Most companies thinking about AI adoption are asking the wrong question. They're still stuck at "can the model do the task?" — as if outputs themselves are the point. They're asking about capability, when the harder operational problem is now budget predictability and workflow design.

Uber dropped roughly 5,000 engineers into Claude Code in December. By February, adoption hit 63%. By April, they'd burned through their annual AI budget. The bill came faster than anyone planned.

That's the shift. The tools work so well that teams don't want to give them up. Finance becomes the bottleneck the moment productivity starts depending on token spend you never built guardrails for.

The Lazy Default

Here's what's happening in most companies right now: someone spins up GPT-4o for a coding task that could run perfectly well on a $0.50 model. Or Claude Sonnet for a rewrite that only needs a cheaper option. Or Opus for something a junior could do with the right prompt template.

There's no routing strategy. No spend visibility. No one asking "is this workflow actually worth frontier-model pricing?"

Neil Patel says his team cut AI costs 21% in 30 days — not by using worse models, but by routing intentionally. Most of their traffic moved to cheaper models. The smaller slice that genuinely needed frontier reasoning stayed there. That's the operational frame that actually works: optimise for ROI on token spend, not raw token minimisation.

The lazy default is using the most expensive model everywhere because it's easier than thinking. That's not a strategy. That's deferred technical debt.

What Winning Looks Like

The companies winning at AI adoption right now aren't the ones cheapest on tokens. They're the ones who built three things most teams skip:

  1. A routing layer — a cost-aware workflow that sends the right task to the right model. A 95% cost reduction from batching system prompts is real, but only if your infrastructure supports it.
  2. Visibility into spend by workflow — most teams have no idea what they're actually spending per use case. Without tracking by workflow, you can't distinguish productive usage from lazy expensive usage.
  3. A default + escalation model — most tasks should default to a cheaper model with explicit escalation to frontier only when the output genuinely needs it.

The market is past "does it work?" The useful question now is: can you control the spend without slowing down?

That's the service angle for an agency like ours. Clients don't need "AI implementation." They need routing strategy, guardrails, and reporting on what their spend is actually buying them. The teams that learn how to measure ROI on token spend and redesign workflows around cost-aware model usage will have the edge. Everyone else is just burning budget and hoping.

The shortcut is not cheaper models. It's intentional model selection.