Are you building for humans or for AI?

Some time ago, the web stopped being made only for people. Since the beginnings of search engines, we have written for machines, so that search engines could read, index and rank our websites.

What changes now is the kind of machine on the other side. If we once optimized titles and meta tags to appear in searches, today we need to create experiences that language models can understand.

We are designing products that are consumed by humans and interpreted by artificial intelligences at the same time.

Some companies already understand this and are adapting their content to be understood by language models, not merely indexed: “LLM Optimization,” pages designed so conversational agents can interpret context, purpose and structure. The challenge is for AI to understand the content.

In e-commerce, the movement is even more evident. Walmart and OpenAI partnered to allow purchases directly through ChatGPT, ushering in so-called agentic commerce: experiences in which the AI agent does all the work. People no longer enter the app — they simply say, “I want to buy coffee and paper towels,” and the agent takes care of everything.

Behind this, MCP (Model Context Protocol) connects e-commerce directly to AI, exposing data such as catalog, price, delivery and payment so that the agent can understand and execute on its own. Google, in turn, introduced AP2 (Agent Payments Protocol), a standard allowing agents to complete transactions securely and auditably.

With the same protocols applied to the financial system, users do not need to request anything. The agent observes spending behavior, understands patterns and acts automatically.

For example, it notices that monthly expenses have fallen and adjusts the plan: it transfers some money to investments, sets some aside for savings and keeps the rest in the account. If it identifies an increase in fixed expenses, it reduces the contribution and says: “I noticed your bills rose this month. I adjusted your savings goal to R$300 and will reschedule the rest if there is money left over.”

All of this happens in the background, without explicit commands. The agent has access to the balance, history, limits, risk profile and goals, and makes decisions based on continuous learning. It does not ask what you want to do with your money; it already knows what makes sense for you.

In this model, the channel stops being the “app” and becomes a service, connected via MCP to a network of agents that manage finances in real time.

The product dilemma is understanding who is really using your system. When an AI agent interacts with an interface, it does not need a beautiful button, but a well-structured API; it seeks semantics, not aesthetics.

The companies that recognize this first will have an advantage: they will be building not only for human users, but also for the agents that represent them.

Because the question that defines your digital product’s roadmap is: is your product being built for humans or for AI?

📌 Originally published in Época Negócios