This path is for analytics people who will use AI on a real file, not a demo. Parts cover what it can and cannot do, tokens and cost as a feel, prompts for data work, checking an answer, retrieval in plain English, agents, what you should not paste, documentation, and a checklist you can reuse. Start with the first part on this page.
- 1 What AI can and cannot do for analysis AI is a sharp intern for analysis, not a signed-off analyst. Part 1 of Practical AI for analytics people maps assist versus replace, where liability for numbers still sits with you, and when to refuse a generated answer.
- 2 Tokens, context windows, and cost (intuition) Treat context windows like a suitcase: pack role, schema crumbs, and the question. Chunk instead of cramming. Tokens are not free, and middle content gets lost first.
- 3 Prompt patterns for data work Build data prompts with role, spec, output shape, and refuse rules. Make grain, filters, and calendar explicit so the model cannot hide a wrong query behind fluent SQL.
- 4 Evals for humans Build human-sized evals: spot checks, golden questions, and a regression set with clear pass criteria. Score analytics outputs before fluent wrong SQL or prose becomes process.
- 5 RAG in plain English RAG retrieves company passages, stuffs them into the prompt, then generates an answer. Use it for definitions and runbooks with citations. Do not treat grounding as a warranty for board numbers.
- 6 Agents, tools, and harnesses Separate chatbot, script, and agent. A harness is the wrapper of tools and guardrails. Put a human in the loop before agents touch revenue or customer data.
- 7 Privacy when pasting data into chat tools Treat every paste as a processing decision. Use a risk ladder and synthetic samples. Consumer chat is not your approved enterprise workspace.
- 8 AI for documentation and data dictionaries Use AI to draft dictionary entries, then verify against a signed source before publish. Draft, verify, publish is the only flow that stays honest.
- 9 Building a personal AI checklist Close Practical AI with one checklist that travels across SQL, charts, Python, and email. Verify every draft against a signed source before it leaves your desk.