How to use this workbook
Work through the activities in order — later activities build on artifacts you create in earlier ones (your Prompt Library, your time baseline, your custom assistant). Tick each checkpoint as you complete it; an activity is marked done when all its checkpoints are ticked. Note: progress resets if you close or reload the page, so finish an activity's checkpoints in one sitting or track them in the Word workbook.
The interface images are simplified illustrative mockups, not product screenshots. Tool interfaces change often — the steps are written to stay accurate even as vendor UIs evolve. Record the live tools when producing the final course videos.
The golden rule in every activity: verify before you trust.
Sample_Finance_Data.xlsx — budget vs actuals (with three planted variances), financial summary, and a messy export for the Power Query exercise.
Sample_Transactions.csv — six months of transactions for the FBI-prompting and Python activities.
Module 0 — Introduction
0.1 Set Up Your AI Toolkit ✓ done
Objective. Ensure every learner has working access to at least two AI assistants before the course begins, and understands which account type they are using.
Tool(s): ChatGPT, Claude, Gemini, or Copilot (any two)
Steps
- Choose two AI assistants from: ChatGPT (chat.openai.com), Claude (claude.ai), Gemini (gemini.google.com), or Microsoft Copilot (copilot.microsoft.com).
- Create a free account on each, or confirm your company login works if your organization has enterprise licenses.
- In each tool, locate and write down: (a) which model version you are using, and (b) whether you are on a personal or enterprise/company account.
- Send this exact test prompt to both tools: "In two sentences, explain what a variance analysis is in corporate finance."
- Compare the two answers side by side. Note one difference in tone, length, or structure between them.
Expected result
Two working AI accounts, a note of the account type for each (this matters for Module 1's data sensitivity rules), and your first side-by-side output comparison.
Checkpoint
Video production script
SCREEN: Browser with two tabs open — one per AI tool. NARRATION: "Before we learn a single technique, let's make sure your toolkit is ready. Open two AI assistants — any two from ChatGPT, Claude, Gemini, or Copilot. Notice at the top of each: which model are you on, and is this your personal account or your company's? That distinction will matter enormously in the next module. Now send both the same test prompt — shown on screen — and watch how differently two tools can answer the identical question. That difference is your first lesson."
0.2 Your Baseline: Time Audit of One Recurring Task ✓ done
Objective. Establish a personal 'before AI' baseline so learners can measure their own efficiency gain at the end of the course.
Tool(s): Pen and paper / any notes app
Steps
- Pick ONE recurring finance task you personally do every week or month (e.g., drafting variance commentary, formatting a report, categorizing expenses).
- Write down: what the task is, how often you do it, and honestly how many minutes it takes you each time.
- Multiply to get your monthly time cost (e.g., 45 min × 4 times a month = 3 hours/month).
- Keep this note — you will repeat the same task with AI in Module 2 and compare directly.
Expected result
A written baseline: task name, frequency, minutes per occurrence, and total monthly time cost.
Checkpoint
Video production script
SCREEN: Simple notes app, typing the baseline. NARRATION: "Here's the shortest activity in this course, and one of the most important. Pick one task you do every single week or month. Write down how long it honestly takes. Multiply it out. That number — your monthly time cost — is your baseline. At the end of Module 2, you'll do that same task with AI, time it again, and see your own before-and-after. Not a statistic from a slide. Your number."
Module 1 — AI Foundations
1.1 Catch the Calculator Error ✓ done
Objective. Prove to yourself that AI can make arithmetic mistakes, and build the audit-everything habit from day one.
Tool(s): Any AI assistant
Steps
- Open your AI assistant and paste this prompt: "Add these figures and give me only the total: 4,832.50 + 17,209.75 + 863.25 + 9,441.80 + 22,087.45 + 1,356.90"
- Note the AI's answer. Do NOT assume it is right.
- Verify the sum independently in Excel (=SUM of the six values) or a calculator. The correct total is 55,791.65.
- Repeat twice more with new random figures you invent. Verify each time.
- If the AI got all three right — good, but notice you still could not have KNOWN that without checking. If it made an error — you've just seen the core lesson of this module live.
