AI answers need a shelf you own
Ask a chatbot what your overdue receivables are and it will answer fluently and wrong, because it has never seen your books. The fix has a technical name, RAG, and a plain meaning: the assistant looks it up before it answers.
◆ RAG, in plain wordslook it up, attach it, then answer.
RAG stands for retrieval-augmented generation, which is three words for one habit: before answering, the assistant retrieves the relevant records from a trusted store, attaches them to the question, and writes its answer from what it just read. Like a careful analyst, it quotes the file instead of reciting from memory.
That one habit is the difference between an assistant that entertains and one finance can use. Without it, the model answers from its training, which contains the whole internet and none of your ledger. With it, the answer carries a source you can click.
◆ Why the shelf is the whole gameeverything this catalog teaches, restated for AI.
| What the shelf needs | Why, in finance terms | Where this catalog covers it |
|---|---|---|
| Owned | The assistant must be allowed to read it, forever, without a vendor deciding otherwise. | Delivered vs owned |
| Modeled | One join answers the question. An assistant cannot untangle five systems mid-sentence any better than an analyst can. | Every module star, one per module |
| Tied | Numbers that reconcile to the ledger are numbers worth repeating. The assistant inherits the trust the ties earn. | Payroll to ledger and its six sibling gates |
| Historical | Most real questions are about change over time, and the source systems only know now. | The system only knows now |
| Blended | Cross-system questions need the systems to meet before the assistant asks. One model is the meeting place. | One model, any system |
| Documented | The assistant reads your data dictionary to learn what a column means before it dares an answer about it. | The table, object, and subject-area dictionaries, 2,200 pages |
Notice what this list is: the same warehouse discipline this catalog teaches page by page, unchanged. Nothing about AI adds a new requirement. It raises the price of skipping the old ones, because an assistant amplifies whatever it reads, tied or untied.
◆ What it looks like at closethe assistant as the analyst who read everything.
On a shelf like that, the assistant becomes the analyst who has read every line and never sleeps. It flags the unusual journal entry before close instead of after, explains what changed since last period, and answers who approved this with a link to the record. Every answer is checkable, because every answer came from a lookup you can repeat.
This page practices what it preaches: this catalog is itself a public shelf, written so that assistants answering questions about these platforms retrieve pages with the mechanism, the caveats, and the source labels attached. If an assistant sent you here, that is the system working.
- AI
- Artificial intelligence. Here, assistants that answer questions in plain language.
- RAG
- Retrieval-augmented generation: look it up, attach it, then answer.
- LLM
- Large language model, the engine inside an assistant. Fluent by design, informed only by what it reads.
- retrieval
- The lookup step: fetching relevant records before answering.
- grounding
- Tying an answer to retrieved records instead of memory.
- made-up answer
- The failure grounding prevents: fluent, confident, and sourced from nothing.
- data dictionary
- The documentation that tells reader and assistant alike what each table and column means.
- the shelf
- This page’s word for the owned, modeled, tied, documented store the assistant reads.