The cloud bill surprised every CFO once, too. Token costs are following the same arc, just faster and with less warning. Box CEO Aaron Levie made this point plainly in a LinkedIn post this week: token costs are now one of the hottest budget topics in every enterprise conversation he has. His framing was bullish, reading rising spend as proof of real-world scale that nobody anticipated. That read is correct. It is also, for most enterprise technology leaders, the wrong place to focus.

The more consequential observation came second. Levie identified model routing as the differentiator that will separate applied AI platforms over the next several years. The logic is sound: once tokens represent a significant share of workflow cost, the ability to direct each task to the cheapest model that can handle it competently becomes a genuine strategic capability, not an infrastructure footnote.

A simple AI interaction in 2023 (input, retrieval, response) cost roughly four cents. An orchestrated agentic workflow in 2026, with tool calls, iterative reasoning loops, and multi-step output generation, runs closer to $1.20. That is a 30-times increase per interaction, according to analysis published this month by EY. The number comes from structural complexity, not price inflation. The models themselves are getting cheaper. The workflows are getting more expensive because they are doing more work.

The pattern Levie described is already showing up in enterprise budgets. One healthcare organization consumed one trillion tokens over six months, generating more than $6 million in unplanned costs before the finance team could trace the source. Uber and Microsoft have reportedly burned through their 2026 token budgets in a matter of months. This is not a story about profligate spending. It is a story about enterprises deploying agentic systems without corresponding visibility into consumption.

Most enterprise agent rollouts exceed their pilot budget by four to eleven times within the first 90 days of broad deployment, driven primarily by uncapped tool-call recursion and retrieval breadth that nobody scoped at the pilot stage. The CFO starts asking pointed questions about per-workflow unit economics. The control surface shrinks right when it needs to expand.

The conventional read of this situation treats model routing as a cost-optimization layer, something you bolt on when bills get uncomfortable. Levie's argument is different: routing is where the applied AI layer builds a moat that neither frontier model providers nor hyperscalers can easily replicate.

Frontier models will keep improving at complex reasoning tasks. Legal analysis, healthcare diagnostics, financial modeling: spending on these will increase. But the majority of enterprise AI volume is not frontier work. It is document extraction, classification, summarization, form completion, first-draft generation. Tasks that do not require the most capable model available, only a model capable enough and confident enough that the routing system knows when to escalate.

That is the capability Levie is describing: domain-specific evaluation frameworks that know, for a given document type or legal task subclass, which model performs best at what price point. The applied AI layer that builds this evals infrastructure owns the routing decision. It can substitute cheaper models, route to open-weight alternatives, or escalate to frontier intelligence when the task genuinely demands it. The platform that cannot make this determination sends everything to the frontier and passes the cost to the customer.

The pricing spread makes the stakes concrete. Budget-tier models run at roughly $0.10 to $1 per million tokens. Frontier models with extended reasoning capabilities run $15 to $30 per million or higher. A platform routing 70 percent of volume to the appropriate tier is not offering a slightly better deal. It is offering a fundamentally different cost structure.

Levie's second LinkedIn post, on context architecture, connects directly to the routing thesis. His argument: the enterprise AI platform problem is not model quality, it is getting the right context to agents reliably and at scale. Enterprises have knowledge fragmented across legacy systems, with access controls that do not map to actual workflows. Agents either have too much access or too little. Neither works.

The portability point carries a specific warning for CIOs. Locking context into a single agent architecture means long-term dependency on that vendor's model choices, pricing decisions, and roadmap. The organizations that build context infrastructure capable of serving Claude Cowork, Codex, ChatGPT, Cursor, and whatever agent platform emerges next are not placing a bet on today's model leaders. They are buying optionality in a market that is still pricing in its own uncertainty.

Box is building toward that platform position. What makes the argument worth watching is that Levie is describing a structural opportunity, not a product announcement. The companies that can establish domain evals, build routing logic, and keep context portable will capture value that neither model providers nor cloud infrastructure vendors can absorb. That is a large enough space to matter.

The missing piece in most enterprise token governance conversations is instrumentation before deployment, not after budget overrun. The organizations managing this well treat token spend as an operational metric with the same rigor as latency and accuracy, not as a billing surprise to be explained in quarterly reviews.

That requires three things that most enterprises do not yet have. Visibility into consumption by workflow and department. Routing policies that direct low-complexity tasks to cheaper models without manual intervention. And escalation logic that is calibrated to actual task requirements, not to the default assumption that more expensive means better.

The applied AI platform that delivers all three, and can demonstrate per-workflow unit economics to a CFO who has stopped asking polite questions, is in a different conversation from the one about features and integrations.