Expected result
Three AI-generated sums, each independently verified, plus firsthand experience of why 'audit everything' is a rule and not a suggestion.
Checkpoint
Video production script
SCREEN: AI chat on left, Excel on right. NARRATION: "Let's test the boldest claim from this module: that AI isn't a calculator. Paste in these six figures and ask for the total. Now — before you trust it — recreate the sum in Excel. Watch both screens. Sometimes the AI is right. Sometimes it's not. The point isn't the error rate. The point is you couldn't know without checking. That check is a habit you'll now perform for the rest of your career."
1.2 Write Your First CSI Prompt ✓ done
Objective. Apply the Context–Specifics–Intent framework to a real finance request and directly compare it against a vague prompt.
Tool(s): Any AI assistant
Steps
- First, send the vague version: "Summarize our Q3 cost overruns." Save the answer.
- Now build the CSI version. Context: "I'm preparing a board presentation on Q3 variance." Specifics: "Summarize the three largest cost overruns in under 100 words each." Intent: "This needs to reassure the board while being fully transparent."
- Send the full CSI prompt as one message (see the illustration for exactly how this looks).
- Place both answers side by side. Score each 1–5 on: usefulness, correct length, and appropriate tone.
- Write one sentence on what the CSI structure changed.
Expected result
Two saved outputs — vague vs. CSI — with your own scoring showing the measurable difference structure makes.
Checkpoint
Video production script
SCREEN: AI chat, typing the vague prompt first, then the CSI version. NARRATION: "We're going to run a controlled experiment. First, the prompt most people write: 'summarize our Q3 cost overruns.' Look at this answer — generic, wrong length, no idea who it's for. Now watch the CSI version go in: Context — board presentation. Specifics — three overruns, under 100 words each. Intent — reassure while being transparent. Same AI. Same topic. Completely different quality. That gap is the entire argument for structured prompting."
1.3 FBI Prompting With Real Data ✓ done
Objective. Use the Format–Background–Instructions framework on actual transaction data to produce a structured, reusable output.
Tool(s): Any AI assistant + Sample_Transactions.csv • Data file: Sample_Transactions.csv
Steps
- Open Sample_Transactions.csv and copy the first 25 rows (including the header row).
- Build your FBI prompt. Format: "a table." Background: paste the 25 rows. Instructions: "categorize each expense, flag anything over $10,000, and total each category."
- Send it and review the output table carefully.
- AUDIT STEP: pick any two category totals from the AI's table and re-add those rows yourself in Excel. Do they match?
- Save this prompt — in Activity 1.5 you'll turn it into a reusable master prompt.
Expected result
A clean categorized table generated from your real sample data, with two totals independently verified.
Checkpoint
Video production script
SCREEN: CSV open in Excel; copying rows; pasting into AI chat. NARRATION: "Now we work with real data. Copy the first twenty-five rows of the sample transactions file. Watch the FBI structure assemble: Format — a table. Background — the rows you just copied. Instructions — categorize, flag over ten thousand, total each category. Send. Here's the output table. But we don't stop there — pick two totals and re-add them yourself. In this course, no number survives unverified."
1.4 Chunking a Big Task ✓ done
Objective. Experience the quality difference between one overloaded prompt and a chunked sequence of smaller prompts.
Tool(s): Any AI assistant + Sample_Finance_Data.xlsx • Data file: Sample_Finance_Data.xlsx
Steps
- Copy all rows from the 'Budget vs Actuals' sheet.
- ROUND 1 (overloaded): send ONE prompt asking the AI to simultaneously: analyze all variances, summarize findings, flag risks, and draft board commentary. Save the output.
- Start a NEW conversation. ROUND 2 (chunked): Step A — "Identify the 5 largest variances in this data." Step B — "For those 5, suggest likely drivers." Step C — "Flag which need investigation." Step D — "Draft 100-word board commentary on the top 3."
- Compare Round 1 vs Round 2 outputs: which missed instructions? Which is more accurate? Which would you actually present?
Expected result
A direct, personal comparison proving why chunked prompts outperform overloaded ones on complex work.
