What is an AI model?
An AI model is a mathematical function that maps inputs to outputs, with its behavior set by parameters learned from data rather than written by a programmer. The input might be a row of customer attributes, an image, or a block of text. The output is a number, a category, a probability, or, for generative models, a sequence of tokens that forms text or an image. A model file is the architecture (the shape of the function) plus the trained weights (the values of its parameters).
Models range from a logistic regression with a dozen parameters to a large language model with hundreds of billions. Training fits the parameters to examples by minimizing a loss function. Inference runs the finished model on new inputs. The model does not learn in production unless it is retrained; what it knows is frozen at training time.
How an AI model works
Three things determine a model’s behavior: the architecture, the training data, and the training procedure. A decision tree, a convolutional neural network, and a transformer are architectures. Each has strengths for certain data types, which is why tabular business data is often handled well by gradient-boosted trees while documents and speech go to neural networks. Training data defines what the model can recognize. A demand model trained on three years of a distributor’s shipments knows that distributor’s seasonality and nothing about a competitor’s. The training procedure, including hyperparameters and regularization, decides how well the model generalizes to data it has not seen.
Why an AI model matters
For a business, the model is the asset and the maintenance item at once. A trucking company that builds an arrival-time model owns something competitors cannot copy, because it reflects the company’s own lanes, customers, and history. The same model drifts as routes, fuel prices, and customers change, so it needs monitoring, periodic retraining, and a model registry that records which version is live. For language models, the choice between a closed model accessed through an API and an open-weight model run on company infrastructure decides who controls the asset.
What to watch for
- Ask what data a model was trained on and whether that data resembles your own, because performance does not transfer automatically.
- Distinguish the model from the AI system around it, since most controls, logs, and approvals live outside the model.
- Record the version, training data, and evaluation results of every deployed model so that behavior changes can be explained.
- Budget for retraining and monitoring, since a model is finished only when it is retired.