Model

An AI model turns an input into a prediction, label, or generated result. It can do part of the work, but you still need a system around it that supplies current client data, sets permissions, checks the result, and sends exceptions to someone who can handle them.

And choosing a stronger model isn’t always the answer. In July 2026, one production agent system I run used models across more than 2,000 tasks, the median active conversation cost about eight cents, and the most expensive task cost $16.12.

Moving routine work off an oversized model cut the bill by about 20 times without changing the accepted result. So choose the model inside a real workflow and measure the finished work, because a stronger model can’t supply a record you never gave it or decide who owns a failure.

How does an AI model produce an output?

A model turns an input into an output through a structure and learned values. NIST defines an AI model as a component of an information system that uses computational, statistical, or machine-learning techniques to produce outputs from inputs.

Most modern models learn those values during training, a process that shows the model data, compares what it produces with a target or goal, and adjusts it. Google’s machine-learning guide describes the result as a complex collection of numbers that captures relationships between the patterns going in and the values coming out.

Using a trained model is called inference. A spam model takes an email and estimates how likely it is to be spam, while a demand model uses past sales and current conditions to predict a number.

A language model takes tokens, the pieces of information it processes, and generates more tokens. Each model produces the kind of result it was built for.

The model does not keep one neat row for every fact in its training data. Its learned values capture relationships between patterns. A generated answer can resemble a source without retrieving that source. Current client records must reach the system another way.

What is the difference between a model and an AI system?

An AI system combines one or more models with the parts needed to use them for a purpose.

The OECD’s March 2024 explanatory memorandum draws this line. It calls a model a core component used to make inferences. It describes a system as models and other parts built around an objective.

Think about a project update. The model drafts it from notes, while context supplies the current scope, project status, and the terms your firm has agreed to use.

A tool reads the project record or saves the approved draft, and the workflow sets the order, rules, and stopping points, including sending missing facts to the project owner. Evals and review check dates, scope, unsupported claims, and tone so you can tell whether the result meets the standard.

ChatGPT is an application. The model inside supplies language ability. The application adds the interface, conversation history, file handling, web access, tool connections, and product controls. Two applications can use the same base model and still behave differently because the surrounding parts differ.

That difference helps you find the problem. If a report uses an old scope, a stronger model may write the wrong report more clearly, because the system still hasn’t supplied the current scope.

But if the model sees the right records and keeps failing the same reasoning test, then model choice may be what needs to change.

Is every AI model a large language model?

Large language models are one family of AI models. They learn broad language patterns and generate text or other token sequences. Other models predict numbers, classify records, rank options, detect objects in images, transcribe audio, or generate images.

A large language model starts broad, so it can handle many language tasks. Stanford describes a foundation model as a large model trained on broad data that can be adapted to many tasks and applications.

That training gives you more possible uses, but it doesn’t make the model right for every step.

Exact tax calculation belongs in ordinary code. A stable yes-or-no routing rule may need a small classifier. A client interview summary fits a language model because the language changes from call to call. One workflow can use all three and still present one finished result to the team.

What changes the model, and what changes only the request?

Training changes the model’s learned values. Fine-tuning continues that process with additional examples or feedback to shape behavior for a narrower need.

A prompt usually leaves the base model unchanged, it gives the current run instructions, examples, source material, and a format to follow. The answer may improve because you gave the model better input, while the values it learned stay the same.

Retrieval and memory work outside the model as well. They select information and place it into the current context. This matters for private facts. A firm that needs the current contract amount should retrieve the approved contract. Training that amount into a model would make the fact harder to update, prove, and remove.

How should a business choose a model?

Choose the smallest model that passes your test for one named task. Before asking which model is best, you need to know what it will do, which sources it gets, what a mistake costs, how fast it needs to be, and what you can spend.

Start with real examples from the workflow, including ordinary work, unusual cases, incomplete inputs, and cases that should stop for a person. Write down what counts as acceptable before comparing models, then run the same set through each candidate and record quality, response time, model cost, and human review time.

The costs can vary a lot even inside one system. Across more than 2,000 agent tasks I measured in July 2026, the median active conversation cost about $0.08, while one complex task cost $16.12.

I had put an oversized model on routine work, and changing that choice cut the bill by about 20 times with the same accepted result. The full numbers and their limits are in the agent rate-card analysis.

A cheap model that creates ten more minutes of review can cost more than an expensive model that passes cleanly. A capable model used for every small classification can waste money without improving the finished work. Compare model cost and review cost together.

Keep the model replaceable by storing the prompt, test cases, required output, and workflow rules outside code tied to one provider where practical. When you upgrade the model, rerun those tests rather than making the firm work out the process again.

What must exist around the model before client work?

The surrounding system needs a named task and a source of truth. It also needs the right context, limited permissions, a completion check, and an owner for exceptions.

Suppose a model drafts a client status report. The system has to fetch the right project and current scope, separate completed work from planned work, block unsupported promises, and send missing information to someone who can decide.

The model can write the draft, but your firm has to supply the facts and rules it works from.

Write that process down before you automate it. A model cannot resolve a rule that changes by manager and lives only in memory. Document the process first so every run receives the same source, definition of done, and exception path.

Set permissions by what can go wrong. Research and drafting can tolerate a cheap retry. A client promise or deleted record needs a harder boundary. The decision to delegate by how easily work can be checked and undone still applies when a newer model arrives.

Where does a model sit on the Delivery Model Ladder?

Direct model use sits mainly at Stage 1, Enhanced, on the Delivery Model Ladder. An employee opens an AI application, supplies the input, checks the output, and carries it into the next system. The person moves faster while the delivery path stays mostly intact.

At Stage 2, Augmented, a workflow calls one or more models for specific, limited steps. It supplies approved context, checks results, controls actions, and sends exceptions to an owner, so model choice becomes one setting inside a process you’ve changed.

Stage 3, AI-native, depends on shared operating context and an owned system across the firm. A larger model cannot create those conditions. Treating a model demonstration as a finished operating change is one way firms enter the Production Gap.

When is an AI model the wrong tool?

Use ordinary software when a rule is exact and stable. Arithmetic, database constraints, fixed eligibility rules, and irreversible account changes should not vary because a model produced another plausible answer.

A model can handle the messy language around an exact step. It can extract invoice fields from an email while code calculates tax and checks the total. It can classify a request while a permissions rule decides whether any action is allowed.

This assumes you can name the task and test the result. If nobody can say which source to trust or what counts as acceptable, keep the work manual until you’ve settled those questions.