Plain-language explainer
What is a large language model (LLM)?
A large language model is an AI system trained on large amounts of text to predict and generate language. Tools such as ChatGPT, Gemini and Claude are built on LLMs.
The work to plan together
A scope shaped around this requirement.
- LLMs generate text from patterns in trainingScope
- Some connect to web search for currentReview
- They can be wrong and should beCheck
The useful starting point
An LLM generates language by modelling patterns learned during training. It can produce fluent text without knowing whether every statement is correct. Some applications add search, retrieval or other tools, which changes what information the application can use.
What to review
- LLMs generate text from patterns in training data
- Some connect to web search for current information
- They can be wrong and should be checked
- Businesses use them for chat, content and automation
A practical example
A business assistant may retrieve an approved policy and draft an answer from it. The useful result depends on retrieving the correct policy and checking the response, not simply choosing a model.
The numbers or scenarios in this guide are illustrative. They are not claimed Tangensys client results.
What to avoid or interpret carefully
Fluency is not proof. Check facts, sources and permissions before using model output for a customer decision.
Turn the guide into a small action plan
Record the starting point
Use the example above to identify what is happening in your own website, campaign or process. Save the evidence before changing it.
Choose the first check
Start with the relevant items in the review list. Assign an owner, identify required access and write down what you expect to learn.
Review before expanding
Compare the result with the original evidence. Keep useful changes, explain uncertainty and investigate problems before increasing scope.
Separate language generation from verified knowledge
A language model can produce a plausible response even when the required fact is absent. Applications may add retrieval, search or tools, but those additions have their own limits. Judge the complete application, not only the model name.
Source access
Identify whether the application uses approved documents, current web information or only its existing model. Ask how it selects sources and whether a user can inspect them.
Action boundary
Drafting a reply differs from sending it or changing a record. Name the permitted action and approval owner. A natural-sounding instruction does not establish that the action is authorised.
Evaluation
Use representative correct, ambiguous and out-of-scope questions. Review factual accuracy and failure handling. A successful demonstration on familiar questions is not enough to establish reliable operation.
How to interpret this guidance
For business use, the source owner, interface and human review often matter as much as model choice. Keep sensitive information and access permissions within the client’s approved process.
Why can an LLM answer confidently and still be wrong?
Fluent generation is not the same as checking a fact. Missing context, unsuitable sources or incorrect interpretation can produce an unsupported answer. Use source review and an explicit fallback where correctness matters.
Can Tangensys help with the next step?
Yes. Share your website, goal and the issue you are investigating. We can discuss the relevant service and provide a custom quote within 2 working days.
Related reading and services
Let’s agree the right next step.
Tell us your goals and current setup. We’ll recommend a practical starting point and provide a custom quote.
Custom quote within 2 working days