How this track works
Ten sessions, all hands-on, all on real finance work. Tonight you learn what the tool actually is; then b2 teaches prompting for analysis, b3 builds the flagship month-end close copilot with a live lever simulator, and b4-b10 add forecasting, board-ready narrative, Excel, document Q&A, risk, markets, and a safe path to production. You do not need to code or know machine learning - if you can write a clear email and read a P&L, you can do every session. Two rails run through all ten, starting now.
What an LLM actually is - for a finance brain 6 min live
A large language model is a very good next-word predictor trained on a huge slice of text. That one fact explains everything useful and everything dangerous about it in finance. It is a language engine, not a calculator - so it is brilliant at drafting, reasoning, and structuring, and untrustworthy at silent arithmetic. The whole skill is knowing which is which.
LiveWhy it fabricates numbers - and why that is not a bug you can prompt away3 min▶
The model generates the most likely-sounding continuation, not the true one. Ask "what was Q3 revenue?" with no data attached and it will happily produce a confident, well-formatted, completely invented figure - because a plausible number is a likely continuation. This is called a hallucination, and in finance it is the whole ballgame.
- It is probabilistic, not factual. No database lookup is happening unless you connect one. Confidence in the output tells you nothing about accuracy.
- Arithmetic is text prediction too. "127.4 + 89.1 =" is answered by pattern, not a calculator. It is often right on simple sums and quietly wrong on messy ones. Never let it do the math unseen.
- The fix is grounding, not scolding. "Please be accurate" does nothing. Giving it the actual numbers, and asking it to reason over them and show its steps, does almost everything.
The board pack that cited a number that never existed. An analyst asked a chatbot to "summarize our Q2 performance" and pasted only the narrative, not the numbers. The model filled the gaps with confident, invented figures that flowed straight into a draft board slide. Caught in review - barely. The lesson that opens this course: an ungrounded finance answer is fiction with good formatting.
Self-studyThe can / cannot table - pin this to your monitor2 min read▶
| Trust it to... | Never trust it to (without a tool + check) |
|---|---|
| Draft variance commentary from numbers you supply | Recall a specific figure from memory |
| Explain a concept, standard, or ratio | Do multi-step arithmetic silently |
| Restructure, summarize, and translate text | Cite a source it was not given |
| Extract fields from a document you attach | Know anything after its training cutoff |
| Draft and explain a formula | Be your system of record |
Everything in the left column shows up as a build-along later in the track. Everything on the right is why sessions b3, b6, and b10 lean on tools, spreadsheets, and verification instead of the model's own head.
The finance AI tool landscape, 2026 5 min live
"Should I use ChatGPT or Copilot?" is the wrong first question. The right one is: is this tool safe for the data I am about to put in it? There are three layers, and the line between them is a privacy line, not a quality line.
LiveConsumer vs enterprise - the setting that matters most3 min▶
The capability gap between a free chatbot and its enterprise tier is small. The data-governance gap is enormous, and it is the only thing your risk team cares about.
- Consumer tiers may retain conversations and, depending on settings, use them to improve models. Fine for public filings, concepts, and synthetic data - never for real financials.
- Enterprise tiers (ChatGPT Enterprise, Claude Team/Enterprise, Copilot for M365) contractually do not train on your data, add admin controls, SSO, and retention settings. This is the tier your firm should license before anyone touches real numbers.
- Embedded copilots live inside Excel, your ERP, or an FP&A tool like Datarails - convenient and governed by that vendor's terms. Read them the same way.
Self-studyThe 2026 landscape at a glance2 min read▶
- General assistants: ChatGPT, Claude, Gemini, Microsoft Copilot - the workhorses for drafting, analysis-over-supplied-data, and Q&A.
- In-suite: Copilot in Excel/Word/Teams, Gemini in Google Workspace - AI where your work already lives (session b6).
- FP&A + finance-native: Datarails, Vena, Planful and ERP vendors are shipping copilots for forecasting, close, and reporting.
- The market signal: 59% of finance functions already use some AI, and nearly 4 in 5 CFOs plan to deploy generative AI within 24 months. The question at your firm is no longer "if" but "governed how".
The two rails: privacy, and assist-not-decide 4 min live
Everything else in this course is technique. These two rails are non-negotiable, and they are what separates a finance professional using AI well from a headline about a finance professional who used it badly.
LiveRail 1 - the data-privacy line2 min▶
Never paste material non-public information (MNPI), client PII, or unpublished financials into a consumer AI tool. This is not caution for its own sake - it is securities law, data-protection law (PDPA, GDPR), and your firm's confidentiality obligations, all at once. Use the enterprise or local path for anything real.
