A word is the smallest piece that means something by itself
Individual speech sounds and individual letters do not mean anything on their own. A word does: it bundles a meaning together with grammatical properties, such as how it behaves in a sentence, and that bundle is what speakers actually store and reuse [1]. This is the sense in which words are called the fundamental units of meaning, and why dictionaries can be organized around them and sentences analyzed in terms of them [1].
The same idea shows up in how people describe communication: shared understanding of words is described as the basis on which humans rely to communicate [4]. That framing treats words as the common currency between speakers, the level at which meaning is exchanged before anything larger is assembled.
Reading shows the order of operations: words first, then sentences
Evidence from beginning readers makes the layering concrete. First-grade children must first identify written words accurately before they can use the meanings of those words to understand sentences and text [3]. In other words, the word is the gateway from print to language: until the written form is recognized, the meaning behind it cannot be brought to bear on a sentence.
This is a claim about sequence in reading development, not a claim that words are the only thing that matters for comprehension. The finding is that accurate word reading is a necessary earlier step, which is exactly what you would expect if words are the units that carry meaning into larger structures [3].
The word level is also a practical choice for building language tools
The same logic appears outside human language teaching. A paper on online language practice platforms treats words and their forms as the basic building blocks from which varied language exercises can be constructed, noting that a learner must know how a word sounds as part of knowing the word [2]. Separately, a protein-modeling paper draws an explicit analogy to language models that work at the level of learned sub-word units rather than individual characters, arguing that modeling at a meaningful unit level gives both insight and control [5].
These are design arguments rather than proof about human language, and the protein case is an analogy to language models, not a claim about words in speech [5]. Taken together with the meaning and reading evidence, though, they point the same way: the unit that carries meaning is the natural level at which to organize both learning and analysis [1][2].
Sources used in this answer
WORDS SENTENCES AND DICTIONARIES AS MEANINGFUL BUILDING-BLOCKS OF LANGUAGE AND THEIR ROLE IN LINGUISTIC COMPETENCE DEVELOPMENT
Words are analyzed as meaningful building blocks of language, with attention to their semantic and grammatical properties and their role in developing learners' linguistic competence.
Building Blocks for Online Language Practice Platforms
Online language practice platforms are described as built from basic building blocks centered on words and their forms, including how a word sounds.
The Building Blocks of Reading Comprehension: Evidence from Decoding and Word List Recall
Among first-grade children, accurate written word identification is presented as a necessary earlier step before word meanings can be used to understand sentences and text.
Idiosyncratic Tower of Babel: Individual Differences in Word-Meaning Representation Increase as Word Abstractness Increases
Across 90 words and 20 adults, individual disagreement about word meanings increased with word abstractness and with how much language or sensory experience a word involves.
The Words of Proteins: Motif-Level Language Modeling for Interpretable Protein Generation
Modeling protein sequences at the level of functional motifs, analogous to sub-word units in language models, improved specificity of generated sequences compared with amino-acid-level and BPE-tokenized approaches.