Why a chatbot can't close your books or process a wire on its own — and the three innovations that turn a language model into something that actually acts.
Ask any large language model (LLM) "reconcile today's nostro breaks" and it will happily write you a paragraph describing what reconciliation is. It cannot actually pull the general ledger export, compare it to the bank statement feed, or flag the three transactions that don't match — because a plain LLM has no way to do anything. It only predicts the next word.
The model can call external functions — query a database, hit an API, read a file, run a calculation — and use the result to keep reasoning. This is what turns "describe the reconciliation" into "run the reconciliation."
Instead of jumping straight to an answer, the model breaks a problem into intermediate steps — check the GL balance, then the bank balance, then compute the delta, then classify the break type — improving accuracy on multi-step financial logic.
Modern models can read scanned invoices, screenshots of a trade ticket, or a PDF statement directly — not just typed text — which matters enormously in finance where source documents are rarely clean structured data.
Put those three together and you get a system that doesn't just answer once — it loops: read the input, reason about what to do, take an action (a tool call), observe the result, and decide the next step. That loop is what Module 2 formally names Perceive → Reason → Act.
| Capability | Plain Chatbot | Agent |
|---|---|---|
| Answer a policy question from a document | ✅ Yes | ✅ Yes |
| Pull today's actual GL balance from the source system | ❌ No | ✅ Yes (tool call) |
| Flag a break and draft a disposition memo | ❌ No | ✅ Yes (reasoning + action) |
| Read a scanned wire instruction and extract the beneficiary | ⚠️ Limited | ✅ Yes (multimodal + tool) |
In the offshore/marine version of this course, the go-to example is "a chatbot can't run your weld inspection workflow." For banking & finance, the equivalent is: a chatbot can't run your month-end close. It can explain the close calendar. It cannot pull the trial balance, flag the accounts that haven't been reconciled in two cycles, and draft the variance commentary — that requires tool use, multi-step reasoning, and (often) reading a scanned support document. That's the gap agents close.