In the first quarter of 2026, a quiet signal surfaced at Microsoft's earnings call. Cloud revenue had grown sharply year over year — but the company attributed the growth not to an increase in paid accounts, but to a surge in AI agent calls. Few outside the room paid much attention to the distinction, yet it marked the moment a thirty-year convention in enterprise software began to wobble, right there on an earnings call.
I recently heard a bewildered account from the head of a small company that had adopted Microsoft Copilot. He had expected a fixed bill based on his fifteen-person team. Instead, the invoice carried extra line items tied to how many emails the agent had answered, how many data lookups it had run, how many document drafts it had generated. The very premise that software is a fixed cost, he said, had been shaken.
That unfamiliar feeling is about to become a common one. With the arrival of agents, the way software costs are calculated is itself changing.
When the Agent Works, the Meter Runs
Microsoft has begun running agent-action-based billing in earnest across its AI product line. Every time an AI agent configured in Copilot Studio sorts or answers an email, queries or cleans up a database, or generates a report or document draft, a per-task rate applies. The billing unit varies with the type and complexity of the work. Use more and the bill goes up; use less and it goes down. It is not a flat subscription.
Behind the shift lies a mismatch between usage and value. Under the old subscription model, companies licensed by headcount — so a heavy AI user and someone who barely touched the tools generated exactly the same cost. AI agent usage, by contrast, is wildly asymmetric. One team might push thousands of tasks a day through agents while another never uses them at all. Seat counts simply fail to reflect actual usage.
Apple's results from the same period connect to the same current. As Apple Intelligence features expanded, per-device memory requirements climbed, and demand for high-capacity memory chips began to outrun supply capacity. Mac shipment schedules came under the direct influence of semiconductor procurement. Even as AI reshapes software cost structures, it is reorganizing hardware supply chains alongside them.
The two giants' earnings reports point to one and the same thing: AI has begun operating as a core variable in both IT budgets and product strategy. And this is no longer a concern reserved for big-company CFOs.
What the Collapse of a Thirty-Year Formula Reveals
Per-seat pricing for enterprise software took hold in the 1990s. The logic was intuitive: software is a tool that people use, so charging in proportion to the number of people using it made sense. Microsoft Office, Salesforce, Slack — the products that grew up on this structure became the standard of the enterprise software market. Forecasting was easy for buyers, too. Multiply headcount by the unit price and you had your IT budget.
AI agents upend that premise. An agent is not a tool a human sits down and operates directly. It is an actor that executes work autonomously, around the clock. A twenty-person company can run a hundred agents simultaneously and process thousands of tasks a day. A five-hundred-person enterprise that never deploys agents will rack up almost no metered charges. Seat counts no longer stand in for actual usage.
Hence Microsoft's change of model: the volume of work an agent performs is the value delivered, and the bill should be proportional to that value. Infrastructure services like AWS and Google Cloud have worked this way for years — pay for server hours, for storage capacity, for API calls. That structure is now spreading into the software application layer.
This change in cost structure also changes what matters to manage. In the per-seat world, the IT manager's core job was contract negotiation and license management: sign a deal based on headcount and make sure no unnecessary seats piled up. In the metered world, how an agent is designed to carry out which tasks translates directly into cost differences. Workflow design skill becomes synonymous with cost-control skill.
The more fundamental shift concerns what sits at the center of professional competence. What people need in the AI era goes beyond knowing how to operate the features of AI tools. It takes understanding how agents behave, designing that behavior to fit a purpose, and judging and adjusting the results. Which tasks should a person handle directly, which should be delegated to an agent, and how should the agent's output be verified — building that decision framework is the heart of it. We are now at the point where the gap between using a tool and understanding how the tool works starts showing up as cost differences, and as differences in outcomes.
This is not a matter of technical knowledge. The ability to design what to ask an agent to do, how to ask it, and how to evaluate what comes back — that is the posture and the way of thinking required of anyone collaborating with AI. More than which AI tool you use, we are entering an era where the judgment framework of the person handling the AI is what makes the difference.
The Advantage Goes to Those Who Use Less, but Better
Look at what this shift concretely means for solo entrepreneurs and small teams in Korea, and a structural opportunity comes into view first. The per-seat model penalized people who worked alone: one person shouldered the entire license cost, with none of the negotiating leverage that comes with a large contract. Under metered pricing, someone who designs and deploys agents efficiently pays exactly that much and no more. A one-person business running a well-built agent workflow can plausibly match or outproduce a mid-sized team with no agent strategy — at a lower AI cost. The penalty of being small shrinks.
But that possibility does not realize itself. A few things have to be in place.
Designing agent trigger conditions with precision is a good place to start. Running an agent on an hourly schedule versus configuring it to fire only when specific conditions are met makes a meaningful difference in billable events. Design the agent to trigger only when an email containing certain keywords arrives, or only when a particular file is updated, and unnecessary call costs fall away. This is not coding ability; it belongs to the domain of logical design.
The habit of calculating agent ROI also needs building: compare the cost of putting an agent on a given task against the business value its output creates. Automating every repetitive task is not the answer. Sometimes the cost of automating runs higher than handling the work directly. Having the criteria to make that call is the core skill.
It is also worth re-examining the billing structure of the AI tools you already use. If a service you rely on already includes usage-based components, an analysis of your usage patterns to date becomes the baseline for forecasting future costs. Knowing which tasks were called and how often stops being a technical chore and becomes part of managing the business's costs.
This transition will not complete itself overnight. Even Microsoft is moving through a transitional period in which per-seat and metered pricing run side by side. None of this means you must overhaul every contract and workflow today. But understanding where the current is headed — versus not understanding it — produces different options at contract-renewal time, at tool-selection time, at team-restructuring time. The prepared get to design better terms.
The criterion for choosing AI tools is shifting from “what can it do” to “how does it charge.” In an era where understanding how agents operate and designing your cost structure determines business efficiency, making sense of a single software invoice becomes the starting point of competitiveness.
