What is narrow AI?
Narrow AI, also called weak AI, is an AI system designed to perform one task or a closely related group of tasks. A model that predicts which insurance claims are likely fraudulent cannot also schedule the adjusters. A language model that drafts customer emails does not know how to read a sensor stream. “Weak” is a technical label and says nothing about quality; a narrow system can far exceed human performance within its task.
Narrow AI is the opposite of artificial general intelligence, the hypothetical system that could learn any intellectual task. Even large language models, which handle many kinds of text, are narrow in this sense. They operate on one modality or a few, inside the limits of their training data, and they do not form their own goals.
Why narrow AI matters
Understanding that all deployed AI is narrow leads to better project decisions. A narrow system needs a precisely defined task, examples of correct output, and a clear boundary for what happens outside its scope. Projects succeed when the task is specified as tightly as the technology, for example “classify incoming support tickets into twelve categories with confidence scores” rather than “use AI to improve support.”
Narrowness also simplifies governance. A system that does one thing can be measured against one set of metrics, audited against one set of rules, and switched off without affecting unrelated work.
Narrow AI in practice
A collections agency deploys three separate narrow systems: a model that predicts which accounts are likely to pay after contact, a speech recognition model that transcribes calls, and a language model that summarizes each transcript for the account record. Each is evaluated separately, and each has a human review path. Together they change how the floor operates, but no single component understands the business.
What to watch for
- Write the task definition before evaluating vendors, and reject tools that cannot state what they do and do not handle.
- Expect a system that performs well on one task to fail silently when given adjacent tasks it was not trained for.
- Combine narrow systems through explicit workflow logic and approvals rather than assuming one model will cover everything.
- Measure each component on its own metrics so that a failure can be traced to its source.