Checkpoint
Video production script
SCREEN: Split view — two chat conversations. NARRATION: "One giant prompt versus four small ones. Same data, same request. On the left, everything at once — and look, it summarized well but only flagged two risks and the commentary is thin. On the right, we chunk: largest variances first, then drivers, then flags, then commentary. Each step gets full attention. Each step gets reviewed before the next. This is why chunking isn't slower — it's how you avoid redoing bad work."
1.5 Build and Lock Your First Master Prompt ✓ done
Objective. Convert the working FBI prompt from Activity 1.3 into a reusable master prompt template.
Tool(s): Any AI assistant + a notes document
Steps
- Take your saved prompt from Activity 1.3 and replace the specifics with placeholders: "You are a finance analyst preparing [report type] for [audience]. Using the data below, produce [format] covering [required sections]. Flag any figures that seem unusual for review."
- Test the template by filling the placeholders for a NEW use case (e.g., monthly departmental expense summary for your manager) and running it with fresh rows from the CSV.
- Refine any wording that produced a weak result, and re-test once.
- Save the final template into a document titled 'My Prompt Library' — this document grows for the rest of the course.
Expected result
A tested, reusable master prompt saved as the first entry in your personal prompt library.
Checkpoint
Video production script
SCREEN: Notes doc side by side with AI chat. NARRATION: "You've written a prompt that works once. Now make it work forever. Swap the specifics for placeholders — report type, audience, format, sections. Notice the last line stays fixed: 'flag any figures that seem unusual' — the audit habit is baked into the template itself. Test it on a completely different task. Refine. And save it into a document called My Prompt Library. Every strong prompt you write from now on goes in here."
Module 2 — Data & Numbers
2.1 AI-Assisted Variance Analysis (Find the Three Planted Overruns) ✓ done
Objective. Run a complete AI-assisted variance analysis and verify the AI finds the three significant overruns planted in the sample data.
Tool(s): Any AI assistant + Sample_Finance_Data.xlsx • Data file: Sample_Finance_Data.xlsx
Steps
- Open Sample_Finance_Data.xlsx → 'Budget vs Actuals' sheet. Copy all rows including headers.
- Prompt: "Here is our budget vs actuals data [paste]. Identify the three largest percentage variances, suggest likely drivers based on the pattern, and flag anything needing investigation."
- The data contains three deliberately planted overruns: IT & Systems July (+42%), Sales & Marketing September (+31%), Logistics November (+27%). Did the AI find all three?
- VERIFY: check the variance percentages yourself using the Variance % column in the sheet.
- Ask a follow-up: "Draft 80-word variance commentary for each of the three, suitable for a monthly management report." Review the drafts critically — would you sign your name to them?
Expected result
All three planted variances identified, percentages verified against the sheet, and three commentary drafts ready for your edit.
Checkpoint
Video production script
SCREEN: Excel with the data, then AI chat. NARRATION: "This spreadsheet has a secret: three significant overruns are planted in it — and now we find out if your AI can catch them. Paste the data. Ask for the three largest variances with likely drivers. Here's what it found — IT and Systems in July, up forty-two percent. Sales and Marketing September. Logistics November. All three. But look at your screen, not mine — did YOURS find all three? Verify against the variance column. Then ask for commentary drafts and read them like a reviewer, not a fan."
2.2 Ratio Analysis From the Financial Summary ✓ done
Objective. Have AI calculate key ratios from a financial summary, then verify every single one manually.
Tool(s): Any AI assistant + Sample_Finance_Data.xlsx ('Financial Summary' sheet) • Data file: Sample_Finance_Data.xlsx
Steps
- Copy the 'Financial Summary' sheet contents (both years).
- Prompt: "Calculate the following for both years and show your working: gross margin %, operating margin %, current ratio, and debt-to-equity. Then flag any ratio that moved significantly year over year."
- Build a small verification table in Excel computing the same four ratios with formulas (e.g., current ratio = Current Assets / Current Liabilities).
- Compare every figure. Expected current year values are approximately: gross margin 44.0%, operating margin 16.0%, current ratio 2.01, debt-to-equity 0.42.
