For some time now, I’ve been challenging an assumption that has taken hold across the AI industry: more tokens, more reasoning, and more powerful models automatically translate into better outcomes.
I’ve written about this phenomenon, often called tokenmaxxing, and why I believe it represents one of the biggest obstacles to making enterprise AI economically sustainable.
Now, The Wall Street Journal is bringing broader attention to the consequences. Its recent reporting found that only 11% of nearly 400 companies surveyed could forecast their AI spending within 10%. Another study cited in the article found that cheaper AI models actually cost more to complete a task in 32% of cases. In one particularly striking example, a lower-cost model took 952 steps, consumed $14, and still failed, while a more expensive model completed the task in 85 steps for about a dollar.
Most people will read this as a pricing problem. I see it as an architecture problem, and one we’ve been actively addressing at Autonomize AI.
The issue isn’t simply that AI is expensive. It’s that enterprises are paying for a tremendous amount of unnecessary computation.
We Need to Stop Paying AI to Rediscover What It Already Knows
Consider what happens when an AI agent processes a healthcare prior authorization. It needs to understand the clinical information, identify the applicable coverage policy, interpret the rules, and determine what happens next.
In many implementations, the model reconstructs that understanding every time, even when the underlying policies and procedures haven’t changed. Some systems add more agents to review the work of other agents, increasing the number of model calls without necessarily improving the outcome.
Multiply that across millions of transactions, and the costs become difficult to predict. More importantly, every additional step introduces latency and another opportunity for inconsistent results. In healthcare, that concerns me far more than the invoice.
As I argued in my earlier piece, The Trouble with Tokenmaxxing, results that cannot be reproduced consistently are difficult to trust. They’re also difficult to budget.
The real opportunity isn’t to make AI think harder about every task. It’s to capture what we’ve already learned and apply it consistently across the enterprise.
We’ve Been Engineering for This Reality
At Autonomize, this isn’t just theory. We’ve been actively addressing it through token optimization, adaptive prompt optimization with RLMOpt, and a shared context foundation that allows agents to apply validated knowledge, policies, and decision logic across workflows without starting from scratch.
Our approach starts with a simple principle: Use reasoning where it creates value, and stop paying for it where the answer is already known.
In practice, that means three things:
| Reason Once and Reuse It | Convert validated policies, standard operating procedures, and decision logic into structured, reusable assets that can be applied consistently rather than asking models to reinterpret the same information for every transaction. |
|---|---|
| Match the Model to the Task | Reserve powerful models for problems that genuinely require complex reasoning, while using smaller, optimized models or deterministic execution for routine work. |
| Engineer for Consistency and Predictability | Evaluate not only whether a system produces the right answer, but whether it can reproduce that outcome reliably, with traceability, and at a predictable cost. |
We’ve seen how dramatically this can change the economics of execution. In our token optimization work, we’ve demonstrated how workflows that previously required 21 model calls can be reduced to execution paths that require no additional model calls for the validated procedure. That’s not about reducing capability. It’s about eliminating unnecessary computation once the correct process has been established.
We’re not trying to minimize tokens at the expense of accuracy or safety. We’re building systems that become more efficient and reliable as organizational knowledge is reused across workflows.
The Economics of AI Will Become a Measure of Its Maturity
I believe the WSJ’s reporting reinforces an important shift in how enterprises need to evaluate AI, especially as they plan their 2027 investments. The conversation must move beyond what models and agents can accomplish to whether organizations can operate them reliably, predictably, and economically at scale.
That requires a different set of metrics: cost per successful outcome, consistency of execution, reuse of established knowledge, and the ability to increase transaction volume without a proportional increase in compute costs.
We’ve spent the last few years proving what AI is capable of. Now comes the harder challenge of building systems that get smarter, more efficient, and more economical with every transaction, rather than more expensive.
AI’s next breakthrough won’t come from consuming more tokens. It will come from making what we’ve already learned compound in value. The real measure of AI progress won’t be how much computation we consume. It will be how much work we can accomplish without paying to solve the same problem twice.
For leaders planning their 2027 AI investments, we’ve built a simple AI budgeting calculator to explore how transaction volume, model calls, and token consumption affect costs, and how more efficient architectures can change that equation. The estimates are directional, but they offer a useful starting point for a conversation every enterprise should be having.
About the Author
Ganesh Padmanabhan is the CEO and founder of Autonomize AI, where he is building AI-native solutions for healthcare enterprises. With over 20 years of experience across AI, cloud, and enterprise transformation, he focuses on embedding governed intelligence into workflows to improve outcomes and reduce operational friction. A recognized voice in responsible AI and healthcare innovation, Ganesh is a frequent speaker, advisor, and advocate for using technology to tackle humanity’s biggest challenges.
