- Articles
- Your AI Is Busy. But Is It Actually Working?
Your AI Is Busy. But Is It Actually Working?
Why chasing tokens is the wrong game, and what enterprises should be doing instead.
Somewhere in your organization right now, there’s an AI agent running. It’s processing, retrieving, reasoning, and looping. It’s consuming tokens by the thousands. And someone in your technology team is watching the usage dashboard with quiet satisfaction, because the numbers are going up.
The question nobody is asking loudly enough is: so what?
The Token Trap
Enterprises fell in love with the wrong metric. When Agentic AI arrived, the instinct was to measure success by how much AI was being used: how many agents were deployed, how many tokens were being consumed, how many processes had an AI layer sitting on top of them. Activity became the proxy for progress, and for a while, that felt like enough.
It wasn’t.
Uber exhausted its entire 2026 annual AI budget in four months. The culprit was the rapid rollout of agentic AI tools where costs were multiplying faster than anyone had modelled. This had several factors: Uber incentivized users with an internal leadership tracking AI usage, AI companies switched from fixed cost to consumption-based pricing, and engineers switched to more complex models even for simple tasks. Agentic AI doesn’t behave like a chatbot. It thinks in loops, calls tools repeatedly, and every single one of those steps burns tokens.
Microsoft, watching the same dynamic unfold internally, began canceling most of its Claude Code licenses in mid-May, ending access across entire divisions.
These aren’t naive companies. These are two of the most sophisticated technology organizations in the world. And they both got caught by the same thing: they were measuring the wrong thing.
Tokens Are Not Value
Here’s the uncomfortable truth that a lot of enterprise AI programs haven’t fully reckoned with yet. A token is a unit of compute. It is not a unit of value. An agent can consume millions of tokens and deliver nothing that moves your business forward. And an agent can consume very few tokens and save your organization millions of dollars, if it’s been designed around the right problem.
A Forbes 2025 survey last November found that fewer than 1% of executives report significant ROI from their AI investments. Not because the technology isn’t capable. But because most organizations designed their AI programs around adoption, not around outcomes.
That’s the gap. And it’s a big one.
What Valuemaxxing Actually Means
Valuemaxxing isn’t a complicated idea. It just means that before you build an agent, before you deploy a workflow, before you spin up a model, you answer one question clearly: what business outcome does this exist to produce?
Not “automate this process.” Not “reduce manual effort.” A specific, measurable outcome. Processing time drops by a defined number of days. Error rates fall to a specific threshold. Cost per transaction hits a target that Finance has signed off on. If you can’t name that number before you start, you’re not ready to start (or you are working on the wrong use case).
JPMorgan Chase got this right with their COiN platform, which uses AI to review commercial loan agreements. A task that previously consumed roughly 360,000 hours of lawyer and loan officer time every year now takes seconds. But the key reason it worked isn’t the technology. It’s that the outcome was defined before a single line of code was written. The ROI was calculable before the first model was trained, and the process was redesigned before automation was applied. That’s the sequence that matters: outcome first, technology second.
Walmart did the same thing in supply chain forecasting. Rather than building a broad AI layer across all of operations, they went after one specific problem: inventory prediction at the store level, where even marginal improvements translate into hundreds of millions of dollars in reduced waste. The constraint wasn’t technological ambition. It was business value. And that constraint is precisely what made it work.
The Questions Every Enterprise Needs to Answer
Before the next Agentic AI deployment, these questions aren’t optional:
Name it before you build. If you can’t, stop and go back.
Because an agent running on a broken process doesn’t fix the process. It just executes it faster and at greater cost. Fix the process first, then decide what role an agent plays in it.
In 2023, an Air Canada AI chatbot confidently told a grieving passenger they could claim a bereavement fare retroactively, a policy that didn’t exist. The tribunal didn’t accept Air Canada’s defense that the chatbot was a “separate legal entity” responsible for its own statements. Air Canada paid, and the internet had a field day. But underneath the jokes and the headlines was a genuinely important business lesson: nobody in that organization owned the outcome. IT deployed the tool, the vendor built the model, the customer service team ran the process, and when the agent got it badly wrong, there was no single person who was accountable for what it said, what it promised, or what it cost when it was wrong. That accountability gap is everywhere in enterprise Agentic AI right now. Somebody needs to own the result end-to-end, not just on the deployment date. That’s a different job, it requires a different governance structure, and many organizations haven’t built it yet.
Define it before go-live. Task completion rates, cost per outcome, error reduction: these are the metrics that tell you whether an agent is working or just running.
At Chazey Partners, we help organizations build the operating models, governance frameworks, and transformation roadmaps that turn Agentic AI investment into Agentic AI outcomes. If your AI program is producing activity but not outcomes, that’s the conversation worth having.
Book a free consultation with Chazey Partners today. Get the clarity and frameworks to make sure your Agentic AI investment is actually earning its place. Because impressive agent demos aren’t a strategy. Value is.
About Chazey Partners
Chazey Partners is a global management consulting firm specializing in operations transformation and shared services. As Agentic AI reshapes how organizations think about work, we bring deep expertise in back-office operations and operating model design to help leaders ask the right questions, and act on the answers. Whether you’re at the beginning of your Agentic AI journey or looking to accelerate an initiative already underway, Chazey Partners brings the frameworks, the experience, and the objectivity to ensure you’re thinking about it at the right level. To learn more, visit www.chazeypartners.com.
Author
Chas Moore
Executive Partner - Transformation
Don't forget to share this post!