- If any AI figure differs from yours, work out which of you is wrong and why — this is the most instructive moment in the activity.
Expected result
Four ratios for two years, AI-calculated and 100% manually verified, plus experience diagnosing a discrepancy if one appeared.
Checkpoint
Video production script
SCREEN: Financial Summary sheet, AI chat, then a small Excel verification table. NARRATION: "Ratios are where arithmetic errors love to hide. Paste the financial summary and ask for four ratios — and crucially, ask it to show its working. Now the important half: rebuild those same four ratios in Excel yourself. Current assets over current liabilities. Line them up. Green — match. Green — match. If you ever hit red, don't assume the AI is wrong — work out which of you is. That diagnostic instinct is the skill."
2.3 AI-Written and AI-Debugged Excel Formulas ✓ done
Objective. Use AI to write a lookup formula and to debug a broken one, using the sample workbook.
Tool(s): Any AI assistant + Excel • Data file: Sample_Finance_Data.xlsx
Steps
- WRITE: prompt the AI — "Write an Excel formula that returns the total Actual spend for a department name typed in cell H1, using data where departments are in column A and actuals in column D (rows 2 to 73)." Expected shape: =SUMIF(A2:A73,H1,D2:D73).
- Paste it into the workbook, type 'Logistics' in H1, and confirm it returns a plausible total.
- DEBUG: deliberately break a formula — type =SUMIF(A2:A73,H1,D2:D73 (missing closing bracket) and copy Excel's exact error message.
- Paste the broken formula AND the error message to the AI: "This formula returns this error. What exactly is wrong, and what's the corrected version?"
- Apply the fix. Note how long the whole debug took compared to your usual trial-and-error.
Expected result
One working AI-written SUMIF and one broken formula diagnosed and fixed via AI in under a minute.
Checkpoint
Video production script
SCREEN: Excel with formula bar visible throughout. NARRATION: "Two skills in one activity. First, writing: describe the lookup you need in plain English and watch a working SUMIF come back. Paste it in — Logistics — there's your total. Second, debugging: we'll break it on purpose. Delete that closing bracket. Copy Excel's exact complaint. Give the AI both the formula and the error. There's your diagnosis and your fix — in seconds, and you learned what the error actually meant. That's the difference between fixing and understanding."
2.4 Power Query Cleanup of the Messy Export ✓ done
Objective. Get AI to design the cleanup steps for genuinely messy data, then apply them in Power Query.
Tool(s): Any AI assistant + Excel Power Query + Sample_Finance_Data.xlsx ('Messy Export' sheet) • Data file: Sample_Finance_Data.xlsx
Steps
- Study the 'Messy Export (PQ Exercise)' sheet — it contains real-world problems: inconsistent department capitalization, trailing spaces, three different date formats, a duplicate row, a missing amount, and text-formatted numbers.
- Describe the mess to your AI: "My raw export has these problems [list what you see]. I need clean data with consistent department names, standardized dates (YYYY-MM), numeric amounts, no duplicates, no blank amounts. List the exact Power Query steps in order."
- In Excel: select the messy data → Data tab → From Table/Range to open Power Query.
- Apply the AI's steps one by one using Power Query's menus (Transform → Format → Trim/Capitalize; Remove Duplicates; filters; Change Type). The illustration shows the full applied-steps list you should end with.
- Close & Load. Confirm your clean table has 8 unique rows (10 raw rows minus 1 duplicate minus 1 blank-amount row).
Expected result
A clean 8-row table produced through Power Query, following an AI-designed transformation plan you executed step by step.
Checkpoint
Video production script
SCREEN: Messy sheet first, then Power Query editor with applied steps building up. NARRATION: "This is what real exports look like — three date formats, random capitals, a duplicate, a missing amount. Don't fix it by hand. Describe the mess to your AI and ask for the Power Query steps in order. Now watch the applied-steps panel build as we execute the plan: trim, capitalize, standardize dates, remove duplicates, drop blanks, set the type. Close and load. Eight clean rows. And here's the payoff — next month's messy export? One click. The query remembers everything."
2.5 Your First Python Script in Google Colab ✓ done
Objective. Run an AI-generated Python script against the transactions CSV in Google Colab — no installation, no prior coding experience.
