The tokenmaxxing era, where companies measured AI maturity by how many tokens employees consumed, is hitting a wall. Uber burned through its projected annual AI budget in four months. Meta is capping usage after costs grew exponentially. Walmart set tool-level limits. Salesforce shifted its internal tracking to “agentic work units,” meaning units of completed agent output, rather than raw consumption. The pattern across all of them points to the same structural problem: usage was easy to count, so enterprises used it as a proxy for value<\/a>, and now the finance office wants a better answer.<\/p>\n The companies now scrambling to cap AI spend share a common trait: they funded adoption with experimentation budgets and measured success with activity metrics, which meant nobody had to answer the harder question until the bill arrived. If your AI programs are currently defended by dashboards showing tool adoption rates or token volumes, you are holding a position that the next budget cycle will challenge directly. The CFO asking whether AI spend belongs in the capital plan or the expense line is really asking whether anyone can prove the output justified the cost.<\/p>\n The IBM reframe from “tokenmaxxing” to “valuemaxxing” is the right direction, even if IBM’s own consulting business benefits from the shift toward outcome measurement requiring outside help. The underlying logic holds: cutting token budgets without measuring what those tokens produced can quietly destroy the context that AI agents need to perform well, generating more retries, more human rework, and ultimately more cost through a different line item. The smarter cost lever is routing, not rationing. AT&T’s Chief AI Officer cited savings of up to 90% by reserving frontier models for tasks that genuinely require deep reasoning and routing simpler work to cheaper alternatives. That is not an optimization detail; it is a pricing architecture decision.<\/p>\n For any ERP-adjacent spend on AI in finance, procurement, or supply chain workflows, the renewal conversation is already changing shape. Vendors will come in showing usage growth; the question your team should pressure-test is whether that usage correlates with measurable process improvement, whether exceptions dropped, whether close cycles shortened, whether planning cycles got faster. The enterprises that built outcome baselines before deploying agents will defend their budgets. The ones that didn’t will find themselves cutting usage to control cost and losing the productivity gains at the same time, which is the worst of both outcomes. I’d revise this view if major vendors begin publishing auditable, workflow-level ROI data at scale, but nothing in the current disclosure environment suggests that’s coming soon.<\/p>\n Model routing is the practice of directing each AI task to the model best matched to its complexity and cost profile, rather than defaulting every request to the most capable, most expensive option. Think of it like freight logistics: you don’t ship a letter overnight just because you can. In enterprise AI, routing decisions determine whether a coding agent uses a frontier model to reason through a complex dependency or a smaller, cheaper model to autocomplete a comment. Getting this architecture right is where most of the economic leverage in AI deployment actually lives.<\/p>\nWhat this means for your business<\/h2>\n
Concept deep-dive: Model routing<\/h2>\n