Enterprise software used to be sold by the seat. A company paid for each employee using a productivity suite, CRM platform or collaboration tool. The pricing model was familiar, predictable and easy for procurement teams to understand.
AI agents are beginning to complicate that logic.
An agent is not just a user interface. It is a chain of model calls, tool calls, searches, file reads, code executions, retries and validations. A task that looks simple to the employee who requested it can quietly consume tokens, compute and time across multiple systems. In practice, an agent behaves less like a software feature and more like a cloud workload.
That distinction is becoming more important as companies move from AI pilots to production deployments. A chatbot usually responds once. An agent may plan, act, observe, revise and act again. It may search an internal knowledge base, call a browser, generate code, edit a document, check the result and retry if something fails. Each step adds cost and latency, and each failure can multiply both.
For enterprises, this changes the metrics that matter. The question is no longer only whether an agent can complete a task. It is how much the task costs, how long it takes, how often it succeeds, which model it uses, how many times it retries and whether a cheaper model could have produced the same result.
In other words, the agent is becoming a workload to be scheduled, routed and optimized.
That shift helps explain why the major AI platforms are no longer selling agents as standalone assistants alone. Microsoft is building around Copilot, Copilot Studio, Azure AI and Microsoft 365 workflows. Salesforce is embedding Agentforce into CRM and customer operations. Google is tying Gemini into Workspace, enterprise search and agent-building tools. Anthropic is pushing Claude deeper into developer and workplace contexts. OpenAI is moving from ChatGPT as an assistant toward agents that can browse, use tools and complete tasks across software environments.
The common thread is that enterprise AI is becoming a stack, not a single app. Companies need models, routing, connectors, developer tools, workflow orchestration, monitoring and cost controls. A polished agent interface may be what users see, but the more valuable layer may be the infrastructure that decides how work is executed behind the scenes.
This is where the economics of agents start to resemble cloud computing more than traditional SaaS. A SaaS license tells a company roughly what access costs. An agent workload can vary depending on task complexity, model selection, context length, tool use and retry behavior. Two users may both ask an agent to “prepare a report,” but one request might require a short summary while another triggers a long chain of search, analysis, document generation and verification.
That variability makes model routing a core part of the agent stack. If every task runs on the most expensive frontier model, costs can rise quickly. If every task runs on the cheapest model, reliability may suffer. The practical answer is a routing layer that can match the task to the right model: stronger models for complex reasoning, cheaper models for high-volume routine work, coding models for software tasks, vision models for image-heavy workflows, and multimodal systems when the task requires them.
Tencent’s recent AI rollout fits into this broader shift. Products such as WorkBuddy, QClaw, TokenHub, Agent Development Platform and Hunyuan show how one major technology company is trying to connect model capability, agent interfaces and model-service infrastructure. But Tencent is not alone in moving this way. Across the industry, the leading players are converging on the same idea: agents need more than intelligence at the front end. They need execution infrastructure at the back end.
That may become the more important competition. Consumer agents can win attention with personality, convenience or a polished interface. Enterprise agents have to win on throughput, reliability and cost per completed task. A company does not deploy an agent at scale because it looks impressive once. It deploys one because it can complete thousands of tasks inside a predictable cost and performance envelope.
The vocabulary of AI adoption may change with this shift. Instead of asking how many users an agent has, enterprises may ask how many workflows it completes. Instead of measuring engagement time, they may measure task success rate. Instead of evaluating only subscription price, they may calculate cost per successful execution.
That puts a different kind of pressure on AI vendors. The winners may not be the companies with the flashiest agent demo, or even the single strongest model. They may be the ones that can combine models, cloud infrastructure, workflow tools and application surfaces into a system that makes agent execution measurable and manageable.
The agent may still be what users see. But the workload layer is what enterprises will pay for.
The next AI agent race may not be about who builds the most autonomous assistant. It may be about who can run the most efficient one.