The gap between enterprises winning with AI and those still chasing their investment is no longer a mystery. BCG has the receipts — and the number is uncomfortably specific.
Companies in the top quintile of AI token usage are growing revenue at 16.5% year over year. Companies in the bottom quintile: 5.1%. Same economic environment. Same model providers. The difference is how aggressively they're deploying AI across their workflows.
That 3x performance gap is why Cursor — the $60 billion AI coding platform being acquired by SpaceX — just launched a CFO council to tackle the token cost puzzle head-on. The first roundtable is in August. The members include the CFOs of Asana and SentinelOne. And if you're a finance, technology, or operations leader who hasn't yet built a token cost governance model, this is your signal that the window is closing.
Why "Token Spend" Is Now a Board-Level Metric
For years, enterprise AI conversations lived in the engineering org. Which model to use. Which vendor to partner with. How to structure prompts. Finance leaders nodded along and wrote the checks.
That's over.
Token costs are now appearing in board-level cost reviews at companies scaling AI seriously. A BCG analysis published last week puts it plainly: "Token costs are real and growing fast. They are attracting CEO and board-level attention. CFOs, CIOs, and CTOs need to be ready with answers when those leaders start asking questions."
The reason it's escalating: the per-unit cost of AI is falling, but total spend is accelerating because usage volumes are growing faster than costs are declining. The blended cost of AI per million tokens fell 67% between Q1 2025 and Q1 2026 — from $18.40 to $6.07. But 73% of enterprises reported that AI costs exceeded their initial projections over that same period. Cost per token dropped. Total spend went up. This is the Jevons paradox applied to AI: lower cost per unit drives more usage, not less.
The CFOs who don't have governance frameworks for this dynamic are the ones getting surprised by end-of-quarter AI budget overruns.
The BCG Data Point That Should Change Your AI Investment Thesis
The 16.5% vs. 5.1% revenue growth split isn't a soft correlation. BCG derived it from Cursor's own usage data — real behavioral signals from companies actively deploying AI coding workflows at scale.
The finding fits a pattern that's showing up across multiple data sets. PwC's 2025 AI Metric Survey found that companies investing more than 1.6% of their revenue in AI achieve 9.5% higher EBITDA, 20.2% better total shareholder return, and 3.5% higher revenue than peers investing below that threshold.
The throughline: AI advantage is not distributed evenly. It accrues disproportionately to the companies pushing hardest on deployment — measured not by strategy documents, but by actual consumption.
Jordan Topoleski, Cursor's COO, framed the paradox well: "The productivity value of AI is growing with each major model release, but adoption is uneven, usage is concentrated, and costs vary widely depending on how work is routed."
Translation for leadership teams: the companies getting 16.5% growth aren't using better AI. They're using more of it, more effectively — and they've built the operational infrastructure to manage the costs that come with that scale.
The Token Cost Problem: What's Actually Happening in Enterprise Finance Teams
Most enterprises don't have a single AI budget. They have a scattered patchwork: cloud line items, individual SaaS subscriptions, per-seat AI add-ons, and token-based API costs spread across five different cost centers. Nobody has the full picture, and the finance team is often the last to find out when a department's AI experiment quietly becomes a $50,000/month production workload.
The numbers illustrate how wide the variance is. According to data from April 2026, the median monthly AI token spend for businesses was $2,246 per month. The average was $140,842. That's a 63x gap between the median and the mean — which means a small number of high-deployment companies are pulling the average way up. The middle 50% of companies spent between $3 and $352 per employee per month on AI tokens.
But here's what the raw spend numbers don't tell you: architecture matters more than raw spending. Organizations running a single frontier model architecture paid a median blended cost of $18.40 per million tokens in Q1 2026. Those running tiered multi-model architectures — routing simpler tasks to lighter models and reserving frontier models for complex work — achieved an 87% cost reduction, landing at $2.31 per million tokens.
The same business outcomes. Nearly one-tenth the cost.
That math is why model allocation strategy has become a C-suite conversation, not an engineering footnote.
Agentic AI Is Making This Harder — And More Urgent
If you thought the token cost problem was manageable with current conversational AI deployments, agentic AI is about to reshape your assumptions.
Agentic AI systems — those designed to complete multi-step tasks autonomously — consume 5 to 30 times more tokens per task than standard conversational tools. A single agentic workflow that replaces what a human analyst used to spend an hour on might generate hundreds of model calls, each with its own token overhead.
This isn't a theoretical risk. It's already hitting companies that moved quickly to production agentic deployments. Finance teams approved ROI projections based on chatbot-level token costs, then watched their monthly AI bills spike when multi-agent orchestration went live. Initial models suggested certain dollar amounts, but production agentic systems consumed at an order of magnitude higher cost.
