Sheet 6
Where it fails, and what to do
AI Tools Field Guide
Not affiliated with OpenAI, Google, xAI or Anthropic; your organization’s rules come first.
Most failures are a missing source or an unchecked claim.

It does not know
- Your command. Confident answers about your instruction or roster when nothing was attached. Attach the source.
- This year. Its knowledge has a cut-off, and web search is a periodic index. Attach the current version, or keep it in a project so every chat has it.
- Arithmetic. Watchbill counts and per-diem that look right and were never computed. Make it write the formula, or use a tool that computes: Grok’s skills, Excel, PowerShell.
- Page 60 of 100. Long files are read unevenly. Attach the section, or ask about it by name.
It invents
- References in the right format that do not exist or do not say that. Make it quote the sentence; a second chat checks each one against the file.
- A complete-looking answer with the gaps filled in fluently. Ask “what is missing from the source?”
- A version you rejected, returned to in a long chat. Start fresh with only the source and the request.
- Instructions hidden in a document it read as if you had typed them. Treat outside files as untrusted; a person reviews before anything acts.
Habits that prevent most of it
- Give the source and say that “not in the source” is an acceptable answer.
- Ask for references beside each claim and check the ones you will act on.
- Pick the careful model for anything you will brief or sign.
- Keep to the data rules: public tools take public material, a work account may take more within its rules, secrets and other people’s data never.
Take away
It is a language processor, not a knowledge store. Source in, references out, your eyes on what matters, and most of this list never bites.