Let’s start with something obvious: nothing in business gets a free pass from being measured. Not even magic.
If you’re leading a company in 2025, you’ve likely signed off on a generative AI pilot. Maybe several. The initial excitement—the demos that write marketing copy in seconds, the models that summarize quarterly reports—often gives way to a dull, administrative throb: the bill. And that bill is almost always denominated in tokens.
Bernard Marr’s recent overview hits the nail on the head: tokens are the currency of this new economy. But in my two decades covering Wall Street and corporate finance, I’ve seen a familiar pattern. A novel, technical metric emerges—be it “clicks,” “streaming hours” or “cloud compute units”—and well-intentioned managers swiftly mistake it for a measure of value. Tokens are the latest and perhaps most opaque entry into that dangerous canon.
So, let’s demystify. An AI token isn’t a cryptocurrency. It’s a fundamental unit of processing. When an employee asks a model to draft an email, that prompt is chunked into tokens. The model’s response is generated token by token. As Marr notes, a token can be a short word, a piece of a longer word or a punctuation mark. This tokenization is why pricing from providers like OpenAI, Anthropic and Google Cloud is so often structured per thousand tokens for input and output. It’s a direct meter running on computational effort.
This creates a seductive sense of control. Finance teams love metered usage; it feels prudent, scalable. You can track departmental consumption, compare the cost-per-thousand-tokens of GPT-4o versus Claude 3.5 and forecast next quarter’s AI spend. This is the useful core of tokenomics. It’s the plumbing, as Marr calls it. You need to understand it to manage your capex and opex in this new paradigm.
The peril begins when that plumbing is mistaken for the architecture of success. I’ve spoken with CTOs at mid-sized firms who, under pressure to show AI adoption, have built internal dashboards ranking employees by token usage. The instinct is understandable. You’ve made a major investment; you want to see it being used. But what you’re measuring is raw consumption, not productivity. It’s the digital equivalent of rewarding a factory manager for the highest electricity bill, irrespective of how many units were produced or their quality.
This phenomenon even has an emerging name in Silicon Valley circles: “tokenmaxxing.” Employees learn to game the system, perhaps by asking the AI to elaborate unnecessarily or to process the same document multiple times. The leaderboard ticks up, but output doesn’t improve. It’s a vanity metric. Amazon’s brief experiment with an internal AI usage leaderboard, which was reportedly shut down after an executive warned staff not to use AI “just for the sake of using AI,” is a cautionary tale cited in industry briefings from Forrester and Gartner. The message was clear: activity is not achievement.
The real insight for business leaders isn’t in counting tokens, but in connecting them to outcomes. Token data is a starting point for a more intelligent conversation. For instance, a sudden spike in token consumption from your legal team could mean one of two things. They could be efficiently reviewing hundreds of contracts, saving weeks of manual work. Or they could be using a poorly configured workflow that sends the entire contract library to the model with every query, burning cash without delivering faster results. The token count alone won’t tell you which. You need to marry that data with performance indicators: time-to-review, error rates, outside counsel costs.
This is where finance leadership becomes critical. The CFO’s office must steer the organization away from simplistic token budgeting and toward value accounting. Don’t just ask, “How many tokens did we buy?” Ask, “How many engineering hours did those tokens save in code review?” or “What was the lift in customer satisfaction scores after implementing the AI-powered support agent?” The Federal Reserve Bank of San Francisco’s 2024 research on AI productivity noted that measurable gains are almost never in the raw usage stats, but in the secondary effects on workflow speed and quality.
In practice, this means your AI governance should have two layers. The first is technical and financial, managed by your tech and finance ops teams:
- Monitoring token spend
- Selecting efficient models for specific tasks
- Eliminating waste like redundant processing
- Benchmarking AI performance
- Creating budgets based on usage
- Tracking expense trends over time
The second is strategic, driven by business unit leaders: defining what “value” means for each AI application and measuring it relentlessly. Did the marketing AI actually improve click-through rates? Did the sales assistant shorten the deal cycle?
| Year | Model | Cost per Token | Use Case | Measured Outcomes |
|---|---|---|---|---|
| 2023 | GPT-4o | $0.0001 | Marketing | Increased CTR by 15% |
| 2023 | Claude 3.5 | $0.00005 | Legal Review | Reduced review time by 30% |
| 2024 | GPT-5 | $0.00008 | Sales Support | Shortened cycle by 20% |
Finally, a word on the future. Tokenomics is still evolving. As models become more efficient, the cost per token may fall, but new pricing models—subscriptions, revenue shares—will emerge. The core principle, however, will remain. Any technology investment must be judged by its return, not its consumption. Tokens tell you what your AI costs. Only your business results can tell you what it’s worth. Ignoring that distinction is the quickest way to turn a transformative technology into a very expensive line item on a dashboard, blinking with empty activity.
