Language models and models trained on your own data
Two kinds of AI dominate business conversations: large language models, which power chat assistants, and predictive models trained on an organisation's own structured data. They solve different problems, fail in different ways and need checking in different ways.
Large language models (LLMs)
A large language model is trained on a very large amount of text to predict which words are likely to come next. From that simple objective it learns to answer questions, summarise, translate, draft and write code. Chat assistants are built on these models.
What makes LLMs useful is their breadth: one model can help with thousands of different tasks expressed in ordinary language. What limits them is that the output is plausible text, not a checked fact. The model is not looking anything up unless the product around it has been connected to a source, and even then it can misread or misquote that source.
Predictive models trained on your own data
A predictive model learns from rows of structured records, such as past orders, customers, transactions or sensor readings, with a known outcome for each row. It learns the relationship between the inputs and the outcome so that it can estimate the outcome for new rows.
Typical questions: Will this customer renew? How many units will we sell next week? Is this transaction likely to be fraudulent? The answer is a number or a category, often with a confidence level, and it can be compared directly with what actually happened.
Side by side
| Language model | Predictive model on your data | |
|---|---|---|
| Learns from | Huge amounts of general text, mostly not yours | Your own historical records |
| Output | Free text | A number or category for each new record |
| Best at | Language tasks: drafting, summarising, explaining, extracting | Repeated, specific decisions with a measurable outcome |
| Typical failure | Confident, fluent statements that are false | Quietly getting worse when conditions change from the training data |
| How you check it | Human review, comparison with trusted sources, task-specific tests | Accuracy measured on held-out records with known outcomes |
Choosing the right tool
A useful rule of thumb: if the task is about words, a language model may help. For repeated decisions with reliable historical outcomes, a task-specific predictive model is often worth evaluating against simpler alternatives. Its value depends on whether it performs well on representative unseen data; with too little relevant history, or if it cannot beat a simple baseline, it is not the better tool.
Asking a chatbot "which of my customers will leave next quarter?" invites a confident guess. A model trained on your own customer history, and tested on customers it has not seen, gives an answer whose accuracy you can measure and compare with what you do today.
Remember: the two kinds work well together. A language model can explain a prediction in plain words, but the prediction itself should come from a model whose accuracy has been measured on your data.