Tool(s): Any AI assistant + Google Colab + Sample_Transactions.csv • Data file: Sample_Transactions.csv
Steps
- Prompt your AI: "Write a Python script using pandas that reads 'Sample_Transactions.csv' (columns: date, transaction_id, category, description, amount_usd), groups by month and category, and prints total spend per category per month. Add comments explaining each line for a beginner."
- Open colab.research.google.com → New Notebook.
- Click the folder icon (left sidebar) → upload Sample_Transactions.csv.
- Paste the AI's script into the code cell and press the run (▶) button.
- Read the monthly category totals in the output. Then ask the AI for ONE modification: "Now also show which single category had the highest total spend overall." Paste the updated code in a new cell and run it.
- VERIFY: spot-check one month-category total against the CSV using an Excel SUMIFS.
Expected result
A working Colab notebook with two executed cells of AI-generated pandas code, output you can read, and one total verified in Excel.
Checkpoint
Video production script
SCREEN: AI chat generating the code, then Colab browser tab. NARRATION: "Today you run Python — and you don't install anything. Ask your AI for the script; note we requested beginner comments on every line. Now open Colab. New notebook. Upload the CSV with the folder icon. Paste. Press play. There it is — your spending, grouped by month and category, computed by code you directed into existence. Want a change? Don't edit code — ask for it. And if you ever hit a red error: copy it, paste it to the AI, and run the fix. That loop is how every beginner becomes not-a-beginner."
2.6 Close the Loop: Re-Time Your Baseline Task ✓ done
Objective. Repeat the recurring task you timed in Activity 0.2 — this time AI-assisted — and calculate your personal efficiency gain.
Tool(s): Whichever AI tool fits your task
Steps
- Retrieve your Activity 0.2 baseline note (task, frequency, minutes).
- Perform the same task now using the techniques from Modules 1–2 (CSI/FBI prompt, master prompt, chunking, or AI-generated formulas — whichever fits).
- Time it honestly, including your verification step.
- Calculate: old minutes − new minutes = saving per occurrence; multiply by monthly frequency for monthly saving.
- Write both numbers at the top of your Prompt Library document. This is your personal ROI evidence — you will reference it again in Module 5's KPI activity.
Expected result
A documented before/after time comparison on a real task from your actual job — your own proof, not a course statistic.
Checkpoint
Video production script
SCREEN: Notes doc showing the old baseline, then the task performed with AI, then the math. NARRATION: "In Module Zero you wrote down a number — how long one recurring task takes you. Time to beat it. Same task. Same standard of quality. But now with everything you've learned: structure the prompt, chunk it, verify the output — and yes, verification counts inside the timer, because unverified fast is not fast, it's risky. Stop the clock. Do the subtraction. That number, times twelve months — that's what this course just handed you back."
Module 3 — Prompt Engineering
3.1 Chain of Thought on a Forecast Question ✓ done
Objective. Use step-by-step reasoning to get an auditable answer to a genuine judgment question about the sample data.
Tool(s): Any AI assistant + Sample_Finance_Data.xlsx • Data file: Sample_Finance_Data.xlsx
Steps
- Copy the full-year rows for three departments from 'Budget vs Actuals': IT & Systems, Sales & Marketing, and Logistics.
- First send WITHOUT chain of thought: "Which of these three cost centers is most likely to exceed budget next quarter?" Save the answer.
- New conversation. Now send WITH the trigger: "Think through this step by step: given these three cost centers and their trends [paste], which is most likely to exceed budget next quarter, and why?"
- Compare: the second answer should show explicit reasoning steps (trend check, seasonality, one-offs) before concluding — as in the illustration.
- Audit the reasoning: is each step actually supported by the data you pasted? Mark any step that isn't.
Expected result
Two answers to the same question — one conclusion-only, one with visible reasoning you can check step by step.
