The AI Agent Budget Reckoning Has Arrived. Better Architecture Is the Answer.

The pressure has shifted from having an AI story to having AI results, and most teams never architected their agents to produce them.

In business, nobody wants to be the Blockbuster to someone else’s Netflix.

That fear is what busted open every enterprise checkbook for AI in 2024 and 2025, and few leaders asked where the ceiling was. It was a reasonable approach at the time. Miss a platform shift like this, and a nimbler competitor has eaten your entire lunch before you even see it coming.

So companies signed off on the bills, stood up the agents, and told the board a story about the future. Now the bills are piling up, and a lot of people in 2026 are looking at their token spend the way you’d check your credit card statement after a Vegas vacation.

In fact, early this year, Uber’s COO conceded the company burned through its entire AI coding budget in just four months and needed to update its company AI use parameters. Fortune 500s like Microsoft, Amazon, Walmart, and Uber have also all made changes to better monitor and control employee AI spend.

This ROI reckoning is actually a good sign. It means the conversation has moved past whether AI works in theory to whether it works in practice. A year ago, the bar was hoping the thing worked. Now, it’s making it work at scale, on budget, and across real systems.

The space is maturing exactly the way it should, the same way innovations have matured in the past.

First it was servers, then the cloud, now it’s LLMs

Go back in time, and this exact pattern has played out for transformational technologies many times before. There is the initial deployment, which comes with an explosion in usage. Then, the trough of disillusionment and the reckoning that comes with it when companies finally realize, “Wait, how did we organize all this?”

For example, cheap x86 servers took off in the early 2000s, and companies went from a few hundred machines to tens of thousands in their data centers almost overnight. Utilization was a mess, and nobody could say what half of those servers were doing while burning up money. Then VMware arrived, bringing virtualization and a collective sigh of relief. Suddenly, enterprises could consolidate all that sprawl onto a fraction of the hardware.

That technology eventually evolved into the cloud and repeated the pattern. EC2 made it trivial to spin up a server on demand, and everyone marveled. A generation of software got built on that convenience. But businesses blinked and had millions of instances running. So the industry built DevOps and shipped tools like Terraform to wrangle the sprawl back under control.

Large language models are the next platform on that list. Enterprises adopted agents quickly, and now they’re waking up to agents everywhere and mounting token bills. Coming face-to-face with that reality is progress, and it’s the same thing I tell every leader who tells me they’re worried they’ve fallen behind. History shows we’re still in the early innings of this game.

Value and ROI aren’t the same thing

Every time a new headline about AI spend and ROI hits, there’s a group of people who say, “See, the emperor has no clothes.” People get disillusioned and think there’s no real ROI in AI. The skepticism is understandable, and the measurement problem is real. But ROI is a trailing indicator, and those are always difficult to prove in the moment.

Ask a team to produce a clean ROI number for Notion, Linear, and GitHub, and they’ll struggle. Then ask them to give up those tools, and you’ll get a hell no. Value and measurability aren’t the same thing, and mixing them up is how you end up denying a tool’s return while depending on it every day.

To return to the Uber example, at the same time the company is examining usage cost, Uber CEO Dara Khosrowshahi said, “Whether it’s our legal team or marketing team or developers, we think it’s [AI] creating employees with superpowers.”

Defining that value is a measurement problem. The spending problem is a different animal, and it traces back to production, not to the technology itself.

Agent architecture is expensive to get wrong

Agent architecture covers a wide array of choices: how narrowly you scope its tools, how tightly you scope its permissions, how clearly you define its job, how much you reuse across agents instead of rebuilding every time, and much more.

Get those right and the work happens in a few clean steps. Get them wrong and the agent improvises, burning tokens along the way.

Tool design is an easy place to see how spending and systems architecture are linked, so let’s start there. Arcade.dev’s own research, published in the recent “State of MCP Tools” report, ran eight identical CRM queries through two Attio MCP toolkits. One was built by Arcade, the other by a competitor, to measure how much tool design affects token spend.

• Arcade’s toolkit used 7,426 tokens across all eight queries.
• The other used 747,083 tokens for the same tasks against the same data.
• At a scale of 100 agents running 200 queries a day, that gap amounts to $1.97 million a year in token costs at Claude Sonnet pricing. For Opus, it’s five times that amount, with newer and more expensive model versions now out since those tests.

While I’m not shy about saying Arcade’s MCP toolkit is the best out there, the larger point is that smart design is the difference. An agent with a well-defined tool built for the job does it without extra steps. An agent that can’t reliably call the right tool will guess, retry when it guesses wrong, and continue to do so until it finally stumbles onto an answer, burning tokens the entire time.

Tools are just one lever, and the same logic runs through every other variable. As enterprises get serious about reining in token spend, architecting agents for efficiency will become the common refrain, and the savings will compound the same way costs do now.

If you want to see how much of your spend is influenced by architecture rather than the model itself, that “State of MCP Tools” report I mentioned above breaks down the full methodology and is a good place to start.

Better architecture, higher ROI

The enterprises that get through this reckoning won’t be the ones that pull back on AI, or the ones who spend blindly. They’ll be the ones who evolve to fix the inefficiencies that are inflating spend.

After all, the only place ROI can exist is in production, and the work to be done is in helping enterprises get there safely and efficiently.