AI is changing. Who’s in control?

TL;DR

The problem: Most organizations are pushing AI adoption while leaving cost and governance as problems for later. Later is arriving faster than expected. 

Key insights: 

• AI tools aren’t consolidating around one vendor. Companies are running multiple models across multiple departments, often without central visibility. 

• Token-based pricing is hard to forecast, vendors don’t always surface consumption data by default.

• Shadow AI is already here. It’s embedded in tools employees use every day, whether IT knows about it or not. 

Your next step: Map where AI is being consumed in your organization before someone else sends you the invoice.

More pie, not less

At the start of this year, a reasonable prediction was that the AI market would start to settle. Companies would pick a platform, standardize, and the field would consolidate.  

Well, that hasn’t happened. 

What’s visible across organizations is that AI tool adoption isn’t replacing anything. Different teams are running different models because different models are better at different things. The idea of a single corporate AI assistant has given way to a sprawl of subscriptions, API calls, and agent workflows nobody has fully mapped yet.

Everyone wants everything.

Xensam’s own SaaS data from Q1 backed this up. The major AI platforms weren’t taking market share from each other. In fact, the pie was getting bigger.

The bill you didn’t budget for

The cost picture that comes with this is genuinely hard to manage. Token-based pricing, how most organizations pay for AI consumption via APIs, adds a layer of complexity that traditional software licensing doesn’t have.

A per-user SaaS subscription is predictable. Token consumption is not. Usage depends on which models employees are running, how complex their queries are, and whether anyone is asking those questions at all.

There’s a real budgeting problem here. Organizations that set their AI budgets at the end of last year were doing so before several of the models and capabilities they’re now paying for even existed. And unlike a traditional software overage, AI consumption can spike fast. 

Earlier this year, The Information reported on an internal Meta leaderboard called “Claudeonomics” that ranked the company’s 85,000 employees by AI token consumption. They were handing out titles like “Token Legend” to the top power users. In 30 days, those employees burned through 60 trillion tokens. Yes, read that again: 60 trillion tokens! Priced against Anthropic’s public API rates, that’s roughly $900 million in a single month. Meta likely purchases tokens at a significant discount, but the point stands: when consumption becomes a competition with no cost accountability attached, the numbers get away from you fast. Meta took down the leaderboard after the story broke. 

What makes this harder to diagnose is that the data isn’t always there by default. Pulling consumption from a service like Gemini often gives you a summary total. To understand which agent consumed what, you have to tag every prompt, which itself consumes additional tokens. The visibility that lets you make good decisions comes at a cost, and most organizations haven’t built that infrastructure yet. 

Shadow AI is not shadow IT

The distinction matters. Shadow IT was usually an employee spinning up a tool the organization hadn’t approved. With shadow AI, the features are already baked into tools the organization is paying for.

AI features are in Miro, Canva, and most SaaS products employees use every day. That consumption is happening whether it’s tracked or not, and it sits alongside the dedicated AI subscriptions that are at least visible to someone. 

This creates tension. Organizations have two pressures pulling in opposite directions: push adoption fast to capture the efficiency gains and maintain enough control to stay compliant with regulations like the EU AI Act, which carries real fines for organizations that can’t demonstrate they’re managing their AI responsibly. 

More organizations are responding to this with a dedicated AI committee, reporting to C-level, responsible for both driving adoption and maintaining oversight. But adoption velocity and governance rigor don’t automatically go together, and the cost side tends to be the last piece of the framework to get built. 

Agents, identities, and the next SAM problem 

AI agents introduce a layer of complexity that goes beyond software spend. When an agent operates with its own identity and permissions inside your systems, you’re opening questions that most organizations haven’t fully worked through yet.  

The principle is simple: agents should run on scoped service accounts, not on personal credentials, and their access should be deliberately limited. Giving an agent broad system access is a security risk with a short feedback loop. 

There’s a licensing dimension here too. Microsoft already benefits from the identity model where each employee maps to one M365 account. AI agents don’t fit that structure neatly. If an organization spins up multiple agents, each potentially needing its own account and permissions, the number of identities under management grows fast, and so does the associated cost. 

What comes after tokens

Tokens are the current standard for AI pricing, but they’re obviously not the final one. Fixed-price or outcome-based models are emerging as alternatives that trade precision for predictability.  

A capped plan means you might run out mid-month, and an uncapped plan means the bill can surprise you. Neither is great. 

There are some practical levers available now. Running batch jobs during off-peak hours can reduce token costs by 30-40%. Choosing the right model for the task matters too. The newest, most expensive model isn’t always the right fit, and for the majority of everyday tasks, older versions often deliver comparable quality at significantly lower cost. 

Open-source models are also closing the gap faster than expected, and the prospect of running a capable model in-house on local infrastructure is becoming realistic for more organizations. If that shift gains momentum, it will reshape demand for commercial AI subscriptions in ways the market is still working out. 

And what now?

AI spend is growing fast and, for most organizations, without the visibility that software spend has built up over years of SAM practice. The habits that help here are familiar ones: know what you have, know what it costs, know who’s responsible. The difference is that AI moves faster and the surprises arrive monthly, not at audit time. 

Getting visibility in place now, before consumption scales further, is a lot easier than untangling it later. 

If any of this resonates with what you’re seeing in your organization, the full conversation is worth your time. Joel Sanderi and Martin Tureniec go deeper on AI pricing models, shadow AI, agent governance, and where all of this is heading.