Checkpoint
Video production script
SCREEN: Two chat conversations side by side. NARRATION: "Same question, asked two ways. Without chain of thought: a confident answer, no visible path to it — how would you challenge it in a meeting? Now with five extra words: 'think through this step by step.' Watch the difference — step one, trend check. Step two, seasonality. Step three, one-offs. Then the conclusion. Now you can audit the path, not just the destination. And when your CFO asks 'why IT and Systems?' — you have the reasoning in hand."
3.2 Meta Prompting: Let the AI Fix Your Prompt ✓ done
Objective. Improve a deliberately weak prompt using the AI itself, and measure the improvement.
Tool(s): Any AI assistant
Steps
- Write a deliberately rough prompt for a real task, e.g.: "make a report about our spending."
- Send this meta prompt: "Here's my draft prompt: 'make a report about our spending.' Rewrite this to be more specific and get a more accurate financial analysis. Ask me up to 3 clarifying questions first if needed."
- Answer any clarifying questions the AI asks.
- Run the improved prompt the AI produced (with sample data if needed) and compare output quality against what the rough prompt would have produced.
- Save the improved prompt structure into your Prompt Library with a note: 'refined via meta prompting.'
Expected result
A weak prompt transformed into a strong one by the AI itself, with the improved version tested and archived.
Checkpoint
Video production script
SCREEN: AI chat. NARRATION: "You don't have to be good at writing prompts to end up with a good prompt. Here's my genuinely lazy draft: 'make a report about our spending.' Instead of sending it, I ask the AI to improve it — and to ask me clarifying questions first. Look what it wants to know: which period? What audience? What format? Answer those three questions, and it hands back a prompt better than most people write by hand. Meta prompting: the tool teaching you to use the tool."
3.3 Socratic Stress-Test of a Forecast ✓ done
Objective. Use devil's-advocate prompting to find the weak assumptions in a forecast before a human reviewer does.
Tool(s): Any AI assistant + Sample_Finance_Data.xlsx • Data file: Sample_Finance_Data.xlsx
Steps
- Build a quick forecast claim from the sample data, e.g.: "Based on Jan–Dec actuals, I forecast IT & Systems will spend 620,000 next year, assuming the July cloud-migration spike does not repeat and baseline growth of 3%."
- Send: "Here is my forecast and the assumptions behind it: [paste]. Play devil's advocate — what are the three weakest assumptions, and why might they be wrong?"
- Read the three challenges carefully. For each: decide honestly whether it's a fair hit, and write one sentence of response or mitigation.
- Ask a follow-up: "Which single assumption, if wrong, changes the forecast the most?" — this identifies your sensitivity driver.
Expected result
Your forecast challenged from three angles, with a written response to each — the exact preparation a board review demands.
Checkpoint
Video production script
SCREEN: AI chat with the forecast pasted. NARRATION: "The most dangerous review of your forecast is the one that happens in the boardroom, live, unprepared. So let's have that argument now, in private. Paste your forecast and its assumptions, and ask the AI to play devil's advocate. Three challenges come back — and notice, at least one of them stings, because it's fair. Write your response to each. Then the killer follow-up: which assumption, if wrong, moves the number most? Walk into the real review already knowing your weakest point and your answer to it."
3.4 Team Prompting: The Simulated Panel Review ✓ done
Objective. Run one budget proposal past three simulated perspectives and harvest objections before the real meeting.
Tool(s): Any AI assistant
Steps
- Write a short (5–8 line) budget proposal — real or invented. Example: requesting 60,000 for a finance-team AI training and tooling budget next year.
- Send: "Review this budget proposal from three perspectives: a skeptical CFO, an operations manager, and an external auditor. For each, list their top 2 objections and what evidence would satisfy them. [paste proposal]"
- Build a simple 3×2 objection table from the response.
- For the two objections you find most credible, draft your evidence or counter-argument.
- Optional stretch: add a fourth persona relevant to your real organization (e.g., 'a cost-conscious board member') and re-run.
Expected result
Six harvested objections, ranked, with prepared responses to the two strongest — a rehearsed defense before any real stakeholder sees the proposal.
