Do AI agents use machine learning models, or just language models?

What is the actual difference between the two?

What they output and how they are checked. A language model produces text, and you judge it by reading it. A trained model produces a number with a confidence figure, and you judge it against what actually happened. Only the second can be wrong in a way you can measure.

That measurability is the whole point. If a model says a deal has a 70% chance of closing this month, you can compare that against a hundred deals it scored at 70% and see whether roughly seventy of them closed. There is no equivalent test for a paragraph of persuasive text.

It also changes how you should treat the output. Text from a language model is a draft to review. A number from a trained model is evidence to weigh, and it comes with an honest statement of how often it is wrong.

Can a language model do lead scoring?

It can produce a ranking, and it will sound convincing. What it cannot do is calibrate: tell you that leads it scores at 80 convert about 80% of the time. Without calibration a score is an opinion with a number attached, which is worse than no number at all.

Worse, because a number implies precision that is not there. A founder given a ranked list assumes something measured it. If nothing did, they will work the list top-down and never discover that the ordering was arbitrary, because there is no feedback loop that would tell them.

This is not an argument against language models. Writing the follow-up email, summarizing a support thread, drafting an article — these are exactly what they are for, and they do them well.

How do you tell which one a platform is using?

Ask what it predicts, what it takes as input, and how accurate it is. A platform running trained models can answer all three. A platform running a language model with a confident prompt will answer the first, hedge the second, and change the subject on the third.

A second test: ask for the list. A platform with a real model library can show you one, with names and what each model does. A platform advertising a large number of models but unable to name them is advertising a number.

newc0 publishes its model library with two confirmed domains named, each with what it predicts, what it needs as input and what it will not tell you. There is no model count anywhere on this site, deliberately, because a count with nothing browsable behind it is unverifiable.

Where does a trained model get its training data?

From outcomes the business has already recorded — deals that closed or did not, leads that converted, tickets that were resolved. That is why these models improve with time and are weak at the start, and why a vendor promising strong predictions on day one is describing something that cannot exist.

This has a practical consequence for how you set a company up. The value of a prediction later depends on the quality of the record now, which means recording outcomes properly from the beginning is an investment rather than administration. A business that never marks why a deal was lost has nothing to learn from.

It also explains why the application matters as much as the agents. Agents with nowhere to write outcomes produce activity that leaves no trace, and a model cannot learn from activity it cannot see. The system of record is what makes prediction possible at all.

Ask any vendor two questions: what data does the model learn from, and what happens before there is any. Vendors with real models answer both readily, including the unflattering half.

Why a company needs both

Language agents do the work that is written: outbound, replies, drafts, summaries. Trained models do the work that is counted: which lead, how much, how likely. A company running only the first produces a lot of text and no idea which of it mattered.