An AI coding assistant is an AI model built into a developer's code editor that suggests the next line of code, or a whole function, as the developer types. A code editor is the program in which developers write code, as a word processor is the program in which people write letters. The assistant is also called an AI code completion tool or an AI pair programmer.
The assistant runs on a large language model (LLM) that was trained on huge amounts of public code and text. The model usually runs on the servers of the company that sells the assistant, not on the developer's laptop. A server is a computer in a data centre that runs programs for other computers over the internet. When the developer pauses, the editor sends the model a few hundred lines from around the cursor and from other open files. The model predicts the code most likely to come next. In under a second, the suggestion appears in faint grey text. The developer presses the Tab key to accept it, or keeps typing to ignore it.
Consider a developer at a telecom operator in Gurugram. She is writing the recharge form, where a customer enters a mobile number. A function is a small, named block of code that does one job. She types a comment: "check the number has 10 digits and starts with 6, 7, 8 or 9". The assistant suggests a six-line function that does exactly that. She reads it and accepts it. Then she asks the assistant's chat panel for tests. It suggests three: a correct number, a 9-digit number and a number that starts with 5. She runs them herself, and all three pass.
Most assistants now do more than complete lines. A chat panel explains unfamiliar code, and some assistants can edit several files when asked. When the tool also runs commands and tests by itself, it becomes an AI coding agent. The difference is who decides each step. With an assistant, the developer decides. An agent carries out a whole task and returns a finished change.
What can go wrong? The model predicts likely code, not checked code. A suggestion can contain a bug, a security flaw copied from the public code it learned from, or a call to a library that does not exist. So many companies add AI code governance, which checks AI suggestions as they appear. The developer is responsible for every line she accepts. Her finished change still goes through code review before it joins the main code in the code repository.
There is also a privacy question. The company's code is sent to an outside model. So companies usually buy a business plan, under which the provider agrees not to keep the code or use it to train its models. They pay a monthly fee for each developer.
The same tools have changed hiring tests. A coding assessment platform must now detect answers that an assistant wrote instead of the candidate. The models inside these assistants are foundation models, built by a few large AI companies and used by many products.