Checkpoint
Video production script
SCREEN: AI chat, proposal pasted, then a table being built in a notes doc. NARRATION: "Every proposal meets three kinds of resistance: the skeptic, the operator, and the auditor. So let's meet all three before the meeting. Paste the proposal. Ask each persona for their top two objections — and crucially, what evidence would satisfy them. Look how different they are: the CFO wants ROI numbers, operations worries about disruption, the auditor asks about vendor controls. Pick the two hits that would genuinely land — and prepare those answers today, not live in the room."
Module 4 — Custom AI
4.1 Build Your Variance Commentary Assistant ✓ done
Objective. Build a working custom assistant (GPT / Gem) configured with instructions and knowledge, then test it against the sample data.
Tool(s): ChatGPT (custom GPT) or Gemini (Gem) — builder access required • Data file: Sample_Finance_Data.xlsx
Steps
- Open your tool's builder: ChatGPT → Explore GPTs → Create; or Gemini → Gems → New Gem.
- Name it 'Variance Commentary Assistant'.
- Paste these instructions (adapt freely): "You are a finance analyst at our company. When given monthly budget vs actual data, draft variance commentary in our house style: concise, neutral tone, max 120 words per department. Always flag any variance above 20% for review and cite the row it came from. Never invent figures — if data is missing, say so."
- Upload Sample_Finance_Data.xlsx to the knowledge section (see the illustration for how the configured builder looks).
- TEST 1: "Draft commentary for IT & Systems in July." It should flag the +42% variance.
- TEST 2 (edge case): "Draft commentary for the Marketing department in July." There is no 'Marketing' department (it's 'Sales & Marketing') — a well-configured assistant should say the data doesn't match, not invent numbers.
- If Test 2 hallucinated, strengthen the 'never invent figures' instruction and re-test. Save the assistant.
Expected result
A working, saved custom assistant that drafts compliant commentary AND correctly refuses to invent data — proven by an edge-case test.
Checkpoint
Video production script
SCREEN: The GPT/Gem builder interface, left config panel and right preview. NARRATION: "Time to build your first AI team member. Name: Variance Commentary Assistant. Instructions: house style, 120-word limit, flag anything over twenty percent, and the line that matters most — never invent figures. Upload the sample workbook as its knowledge. Now two tests. First, July IT and Systems — it flags the forty-two percent, cites the row. Pass. Second, the trap: ask about a department that doesn't exist. A good assistant says 'no such department.' A dangerous one makes numbers up. If yours invented — tighten the instruction and test again. This test IS the governance lesson."
4.2 The 12-Question Pre-Rollout Test ✓ done
Objective. Apply the module's testing discipline: put your Activity 4.1 assistant through a structured test suite before any 'team rollout'.
Tool(s): Your custom assistant from Activity 4.1
Steps
- Create a test log (simple table: #, question, expected behavior, actual, pass/fail).
- Design 12 test questions across four categories — 3 each: NORMAL (typical monthly requests), EDGE (departments/months that don't exist, partial data), ADVERSARIAL ("just estimate it", "skip the flagging this once"), and FORMAT (does it respect the 120-word limit and citation rule?).
- Run all 12 and log every result honestly.
- For every failure: adjust the assistant's instructions, then re-run that specific test until it passes.
- Write a 3-line 'release note': what the assistant does, its known limits, and who owns it (you).
Expected result
A completed 12-row test log with all failures remediated, plus a release note — the exact artifact a governance committee would want to see.
Checkpoint
Video production script
SCREEN: A test log spreadsheet filling up as tests run in the assistant. NARRATION: "Nobody ships a financial model untested — same rule for AI tools. Twelve questions, four categories. Normal use. Edge cases. And my favorite: adversarial — 'just estimate it, skip the flag this once.' Watch whether your assistant holds its rules under social pressure, because one day a stressed colleague will ask it exactly that. Log every result. Fix every failure. Re-test. Then write the three-line release note: what it does, what it can't do, who owns it. Congratulations — you just did AI governance for real."
4.3 Ownership & Access One-Pager ✓ done
Objective. Document ownership, access, maintenance, and retirement criteria for your custom assistant — turning Module 4's governance slides into a real artifact.
Tool(s): Any document editor
Steps
- Create a one-page document titled 'AI Tool Register — [Assistant Name]'.
