The AI Cost Problem Worth Having
Three things shaping my thinking about AI cost management.
The reporting on AI cost management has intensified in recent months. Stories about large technology firms struggling with infrastructure costs at scale—Uber is a frequently cited example—have moved cost control toward the top of the CIO agenda. I read these accounts carefully. But for firms like ours, they feel less like a warning and more like a dispatch from a different country: our AI cost problem is not cost control, but discovery.
Three things have shaped my thinking.
The Adoption Phase Changes the Calculus
At Cornerstone, like most professional services firms I talk with, we are still figuring out how to use AI effectively. That means preserving client confidentiality, designing workflows that pair machine output with the human auditing our work requires, and combining genuine exploratory brainstorming with carefully executed final analysis. So, the priority is not cost optimization. It is discovery: learning which uses of AI create genuine value, which require significant human correction before they are usable, and which are not yet ready for production.
Cost management matters. Of course, I want to avoid significant unexpected overruns. But it is not the binding constraint right now. Finding the right workflows is key because they tell us where AI creates value and where it does not.
Said differently: you optimize the thing you understand. We do not yet fully understand what we are building, so optimizing prematurely has its own cost.
Advanced Users Are Driving Almost All of the Spend
This is the most practically important thing I know about our current AI cost structure: approximately 80% of our AI spend comes from 10% of our users. And those users are not running low-value tasks. They are the people who have gone furthest in figuring out what agentic AI can actually do—the ones who have learned to orchestrate multistep workflows, who push the tools well beyond what most users attempt.
A typical knowledge worker using AI mostly for chat and drafting is not incurring high costs. That will change as AI capability diffuses more broadly. But today, cost is highly concentrated in a small group of advanced practitioners who are also getting the most out of the tools. That is not a coincidence; it is the payoff in its current form.
This configuration will not persist. But right now, it is the reality—and it matters considerably for how you think about managing it.
Smaller Firms Can Be Surgical in a Way Larger Ones Cannot
We have fewer than 2,000 employees. Not a tiny firm, but not Uber. Today, roughly 20 people in the firm are genuine AI power users—people running agentic workflows that generate meaningful spend at scale. That is a small enough number that I can talk with each of them directly. And I do.
Those conversations are genuinely interesting. The people generating the highest AI spend are not wasting money. They are doing things with AI that are hard to capture in a policy memo and easy to understand in a 15-minute conversation. When I read about firms worried about rising AI costs, I sometimes wonder what the high-spend users in those firms are actually doing—because the right response to expensive AI usage depends entirely on whether it creates value. If it does, the payoff is worth the spend.
In firms our size, we have something that larger firms do not: the ability to be high-touch and specific. I do not need a policy designed to handle 10,000 edge cases. I need to understand roughly 20 people. That is a different problem, and right now it is tractable.
Rising AI cost is not a problem I want to solve by suppressing usage. In the current moment, rising costs are mostly evidence that certain people in the firm are discovering genuinely new ways to do our work. That is worth encouraging, not auditing into submission.
What I am watching for is the moment when this changes—when cost diffuses from power users to the broader workforce, when usage is less concentrated and harder to evaluate person by person. That moment will require different tools. We are not there yet.
For now, the most useful thing I can do when AI costs rise is to ask two questions:
Who is spending?
What are they spending on?
The answers, so far, have been encouraging. I feel genuinely good about rising AI costs when I understand what they are buying.
The views expressed herein are solely those of the author and do not necessarily represent the views of Cornerstone Research.