Dr. Chih-Han Yu is the CEO and co-founder of Appier, an AI-native agentic-AI-as-a-service (AaaS) company.
This spring, a term briefly took hold in Silicon Valley: tokenmaxxing. The idea was simple. Many were trying to push token consumption as high as possible and treat usage volume as a proxy for AI adoption. Watching those conversations unfold, one question kept coming to mind for me: Does using more tokens mean producing more value?
Think of an AI agent as an employee and tokens as its salary. A higher salary doesn’t guarantee better work, and token volume alone tells you nothing about the value AI has created. At my company, we care less about how many tokens a task used and more about what it accomplished.
Token spend surged rapidly at some tech companies this year. At least one reportedly burned through a year’s AI budget in four months, as cited by Fortune. In that same article, Fortune noted that the tokenmaxxing craze is already fading.
None of this means companies should use less AI. It means token volume can no longer stand in for productivity. The bigger bill isn’t really the problem. The real question is whether token consumption has drifted from the task’s objective, its output quality and the business outcome it was meant to serve. What a company measures shapes what its people optimize for, and that shapes its AI culture. Reward using more tokens, and you’ll get more tokens used, whether or not any of them mattered.
My company has always been an ROI-driven business. We want every dollar our customers spend to generate a return, and we hold internal investment to the same standard. Calculating AI’s ROI is genuinely difficult, though, and it’s something we’re still refining.
How Do You Actually Measure The Return?
Start by separating what the investment is meant to achieve. Is this deployment meant to cut costs or to grow revenue? The two need very different approaches.
Cost reduction is the easier one. Take a software project, for example: If AI is introduced without changing scope or quality, total hours should go down. Say a project required three engineers for eight weeks. With AI, it takes six. That means six engineer-weeks saved. Convert that into dollars based on engineering seniority, then compare it against your incremental token and infrastructure costs.
What matters is the gap between two numbers: How much extra did you spend adopting AI, and how much cost did you actually eliminate? Compare a token bill to your team’s entire payroll, and AI will always look like a bargain. That tells you nothing real. The investment only pays off when added spend is lower than realized savings and quality hasn’t slipped.
Revenue growth is harder, and it comes in stages. A company might roll out AI to boost sales productivity with an eye toward revenue, but looking at the top line on day one is a mistake. First, check whether each salesperson can serve more customers per day. Then check whether conversion rates have improved. Only after that should you look at whether revenue and lifetime value have grown. Each stage needs its own metrics, tied to your company’s processes. There’s no formula that works for every business.
Building The Infrastructure To Measure It Properly
Companies with the capacity can build more complete AI governance and performance-tracking systems. At my company, we built an internal LLM gateway, originally just to centralize cost management and security. Over time, it grew into a platform tying together AI tools, model calls and usage tracking, so we can see who’s using what and for which task, as well as what it produces.
When cost and business metrics live in the same framework, every dollar of AI spend maps to a KPI. For engineering teams, the real benchmark isn’t token consumption. It’s whether AI is shortening development cycles, improving quality or reducing rework.
Measurement is a cultural discipline as much as a technical one. AI ROI is hard to calculate, but if difficulty becomes an excuse not to measure, your company loses its basis for knowing what to improve next. Once you set a baseline, your estimates get sharper with each round.
Put Resources Where They Create Value
Measuring token consumption doesn’t mean hitting the brakes. Companies shouldn’t restrict AI use just because usage spiked in the short term. At my company, when we see an unusual jump, we bring in people with relevant expertise to see what’s driving it and help the team find a better way to get the task done.
Tokens don’t need to be maximized, and minimizing them isn’t the goal either. What you should be after is optimization. In the early exploration phase of product development, restricting usage too aggressively is a mistake. The priority is to validate, in a fairly unconstrained environment, which use cases are feasible and which create real value. Treat cost as the primary constraint too early, and you risk shutting down innovation before you’ve understood what the technology can do.
Before a product goes live, it’s time for another round of cost optimization: reviewing which tasks were handled by higher-cost models and testing whether a leaner one can do the job just as well. After launch, tuning should keep going based on real usage. None of this is new. Teams have always validated feasibility first, then tightened performance and cost. What’s changed is that this now has to be built into the product lifecycle from the start, not scrambled together after the bill spikes.
For companies on fixed-fee AI subscriptions, the math works differently. Since cost stays flat regardless of volume, the priority is getting as much out of what’s already been paid for.
Every conversation about AI, agents and tokens comes back to one question: Is this generating a positive return in real-world applications? Only when it consistently does should companies expand investment, and only then does it make sense to put more into model development and computing infrastructure.
Spending the most on AI was never the point. What matters is whether every token spent comes back as something real.
Forbes Technology Council is an invitation-only community for world-class CIOs, CTOs and technology executives. Do I qualify?