UBS published research this month showing 60% of companies actively managing AI budgets have already moved to cheaper models or open-source alternatives. The teams that have not are running into $35,000 monthly bills and quota overruns at 200% of plan.
The pattern is consistent. A team adopts an AI assistant or builds a workflow on a frontier model, usage grows faster than expected, the bill arrives, and someone in finance asks a question that nobody can answer cleanly: what exactly are we spending this on?
The model routing shift
The UBS research identifies model routing as the dominant cost-control response. This means sending straightforward tasks (classification, short-form Q&A, simple summarisation) to cheaper models such as Claude Haiku or Gemini Flash, and reserving frontier models (Claude Opus, GPT-4.1, Gemini Pro) for tasks that genuinely require them: complex reasoning, multi-step code generation, long-context analysis.
This is sound engineering practice. A frontier model on a classification task is the equivalent of hiring a consultant to file receipts. The capability exists, but the cost-to-value ratio is wrong.
The problem is visibility, not intention
Most engineering teams know model routing is the right answer. The problem is identifying which specific workflows have the mismatch. Without workflow-level cost attribution, the routing decision is a guess. You might switch the wrong workflow, cause quality problems, and conclude that optimisation is not worth the effort.
The teams doing this well start with a diagnostic. They know exactly which API calls are costing them, which task types are running on overspecified models, and where caching would eliminate repeat spend. That visibility is the foundation that makes the routing decision safe and quantified rather than directional and risky.
Teams without that visibility face a different problem. They can see the total bill but not what is driving it. They suspect there is waste but cannot quantify it precisely enough to act with confidence. The result is inertia: the spend continues, the bill grows, and the finance conversation becomes harder.
Chinese open-source and the cost curve
The UBS research also identifies Chinese open-source models (Qwen, DeepSeek, MiniMax) as entering the enterprise cost curve for lower-complexity tasks. These models can be run locally or accessed through cloud catalogs, which changes the make-versus-buy calculation for teams with high-volume, lower-stakes workloads.
Token spend diagnostics need to account for this expanded option set, not just the Anthropic and OpenAI pricing tiers. A classification or summarisation workflow that currently runs on a mid-tier commercial model may be a candidate for a Chinese open-source alternative at a fraction of the cost. Assessing that requires knowing what the workflow is actually doing and what output quality it requires — both of which come from diagnostic data, not assumptions.
What the 40% are waiting for
The 40% of companies that have not yet switched models are not necessarily doing the wrong thing. Some have workflows that genuinely require frontier capability. Some have compliance or data residency constraints that limit their model options. But a significant proportion are simply waiting for clarity they do not yet have.
They know they are probably overspending. They do not know by how much, on which workflows, or what the routing change would actually cost them in terms of implementation effort and quality risk. The decision to switch feels too large to make without better data, so it stays on the backlog.
What this means for your team
If your AI spend has grown to a point where it appears in budget conversations, the question is no longer whether to optimise. It is where to start. The teams moving fastest are the ones who began with a diagnostic rather than a guess: map the spend at workflow level, identify the mismatches, and make routing decisions based on actual data rather than assumptions about what your team is doing.
The diagnostic also gives you the finance conversation you need. When the CFO asks what you are getting for the AI bill, "we've mapped it and here is the ranked list of savings" is a materially different answer to "we think there might be some optimisation opportunity."