Over the past two years, the enterprise conversation has been dominated by AI capabilities, productivity gains and adoption rates. I believe the next major conversation will be about something less exciting but far more consequential: AI economics. Not which models to use or which vendors to partner with, but whether organizations know what their AI is costing them, who is responsible for that spend and whether it is delivering the outcomes the business expected when it approved the investment.
Unlike traditional software licensing, AI introduces a consumption-based model where every prompt, every agent action and every inference carries a cost. A single interaction may cost only pennies. But at enterprise scale, those pennies add up to millions of interactions per month, creating a category of technology spend that is genuinely difficult to forecast, attribute or explain. In some cases, the value is obvious and measurable. In others, the investment sits in a grey area where the technology is clearly being used, but nobody can say with confidence what it has returned.
Uber exhausted their entire 2026 AI coding budget within four months. What struck me about that story was not the scale. It was the familiarity. I have worked on teams where AI was saving hours on document review and summarization every single week. The time savings were real, and everyone felt them. But the cost per interaction had never been logged, so the business case lived in people’s heads rather than in any report. The rideshare giant’s COO put it plainly: “It’s very hard to draw a line” between rising AI costs and useful features for customers. That gap between AI adoption and AI accountability is one most organizations are still navigating.
In my experience, some of the increase in AI costs organizations cannot explain comes down to a rise in the cost per interaction that nobody planned for. The model changes, the per-token price jumps and usage continue scaling as if nothing happened.
A team builds a workflow on a capable, cost-efficient mid-tier model. It performs well. At some point, someone upgrades to a frontier reasoning model, either because the output felt noticeably better or simply because it was available. What nobody checks is that frontier models are dramatically more expensive per token, generate significantly more verbose responses and hit usage limits far faster. The model did not just get better. It got hungrier, and the budget absorbed that quietly.
I have seen this play out even at the individual level. On a personal AI subscription, switching from a mid-tier to a frontier model can exhaust a monthly message limit in a fraction of the usual time, not because the user is doing anything differently, but because a more powerful model thinks longer, responds at greater length and consumes far more tokens per interaction. The behavior of the model changes the cost profile entirely, even when the task stays the same.
Now multiply that across an engineering team, an operations group using an internal AI assistant and a customer-facing product, all running the upgraded model simultaneously. Nobody made a budget decision. Nobody ran a cost comparison. Someone changed a single line in a config file and the spend profile of the entire organization shifted overnight. In my experience, this is not an edge case. It is how AI cost surprises happen inside organizations today, quietly and without any paper trail.
The organizations handling AI economics well are making architectural decisions up front that build cost intelligence directly into how their systems operate. One of the most effective approaches I have seen is model routing, sometimes referred to as the orchestrator-subagent pattern or tiered model architecture. Rather than routing every task through the most powerful and expensive model available, you assign a lightweight model to handle routine execution and only escalate to a frontier reasoning model when the task genuinely requires it.
Think of it like any well-run team: a junior resource handles the day-to-day work and escalates to a senior manager only when the problem genuinely requires that level of judgment. You do not pull a senior manager into every task. You reserve that capacity for the decisions that need it. In practice, a team building an internal contract review tool might configure a lightweight model to handle the initial pass, extracting key clauses, flagging standard terms and formatting the output. When that model encounters an unusual clause requiring deeper reasoning, it escalates to a frontier model for expert-level analysis. Once resolved, execution returns to the lightweight model. The result is near-frontier quality on the hard cases at a fraction of the cost of running an advanced model across every document.
What I value about this approach is the discipline it forces. It requires teams to think deliberately about which tasks need the most capable model and which do not. That thinking, applied consistently, is what separates organizations that govern AI spend from those that simply absorb it.
I have been in rooms where a team demos an AI agent and the energy is infectious. It reads documents, drafts responses, pulls data from internal systems and hands off to the next step in the workflow. Then the question comes up: what data does this agent have access to? In most of those rooms, the answer is silence. Teams think about capability before they think about boundaries, and that silence has consequences. Without clearly defined limits, an agent can inadvertently process personally identifiable information or protected health information never approved for AI use. Regulations like GDPR, HIPAA and CCPA do not make exceptions for unintentional exposure. Beyond data, there is also the risk of prompt injection: malicious directives embedded inside a document or email that hijack what the agent does next. The organization’s liability does not change because the breach was caused by an AI agent rather than a human.
Access is one side of the problem. Output is the other, and in my experience, it is the one that catches organizations off guard more often. A model that hallucinates does not announce itself. It produces a confident, well-formatted answer that reads as authoritative until someone with the right knowledge examines it carefully. When that review step is missing, the output moves forward as fact. The Alabama Supreme Court sanctioned an attorney who had filed legal briefs containing inaccurate AI-generated citations, including references to cases that simply did not exist. The attorney did not intend to mislead. The model was not asked to fabricate. But there was no human in the loop to catch what the model got wrong before it reached the court. That is the risk. Not that AI produces errors, but that those errors reach consequential places when no one is checking.
Human-in-the-loop is not a technical feature. It is a governance decision: designing workflows so that a person with the right knowledge reviews outputs for accuracy and completeness before they influence real decisions. It is also the first thing cut when teams are under pressure to move fast. Organizations that build that review step in from the start treat it not as a check on the technology but as a check on the consequences of trusting it without one.
Governance, cost architecture and responsible AI practice are not separate conversations. They are three dimensions of the same challenge, and the organizations that bring them together will be best positioned to scale AI with confidence. The shift from AI capability to AI economics will become one of the defining leadership conversations of the next decade. Getting governance right is not just about cost. It is about building AI that people inside and outside your organization can trust.
This article is published as part of the Foundry Expert Contributor Network.Want to join?