01The Numbers Are In
Uber burned through its entire 2026 AI budget in four months. A three-person team spent over $1.3 million on tokens in a single month. Goldman Sachs published a report this spring estimating that agentic AI — the autonomous, multi-step systems that have become the dominant enterprise deployment model — can increase token demand by 24 times compared to conversational tools. The average business is spending 13 times more on AI tokens than it was in January 2025.
The paradox: token prices dropped 67% year over year. Costs are still exploding. When volume grows 24 times and price drops 67%, the bill increases by a factor of roughly eight. The math is not complicated — but it turned out to be a surprise to a lot of organizations that made volume assumptions based on chatbot usage and then deployed agents.
JPMorgan published a note entitled “AI Token Costs Are Eating Internet Profits Alive.” Shopify, Spotify, ServiceNow, and Roku flagged AI as a surging share of operating expenditures in recent earnings calls. Fortune 500 companies are reporting monthly AI inference bills in the tens of millions.
Volume assumptions that were reasonable for a conversational tool are wrong by an order of magnitude for an agent that orchestrates a workflow.
The question being asked now is no longer “Can AI do this?” It’s “Is any of this actually generating a return?”
02The Familiar Pattern
If you’ve been in manufacturing long enough, you’ve seen this before. Not with tokens — with the software that was supposed to fix the process.
The ERP that was going to eliminate the spreadsheet. The MES that was going to give you real-time floor visibility. The quality system that was going to close the SCAR loop automatically. Each one cost more than planned. Each one delivered less than promised. And in most cases, the gap between expectation and result came down to the same root cause: technology deployed on top of an unredesigned process, with no defined financial objective and no measurement method established before the work started.
The AI token cost crisis is the same failure mode. Faster, more expensive, and at a scale that makes the P&L impact impossible to ignore.
Uber’s operations chief noted that token usage “didn’t seem to have a direct correlation with useful consumer features.” That sentence should be recognizable to anyone who has ever reviewed a software implementation budget and asked what exactly was delivered. The tool was running and the spending was real. The outcomes never showed up.
03What the Numbers Actually Say
BCG published research showing that 70% of AI ROI comes from people and process change — not technology. Twenty percent comes from data and infrastructure. Ten percent comes from the algorithm itself.
Most organizations deploying AI are spending in reverse sequence. The models are live. The agents are running. The tokens are accumulating. But the process wasn’t redesigned before the automation was built. The financial objective wasn’t defined before scope was locked. Nobody established a measurement baseline at the start of the engagement, so there is no clean way to measure what changed.
That’s not a technology problem. That’s a discipline problem.
Agentic AI doesn’t amplify good decisions. It amplifies whatever you give it — including broken processes, unclear ownership, and undefined outcomes. At 24 times the token volume.
When the process is unstable, automating it produces faster failure. That has been true since the first workflow platform. AI makes it more expensive and more visible.
04The Manufacturing Version of This Problem
Mid-size manufacturers may not run $50 million monthly token bills, but the failure mode is identical, at a different scale and cost profile.
The AI pilot that was deployed without a defined financial objective. The tool that’s live but the operators have found workarounds for. The automation that broke the moment someone changed a supplier form, and never got fixed because the vendor was gone and nobody internally owned it. The license that gets renewed each year because canceling it would require admitting the ROI never appeared.
The cost shows up differently than a token bill. Wasted license fees. Operator time managing a tool that makes the job harder, not easier. A quality workflow that’s automating a process that nobody standardized first. A procurement automation built around data that nobody trusts.
The enterprise version of this failure is now generating headlines because the spending is large enough to move earnings calls. The mid-market version has been generating waste quietly for years.
05The Discipline That Prevents It
Every 7Flows engagement starts with the same question: what is the specific financial outcome we are trying to produce, and how are we going to measure it?
The question isn’t which tool to use or which process to automate. It’s the measurable result — labor hours eliminated, supplier non-conformance cost reduced, OTIF penalty exposure closed — and what it’s worth on the P&L. That question has to be answered before scope is locked. Before the automation is designed. Before any spending begins.
The financial objective defines what success looks like. It determines which processes are worth automating and which aren’t. It establishes the baseline that makes measurement possible after deployment. And it creates the accountability structure that separates a real engagement from a pilot.
The operational diagnosis comes first, before the technology selection. Stabilize the process. Define the financial target. Agree on how it will be measured. Then design the automation to hit that target. In that order.
The companies now renegotiating their AI contracts didn’t skip the technology — they skipped the discipline. The technology was the easy part. It always was.
Do you know which AI initiative in your operation is generating a measurable return right now?