Enterprise spending on foundation generative AI models is set to more than double this year, from $11.4 billion to $23.4 billion, according to Gartner. Most of that money is still being tracked with the same metrics IT has used for two decades: infrastructure cost, license seats, uptime. None of those tell you what’s actually driving an AI bill.
As AI adoption grows, the metrics have to change too. Token consumption, model usage, and cost per outcome are AI-specific measures that infrastructure budgeting was not originally built to track, and a survey shows what that gap costs: 85% of organizations wrongly estimate their AI spend by more than 10%, and nearly a quarter miss by 50% or more. Visibility into how agents actually use models and resources is what lets a CIO balance performance, quality, and spend, instead of guessing at all three.
You can’t manage what you can’t see
Gartner’s Will Sommer, a senior director analyst at the firm, has a warning for anyone celebrating falling token prices: commodity tokens getting cheaper doesn’t mean frontier reasoning is getting cheap. Agentic AI models need 5 to 30 times more tokens per task than a standard chatbot, and Gartner still expects enterprise bills to climb even as inference costs fall more than 90% by 2030, because consumption is rising faster than price is falling.
The visibility gap is what turns that into a surprise instead of a forecast. Most AI spend is scattered across whatever tool each department adopts on its own, with no single view of which team is spending what, on which model, for which task. Governed orchestration closes that gap: it gives CIOs per-process cost visibility instead of a lump sum that shows up at the end of the month.
Not everything needs to run on AI
None of this is an argument for routing everything through AI. The highest-value moves keep a process running on deterministic orchestration and reserve the model call for the one step that actually needs judgment. AI aimed at the decisions that need it, orchestration handling the rest, is how enterprises get the most value for the least spend.
Enterprise MCP also makes the AI calls themselves more efficient. Giving an agent comprehensive access to the context and data it needs up front means it can prompt once instead of guessing its way through several calls, which is its own form of cost control.
This is ultimately an evolution, not a one-time fix. As organizations get better at understanding how AI is actually being used day to day, usage matures and cost comes down with it. The organizations getting ahead of this right now aren’t the ones spending less on AI, but the ones who can actually see where the money is going.
What that looks like inside with Enterprise MCP
Enterprise MCP builds that visibility at three levels.
Access control governs who can reach which MCP servers, down to the tool, user, and group level, so the right people have access to the right resources and nothing more.
Frequency and volume limits set rate limits and quotas at the MCP server level. That governs behavior, not individuals, so a runaway agent gets capped before it turns into a line item.
Visibility and audit logs track every tool call an agent makes. Independent testing of a widely used developer MCP server found its tool definitions alone consume roughly 42,000 tokens before an agent does anything useful, more than a fifth of a standard context window. Logging every call is how a team catches that kind of waste before it shows up on the bill, instead of after.
Ultimately, not everything needs to be run through AI. The highest-value moves keep a process running on deterministic orchestration and reserve the model call for the one step that actually needs judgment. AI aimed at the decisions that need it, with automation and orchestration handling the rest, is how enterprises get the most value for the least spend.
As organizations get better at understanding how AI is actually being used day to day, usage matures and cost comes down with it. The organizations getting ahead of this right now aren’t the ones spending less on AI. They’re the ones who can actually see where the money is going.
Using a single unified platform, like Workato, allows enterprises greater visibility and control of their enterprise AI usage. Access controls, rate limits, and audit logs handle the governance side. The execution side runs on composable, pre-orchestrated skills that do the actual work an agent would otherwise reason through step by step, burning tokens on every decision along the way. Most tools built for this moment pick one of those halves and call it done. Doing both from the same platform is what turns AI spend from a number finance discovers after the fact into one IT can see, shape, and defend as it happens.
Want to see where your own AI spend has visibility gaps? Talk to our team.