The modern payment infrastructure assumes well-defined inputs by users, involving specifying parameters such as the amount of the transaction, payee identity, and timing of payment execution. This mode of operation causes procedural overhead and constrains the possibility of abstraction, especially with the ongoing evolution of financial products towards greater userfriendliness. This article presents a conceptual framework for intent-driven payments, in which users formulate their financial intentions in natural language and large language models (LLMs) are used to generate corresponding structured payment workflows. While the term has been used informally in prior industry commentary, this work offers a structured architectural treatment of the paradigm. The framework described here is proposed and has not yet been empirically implemented. The intentto- payment system presented is built as a multistep pipeline, incorporating such stages as intent extraction, entity recognition, constraint validation, and orchestration. One significant novelty of this work concerns the development of the financial intent compiler, which is designed to enforce that the outputs generated by the system are deterministic, transparent, and aligned with applicable regulatory constraints. This article touches upon a number of topics relating to the design of systems, such as latencyrelated problems, handling ambiguity, considerations of security, and verifications of computations made by people.