THE INSTITUTE / RESEARCH TOPICS
AI Models
What models do, how versions differ, and how they fit into business software.
Explore this topic
AI Models
Choose a layer to understand the tools and buying questions.
The quick answer
An AI model is a trained system that turns inputs into predictions, classifications, or generated outputs. A language model can draft a reply; an image model can interpret or create images; another model might predict demand. The model is one part of the software your business uses.
For a buying decision, separate four things: the model family, the exact version, the company that developed it, and the provider that runs it. Those choices affect what you can test, where data goes, and what you pay.
What it is
Training adjusts a model’s learned parameters using examples and an objective. Inference applies the trained model to new input. The application supplies the instructions, context, permissions, and user experience around that model.
A useful output is not automatically a verified fact or an authorized business action. Those depend on the surrounding system, its checks, and the responsibilities your organization assigns.
Know the difference
A model is not the whole application
A chat interface, search system, or agent can use a model, but also needs software for data access, tools, identity, records, and controls. A product demonstration measures that combined setup.
A family name is not an exact version
Products within a family can differ in cost, supported inputs, speed, size, and behavior. Record the model identifier and test configuration so another team can reproduce your evaluation.
How it works
1. Define the input
The application prepares the request and the information the model may use.
2. Run the model
The model processes that input and returns an output. Depending on the model, that may be a prediction, generated content, or a proposed tool call.
3. Check and use the result
The application validates the output, applies permissions, and decides whether to present it, take an action, or ask a person to review it.
Types of tools
Platforms and tools
Business uses
For invoice processing, test whether the complete system extracts the right fields and flags uncertain cases. For a support assistant, test whether answers follow current policy and whether account changes require approval. For forecasting, compare the model with your existing approach on representative periods.
These are evaluation examples, not measured customer outcomes.
Benefits and costs
Limits and risks
Models can fail on unfamiliar inputs, produce plausible errors, and perform differently after changes to instructions or deployment. A benchmark can inform a shortlist, but cannot establish performance on your business task by itself. Keep representative test cases and repeat the relevant checks when the model or surrounding system changes.
What the evidence shows
How we assess it
This guide combines primary documentation with the Institute’s questions for a buying team. It does not report a survey, independent benchmark, or vendor score.
Questions before you buy
- Which exact task and model version are we evaluating?
- What counts as a correct, usable result?
- What data and tools does the surrounding application need?
- Who runs the model, and what data terms apply?
- What happens when the model fails or the provider changes a version?
- What is the cost per accepted result, including review and retries?
Common questions
Sources and reading
IBM — What is an AI model?: model concepts and training.
IBM — What is LLM inference?: running a trained language model.
OpenAI model documentation: examples of version-specific capabilities.
Google model documentation: examples of model variants and supported features.
The buying questions are the Institute’s synthesis. This introductory guide is not a model leaderboard.
What changed
October 24, 2026 — First publication as part of the Models research category.