Samsung's ban. Engineers pasted confidential source code into a consumer chatbot to debug it; the company banned generative AI on internal machines within weeks. The finance equivalent - pasting a pre-release earnings table to "help write the commentary" - is worse, because it can also be insider-information mishandling. Rail one exists so you never make that headline.
LiveRail 2 - AI assists, the licensed human decides2 min▶
AI drafts, suggests, and flags. A qualified human reviews and owns every number, filing, credit decision, and piece of advice that leaves the building. Nothing in this course is investment, tax, accounting, or legal advice, and no workflow here makes an automated decision about a person's money without a human in the loop.
- Accountability does not transfer to the model. "The AI said so" is not a defence to an auditor, a regulator, or a client.
- Regulators are watching the claim, too. The SEC has penalised "AI-washing" - overstating what your AI does. Be as honest externally as you are internally.
- Design for it. Every workflow you build gets a review step. In b3 that is a human sign-off on the variance narrative; in b8 it is a human decision on every flagged item.
Ungrounded vs grounded - see the fabrication die ★ 12 min · everyone builds
The single most important habit in the course, in one experiment. Ask the same analytical question twice - once from the model's memory, once with the actual numbers attached - and watch a confident invention become a checkable answer.
Open any chat AI. Pick a small, non-confidential number set: a public company's headline figures, or make up a tiny 6-line P&L. Keep it safe for a consumer tool (rail one).
Ungrounded ask: without pasting any data, ask it to "explain the main driver of the change in operating margin this quarter, with the numbers". Note how specific - and how invented - the figures are.
Grounded ask: paste your actual number set, then ask the same question - and add "use only the numbers I gave you, show the subtraction, and say 'not in the data' if something is missing".
Compare. The grounded answer ties to your figures, shows its arithmetic so you can check it, and admits gaps. That delta - fiction to checkable - is the value of everything that follows.
Now stress it: remove one number it needs and re-run the grounded prompt. A good prompt makes it say "not in the data" instead of inventing. That is the behaviour you will engineer on purpose in b2.
Write your team's one-page AI ground rules ★ 10 min · build your own
A reusable artifact you will actually keep: a short, concrete "how our finance team uses AI" card. You draft it with the AI, then you edit it - which is itself rail two in miniature.
Give the AI the two rails and your context: team, tools you have licensed (or not), and the data types you handle (management accounts, client data, MNPI).
Ask for a one-page card with three sections: Green light (safe uses + tools), Red light (never do this), and Always (verify math, human sign-off, cite the source data).
Edit it hard. Cut anything vague, add your firm's actual tool names and approvers. The AI draft is 70%; your judgement is the 30% that makes it real - and the 30% that matters.
Save it where your team will see it. You will refine it again in the leader track's session a3 (governance) and the builder finale b10 (production).
The card that unblocked a team. A controller's group was frozen - half afraid of AI, half already pasting into free chatbots. A one-page card like this, agreed in a 30-minute meeting, replaced fear and free-for-all with a shared green light. Adoption went up and risk went down, because "what's allowed" stopped being a guess.
Try it yourself - this week ◐ 30-45 min total
- Finish both demos if you did not complete them live - especially the grounded-vs-ungrounded experiment. Feeling the fabrication yourself is worth more than being told about it.
- Run the ungrounded prompt on three questions you actually get asked at work. Screenshot the invented specifics - that is your reminder of why grounding matters.
- Finalise your one-page AI ground rules and share it with one colleague for a sanity check.
- Confirm which tool tier you are cleared to use for real data. If the answer is "none yet", that is a great thing to raise - and the leader track's a5 gives your manager the business case.
- Optional: skim your firm's existing AI or data-handling policy. Bring one question it does not answer to session b2.
Official sources covered
This track teaches from public finance-AI curricula and vendor documentation. Certificates, graded assessments, and paid tooling stay with the official programs - all noted honestly. This session covers:
Three questions before you go 🎯 ◐ 90 seconds
1 · You ask a chatbot "what was our Q3 revenue?" with no data attached. It gives a precise figure. What do you know?
An LLM predicts likely text, not true facts. With no data attached, a precise figure is a hallucination with good formatting. Ground it, then verify.
2 · You need to draft variance commentary from your real, pre-release management accounts. Which tool is acceptable?
Rail one. Real, non-public financials go only into enterprise or local tools. The capability is similar across tiers; the data-governance is not.
3 · What is the reliable fix for an AI that fabricates or fumbles numbers?
Scolding does nothing; grounding does almost everything. Supply the numbers, make the steps visible, and keep a human check. That is the whole track in one line.