, AI offers perception across modalities, reasoning within trained domains, tool use, and generation. These are powerful, general capabilities. But mapping them to specific enterprise value requires both a solid understanding of what these capabilities can and cannot do, and an equally deep understanding of where the business needs them most. Learning about AI from the perspective of applying it to your enterprise — grounded in your domain’s properties and your organization’s economics — is what transforms general awareness into actionable direction.

#3 — Developing a Long-Term Memory Strategy

Data readiness is important — having clean, accessible data to support an AI initiative is a real prerequisite. But it is not complete. The deeper question is whether your organization is building the capability to capture, curate, and represent its expertise in a form that AI systems can consume and compound over time — what we think of as the organization’s long-term memory. AI systems, particularly agentic workflows, need more than transactional data. They need experiential knowledge: how decisions are actually made, why proposals get rejected, what senior practitioners know that has never been captured in any system.

The co-pilot paradigm turns out to be more than an adoption strategy — it is a knowledge capture mechanism. When a human expert reviews an AI system’s output, corrects its reasoning, and explains why a proposal was rejected, that interaction generates exactly the kind of experiential data the system needs to improve. Rejection decisions, near-misses, and risk-based denials — knowledge organizations almost never document — become structured learning inputs. Over time, this creates a durable representation of institutional expertise that compounds with every deployment cycle. Organizations that recognize this early and design their AI adoption to build long-term memory as a byproduct — not as a separate initiative — create an advantage that deepens with time.

#4 — The Broken Ladder

We are already seeing enterprises struggle to recruit and develop talent suited for an AI-transformed workplace. In most professions, expertise develops through progressive exposure — simpler tasks leading to increasingly complex scenarios over years, building the judgment that defines senior capability. This progression is also the training system that produces the next generation of practitioners. AI disrupts it by assuming the work that once formed the apprenticeship path, effectively pushing everyone up the capability order — senior builders straddle product design, solution architecture, and implementation with the assistance of coding agents, while the entry-level work they once learned on is increasingly handled by AI. The broken ladder is now an industry-understood reality in software, and it will extend to every domain where agentic systems are deployed.

Co-pilot systems can help. The same system that captures expertise from senior practitioners — learning from their corrections, absorbing their reasoning — can turn around and teach it to the next generation. A well-designed co-pilot has two personas: on one end, it learns from experienced humans in the loop; on the other, it accelerates domain onboarding for newer practitioners by making that captured expertise accessible and interactive. This dual role is already a natural outcome of agentic systems — the same tools practitioners use to execute at a higher level also serve as powerful learning environments. Organizations that design for both sides of this interaction create a self-reinforcing talent development loop alongside their operational AI systems.

The Takeaway

Enterprise AI adoption is still early, and the technology’s potential remains significantly ahead of its current utilization. Developing a deeper, enterprise-specific understanding of AI — across functions and with rigor that matches the scale of the opportunity — is what separates productive adoption from expensive experimentation. And because AI is fundamentally new as a technology class, we should expect unfamiliar speed bumps to continue surfacing as adoption deepens — staying alert to them is as important as clearing the ones we already know.