Interaction Model
An interaction model describes the fixed rules governing how a human and a technical system deal with each other: which inputs are allowed, what they mean, and what happens as a result. In voice assistants and chatbots, it is the structure that turns a spoken or typed sentence into a concrete action.
When you talk to a device or press a button, both follow a fixed pattern. You provide some input, the device maps that input to a particular intent, and then it responds. This exact pattern is called an interaction model. It defines which inputs may occur at all, how they are interpreted, and what response or action follows. An interaction model is therefore not a program, but rather the blueprint for the conversation between human and machine. You can think of it like the rules of a board game: they state which moves are allowed and what each move does.
Why voice assistants fail without a fixed blueprint
People say the same thing in many different ways. “Turn on the light,” “lights please,” “can you make it brighter in here” all mean the same thing. A system needs a structure that maps all these variants back to the same command. Without this structure, you would have to write a separate rule for every possible sentence, and that is impossible.
A clean interaction model also determines how a product feels. If a system politely asks a follow-up question when information is missing, it comes across as competent. If it instead aborts with an error message, it feels broken. Many complaints about voice assistants are not actually about speech recognition, but about a poorly designed interaction model behind it.
For companies, this is also a matter of cost. An interaction model that reliably covers the twenty most common customer requests takes a noticeable amount of work off a support team’s plate. A vague model, on the other hand, produces incorrect responses that someone later has to correct.
Intents, slots, and sample utterances
Classic interaction models consist of three building blocks. The first is called an intent, meaning the purpose behind an input, for example “set an alarm.” The second are slots, the variable pieces of information within that intent, such as the time or the day of the week. The third are sample utterances, which developers use to show the system how users typically express the intent.
If a required slot is missing, a follow-up question is triggered. If you only say “set me an alarm,” the system responds with “for when?” These follow-up questions are part of the model and not just a coincidence. This is also called dialog management: the system knows which piece of information it is missing and asks for it specifically.
With modern language models, meaning AI systems that generate freely formulated text, this rigid grid becomes blurred. Such systems no longer need pre-written sample sentences. Nevertheless, an interaction model still exists, it is just described differently: through instructions given to the model and through a list of tools it is allowed to call. A common misconception is the assumption that with language models, structure becomes unnecessary altogether. Without clear boundaries, a system will respond even in cases where it would be better to stay silent or hand off to someone else.
From Alexa Skills to the support chatbot
The term is most visible in voice assistants. Anyone building an extension for Amazon's Alexa actually uploads a file there literally named Interaction Model. It contains the intents, the slots, and the sample utterances. Google and Apple also work with very similar concepts in their developer tools.
In everyday life, you encounter the same principle in the chatbot of a bank or an online shop. When a menu with four options appears there, that is a deliberate decision of the interaction model: it narrows down the inputs so that less can go wrong. The voice menu of a phone hotline follows the same pattern too, just using number keys instead of text.
In news about AI products, the term usually comes up when companies rebuild their assistants. If a report states that a provider has overhauled its interaction model, that means: the rules for inputs, follow-up questions, and permitted actions have been redefined. The term should be distinguished from the user interface, meaning the visible surface with buttons and colors. The interface is the surface, the interaction model is the logic behind it.