For context: Gartner predicted that 40% of enterprise applications will incorporate task-specific AI agents by end of 2026. If that projection holds — and current deployment trajectories suggest it will — the token cost governance problem becomes the defining CFO challenge of the next 18 months.
What Cursor's CFO Council Is Actually Building
The announcement of Cursor's CFO Council is worth reading as a market signal, not just a networking initiative.
Cursor is acquiring its own product credibility from an unlikely direction: SpaceX, which agreed to acquire the company for $60 billion in a deal expected to close in Q3. That's a company that takes operational rigor seriously. Elon Musk built SpaceX on cost-per-outcome thinking: what does it cost to put a kilogram in orbit, and how do we drive that number down systematically? Applying that mindset to AI token costs — treating intelligence as infrastructure with measurable cost per unit of work — is exactly what the CFO council is attempting to formalize.
The council's stated goals are concrete: shared benchmarks for AI productivity, frameworks for measuring returns on intelligence, and practical approaches to model allocation and cost management. The inaugural partners — Asana CFO Aziz Megji and SentinelOne CFO Sonalee Parekh — represent the kind of fast-scaling SaaS companies where AI cost dynamics hit first and hardest.
The working group will meet quarterly in rotating cities. August is the first session.
What they're trying to build is the enterprise equivalent of the FinOps framework that emerged to manage cloud cost complexity a decade ago. Cloud vendors gave enterprises cheap per-unit pricing that encouraged adoption, then companies discovered they needed an entire discipline — people, processes, tooling — to manage the resulting spend. The same dynamic is now playing out with AI tokens, and the CFO Council is a bet that the framework needs to be built collaboratively rather than each company reinventing it independently.
The Practical Framework: What Finance Teams Should Build Now
The data is clear on the direction. Here's what that translates to operationally:
Map every AI cost center. Before you can manage token costs, you need visibility. Most enterprises have AI spend distributed across direct API contracts, embedded costs in SaaS tools, and cloud-hosted model inference. Build a unified view. The monthly median of $2,246 and average of $140,842 tell you the distribution is extreme — you need to know where you sit on that curve.
Separate model tiers for different workloads. The 87% cost reduction from tiered multi-model architectures is not an edge case. It's the difference between AI that scales profitably and AI that creates exponentially rising costs as deployment expands. Engineering and finance teams need to collaborate on routing logic: which tasks genuinely require frontier models, and which can be handled by smaller, cheaper options.
Project agentic costs separately from conversational costs. If your organization is moving toward agentic AI — and most enterprise AI roadmaps are — your token cost models need to reflect 5-30x multipliers on per-task consumption. ROI projections built on chatbot economics will fail when agentic systems go to production.
Set consumption guardrails at the department level. Token-based pricing introduces a new kind of operational risk: a single runaway workflow or aggressive new tool can generate thousands of model calls before anyone notices. Department-level consumption alerts, similar to cloud spend alerts, are a minimum governance requirement.
Define your AI investment threshold. PwC's data suggests 1.6% of revenue as the threshold where outcomes shift materially. That's a planning target, not just a coincidence. Finance teams should calculate where their current AI investment sits as a percentage of revenue and model the gap.
The CFO's Read on 2026
The conversation I'm hearing from finance leaders at enterprise AI events consistently lands on the same themes: token costs are becoming visible to boards, ROI measurement is still immature, and the gap between AI leaders and laggards is widening faster than anyone expected.
The BCG data quantifies what many finance leaders already sense: this isn't a marginal productivity story. 16.5% versus 5.1% revenue growth is a strategic divergence. In most industries, a sustained 3x differential in revenue growth rate is the kind of gap that reshapes competitive landscapes within three to five years.
The companies that figure out token cost governance — who can deploy AI aggressively while managing costs with discipline — will be the ones on the right side of that divergence. The companies that either underinvest in AI or let token costs spiral out of control without governance will be on the other side.
Cursor's CFO Council is a bet that the right answer is a shared framework. That the complexity of AI cost management is a collective problem, and that enterprises benefit from solving it together rather than individually.
Given that the first meeting is in August and the founding CFOs are from companies actively scaling AI, it's worth watching what framework emerges. That's likely to become the de facto governance standard for the next cycle of AI deployment — the same way FinOps became the cloud cost management template.
Cursor's CFO Council first roundtable takes place in August 2026. CFOs interested in participating can apply through cursor.com. The BCG report "The Era of Token-Based Competition Is Here" and "Managing AI Token Costs" are available at bcg.com.
