A BCG official is telling enterprise leaders to start spending their AI capacity, not just buying it, as new data reveals that companies plan to double AI budgets in 2026 to roughly 1.7% of revenues while fewer than one in three have moved meaningfully beyond pilots, exposing a gap between commitment and execution that threatens the near-term revenue story for every company in the AI supply chain.
The metaphor of starting the pump captures something real about the current state of enterprise AI. Water fills the pipes. Nobody turns the tap. Companies have signed agreements with Microsoft, Anthropic, Google, and OpenAI. They have Copilot licences sitting idle. They have agentic AI infrastructure that cost significant budget to procure, sitting in the equivalent of a closed valve. The Business Insider framing, urging firms to start the pump on AI tokens, is a shorthand for a specific failure mode: investment without activation, spending without use, capability without deployment. The tokens in question are not a single technical construct. They refer collectively to the purchased capacity: API usage credits billed per thousand tokens of context processed, Microsoft 365 Copilot seats charged monthly regardless of engagement, enterprise licence tiers on AI platforms that meter output against prepaid budgets. You pay for the pipe. You still have to open it.
BCG's own research documents the scale of the problem with uncomfortable precision. The firm's January 2026 global survey of 2,360 executives found that companies plan to double AI spending from 0.8% of revenues in 2025 to approximately 1.7% in 2026, and that 94% of CEOs will continue investing at current or higher levels even if those investments do not pay off this year. That is an extraordinary statement of commitment from executives who are simultaneously being held to quarterly earnings targets. It reflects genuine belief that AI is a strategic necessity rather than a discretionary experiment. But BCG's own adoption data from the same period tells a different story at the operational level. Only 31% of organisations report meaningful progress scaling generative AI beyond pilots. The remaining 69% are still in experimentation mode, seeing little to no measurable value. In functions like corporate affairs and communications, BCG found that 72% of teams do not yet meet the productivity potential available at the task level, which is a prerequisite for unlocking the process-level improvements that actually move EBITDA.
The mismatch between executive confidence and middle management readiness is the most consistent finding across BCG's 2025 and 2026 AI research. Seventy-eight percent of C-suite leaders expect significant AI returns within 18 months. Only 23% of middle managers feel adequately prepared to deliver those outcomes. That gap is not a technology problem. The tools exist and largely work. The gap is a change management problem, a skills problem, and a measurement problem. BCG's own analysis from its AI at Work research concluded that technical implementation accounts for only 30% of the total effort required for successful AI adoption. The remaining 70% is helping employees adapt to new workflows, developing new skills, and adjusting how organisations define performance and make decisions. Most companies investing heavily in the 30% are dramatically underinvesting in the 70%.
The practical implication for model labs and SaaS providers is significant. Enterprise deals are increasingly structured as committed spend agreements: Microsoft's Copilot, Google's Workspace AI, Anthropic's Claude for Enterprise, all of these are typically purchased as monthly per-seat or usage-tier arrangements. Revenue is recognised whether or not the seats are actively used. But utilisation data feeds renewal conversations, expansion conversations, and the case studies that drive new sales. Low utilisation means weak evidence for ROI, which means procurement teams question renewals, budget holders escalate for justification, and the CFO's patience for unrealised potential eventually runs thin. The BCG official urging companies to start the pump is, among other things, making a case for protecting the AI industry's own revenue cycle from the adoption lag that threatens it.
Consulting firms are positioning themselves as the answer to this problem, and that positioning is worth examining directly. BCG, McKinsey, Accenture, and Deloitte have all built significant AI transformation practices in the past 18 months. Their pitch is that they bridge the gap between purchased capability and deployed workflow, providing the change management, process redesign, training, and measurement frameworks that large enterprises cannot build internally fast enough. That value proposition is real. BCG's five barriers framework, identifying talent gaps, unclear ownership, legacy infrastructure, measurement inadequacy, and insufficient change investment as the primary blockers to AI value, is accurate as a diagnostic. The question is whether the consulting model of external advisory is the right delivery mechanism for what is fundamentally an internal capability building problem. You cannot outsource the organisational learning that makes AI adoption durable. You can only facilitate it.
For startups building in the enterprise AI layer, the adoption gap creates both opportunity and risk. The opportunity is in tooling that reduces the activation cost of AI for non-technical users: workflow automation that embeds AI at the point of task completion rather than requiring a separate AI chat interaction, measurement dashboards that give managers visible evidence of productivity changes, change management software that tracks adoption and surfaces coaching needs. The risk is that enterprises in experimentation mode are poor customers. They buy pilots, not platforms. They need proof before they commit to the scale that justifies startup unit economics. The companies positioned best are those that can demonstrate measurable workflow impact within a defined time window, short enough to survive budget review cycles, clear enough to survive CFO scrutiny. The pump metaphor is useful. Pumps work when the infrastructure is connected end to end. Most enterprise AI stacks right now are still a collection of pipes pointing at each other without joints.
Also read: Waymo's Child Passenger Problem Is a Preview of Every Messy Regulation Autonomous Vehicles Will Face • A Developer Ditched Cloud AI for a Local Model on an RTX 6000 Pro and the Results Are Hard to Argue With • Unsloth found and fixed a bug in Mistral Medium 3.5 and the story reveals how much of the open-weight model race is actually a tooling race
