A language model learns patterns from training data and generates tokens in response to the input it receives. It does not automatically see your organisation’s current systems, know which document is approved or verify every sentence it produces. The surrounding application determines what information and tools are available.
What the model can see
The model receives a context assembled from instructions, user messages and any supplied material or tool results. A file on your computer is not available merely because you know it exists. A website may not be accessible unless a browsing tool retrieves it. A screenshot exposes visible information, not necessarily the underlying database.
Multimodal models can process supported images and other inputs, but interpretation still depends on the model and the quality of the input. Small text, unclear charts and missing context can lead to mistakes. Provide the original data when precise calculation matters.
Why a confident answer can be wrong
Fluent text is not proof that a claim was checked. The model can produce a plausible continuation when evidence is missing or ambiguous. Asking it to sound authoritative makes that problem worse if you do not also require evidence and uncertainty handling.
For a company policy question, supply the current policy or use retrieval. For a numerical calculation, use validated code or a calculator and inspect the inputs. For current information, use appropriate sources and check dates. Match the verification method to the task.
Give it an explicit success test
Instead of “analyse these sales,” define the period, currency, treatment of refunds and the decision you need to make. Ask the assistant to identify missing fields before calculating. Provide an example of the expected output and require assumptions to be visible.
For a writing task, explain the reader and the action the text should enable. For a software task, describe the user journey and failure cases. The model can infer many things, but your business requirements should not depend on a lucky inference.
Treat external content as evidence
Documents and web pages can contain misleading text or instructions unrelated to your task. A well-designed system separates user-authorised instructions from retrieved material. Tools and permissions should enforce boundaries even if the model is presented with hostile or confusing content.
Keep sensitive actions behind appropriate application controls. The model’s confidence should never be the only authorisation for changing a record or exposing restricted information.
Work in a verification loop
Provide relevant context, request a bounded result, check the evidence and correct specific errors. Preserve the corrected decisions for subsequent work. This makes AI useful as a collaborator while keeping responsibility for the outcome clear.
Sources & further reading
Primary references for the technical background and regional statements in this guide. Planning examples and checklists are Novacom’s practical guidance; examples are illustrative unless explicitly identified as project experience.