You Want Your Token Costs To Explode (But That's Also Why You Need A Governance Layer)
Early on, one of the top 10 most-used agents at one of our customers was an agent an employee had built to help him write a fantasy novel. Good for him, but it’s not something the company should be paying for. The customer caught it, and the employee moved that work out of the company workspace.
The lesson is that AI has to be managed, and you can only manage it if seeing what’s going on is easy. The first question is “What are they actually using it for?”. Once you’ve answered that, the next one is “Are they doing it efficiently?” After that comes “What more could we be doing?” The companies creating the most value for their customers are the ones working on that third question.
Why costs are going up
Three things are happening at once on every serious AI platform.
First, AI tools are getting better at gathering the right context before they answer. Instead of counting on the user to put everything relevant in the prompt, platforms pull in the person’s role, company knowledge, related data sources and earlier work on their own.
One caveat: this is hard to measure. More context means bigger individual requests, but it also means much better output on the first try, which cuts down the back-and-forth it takes to get something useful. One bigger request that gets it right costs less per task than ten small exchanges that slowly get there. The average per request goes up while the average per task may go down, and the two move loosely enough that either number on its own will mislead you.
Second, scheduling means AI runs much more often than it used to. Agents kick off on a schedule, work through their inputs and post results around the company all day. Usage adds up in ways that are easy to underestimate if you only picture people sitting in individual sessions.
Third, it’s now much more practical to chain together complicated workflows that use many tools. One goal can set off a series of integrations, data pulls, drafts and updates across several systems. Every step costs something, and the chains keep getting longer as reliability improves and people try bigger things.
Put that together and the forecast is clear: token usage will go up, probably a lot, and roughly in step with how well your AI rollout is going.
Rising costs are the right outcome
Most companies get this backwards. Holding down AI spend by limiting usage doesn’t manage cost. It holds back what your people can do.
The whole point of enterprise AI is that employees get more meaningful work done, faster and with less friction. If that’s happening, they’re using more AI, and if they’re using more AI, they’re spending more tokens. The two numbers rise together.
A workforce using AI heavily to turn out better work in greater volume is exactly what you paid for. Rising token costs are the proof that you’re getting it.
The question is whether costs are going up for the right reasons
This is where the asterisk comes in.
More token usage is a good sign if employees are taking on ambitious, high-value work. It’s a bad sign if they’re sending a simple task to an expensive model when a cheaper one would do just as well. It’s waste if someone is running a personal project on the company’s AI subscription. And it tells you nothing if half your team uses AI through a dozen disconnected SaaS integrations that nobody can see into.
These worries are real. They’re what you’d expect when AI use spreads around a company on its own, with nothing underneath to support it.
You can improve the efficiency of token use
Is each task going to the right model? Are agents using context well? Is there a scheduled task that runs every day and produces output nobody reads? Is usage piled up in a few power users while everyone else does nothing? Are the same workflows running in several tools at once? Is anyone doing something they shouldn’t?
Without a governance layer, you can’t answer any of those questions.
Why a fragmented stack makes this unsolvable
A lot of companies have ended up with AI that’s technically available everywhere and governed nowhere.
There’s the Microsoft Copilot license that covers Office. There are the AI features built into the CRM. There’s the model subscription a few employees signed up for on their own. There’s the automation tool the ops team started using, and the AI features inside the project management software.
Each has its own usage, its own costs, and its own audit trail, or none at all. None of them talk to each other, and none give anyone a clear picture of what the company is doing with AI. You might have a separate API key for every tool or every team. That still tells you little beyond the size of the bill.
Signing employees into a dozen products that happen to have AI features doesn’t give you enterprise AI. It gives you enterprise shadow IT with better PR.
Committing to a single model provider or a single software suite has the same problem. As soon as a better model for your use case shows up somewhere else, you either miss out on it or add one more disconnected tool to the pile.
What governance actually requires
You can only govern AI if its use is centralized, and centralizing only works if the platform can do everything employees actually need.
That means any model, so people always have the best option for the task and you can see which model is used for what. It means any integration, so workflows run inside the governed environment instead of drifting outside it because the needed tool isn’t there. And it means any employee, technical or not, so the people on the system represent the whole company and the usage data actually means something.
When everything goes through one platform, seeing your costs and seeing your workflows become the same thing. You can tell where the tokens went, what they produced, whether it was worth it, and where you’d like people to work differently.
The goal is high usage you can understand
The companies that get the most out of AI will be the ones that push people to use it while building the means to see what that use produces.
Seen that way, rising token costs are a sign the company is healthy, not a budget line to squeeze. You want them going up. You just want to know exactly why.