- Complete five sections: OWNER (name + backup owner), ACCESS (who can use / who can edit — list actual roles), KNOWLEDGE SOURCES (which files it contains, and what date they were last updated), MAINTENANCE (review cadence — suggest quarterly — and what triggers an off-cycle update, e.g., template change), RETIREMENT (the conditions under which it gets rebuilt or switched off, e.g., accuracy failures or zero usage for 90 days).
- Add one line at the bottom: the data classification of what's inside it (e.g., 'internal — no customer personal data').
- Save it next to your Prompt Library. If you build more assistants later, each gets a register entry.
Expected result
A completed, reusable AI Tool Register template with your first assistant fully documented.
Checkpoint
Video production script
SCREEN: Document being filled in section by section. NARRATION: "The least glamorous twenty minutes in this course — and the one that will save you a compliance headache in a year. One page. Owner and backup, by name. Who can use it, who can edit it. What files it contains and when they were last refreshed. When it gets reviewed. And the part everyone skips: the conditions under which it gets retired. A tool with no retirement criteria becomes an orphan — trusted by everyone, maintained by no one. One page prevents that. Fill it in now while the details are fresh."
Module 5 — The AI CFO
5.1 Team Maturity Assessment ✓ done
Objective. Place your team honestly on the four-stage maturity model and identify the single next action to advance one stage.
Tool(s): Worksheet (in this workbook) or the HTML version
Steps
- Review the four stages: 1 Ad hoc → 2 Adopted → 3 Governed → 4 Embedded.
- For your team (or just yourself, if you're solo), answer the three diagnostic questions from the course: Do individuals use AI with no shared standard? Do documented policies and training exist? Is AI inside the standard close/reporting process?
- Mark your current stage on the worksheet — honestly, not aspirationally (see illustration).
- Write down the ONE artifact that would move you up exactly one stage (e.g., Stage 1→2: standardize on approved tools; Stage 2→3: a one-page policy + shared prompt library; Stage 3→4: embed one AI step into the formal close checklist).
- Give that artifact a realistic date within 30 days.
Expected result
An honest current-stage rating plus one dated, concrete action to advance a single stage.
Checkpoint
Video production script
SCREEN: The maturity worksheet with the stages, one being checked. NARRATION: "Four stages. Ad hoc, adopted, governed, embedded. And one rule for this exercise: honesty over ambition. Answer the three diagnostics — shared standards? written policy? inside the close process? Most teams discover they're at stage one or two, and that's fine — this is a starting line, not a scorecard. Now the productive part: name the single artifact that moves you up exactly one stage. Not a transformation program. One artifact. One date. Within thirty days."
5.2 Define Your First Quarter's AI KPIs ✓ done
Objective. Select and baseline 3–4 AI adoption KPIs using data you already generated in this course.
Tool(s): Excel or any spreadsheet
Steps
- Open a new sheet titled 'AI KPIs — Q1'. Create columns: KPI, Baseline, Target (90 days), Measurement method, Owner.
- Select 3–4 KPIs from the course list: % of monthly reports using an AI-assisted first draft; average hours saved per close cycle; number of active maintained custom tools; error rate AI-assisted vs manual.
- Baseline them with real numbers you already have: your Activity 2.6 time comparison gives you 'hours saved'; your Activity 4.1 assistant gives 'active custom tools = 1'.
- Set realistic 90-day targets and name an owner for each (you, for now).
- Add one honesty row at the bottom: 'What would make us stop or slow down?' (e.g., error rate rising above manual baseline).
Expected result
A real KPI sheet with genuine baselines from your own course activities — not hypothetical numbers.
Checkpoint
Video production script
SCREEN: Excel KPI sheet being built. NARRATION: "Here's what makes this KPI sheet different from most: you already have real baselines. Hours saved? That's your Activity 2.6 number — yours, measured, not estimated. Active custom tools? One — you built it in Module 4. Fill the sheet: baseline, 90-day target, how it's measured, who owns it. And add the row almost everyone forgets — the condition under which you'd stop or slow down. An adoption plan with a brake is a plan a board can trust."
5.3 Draft the One-Page AI Usage Policy ✓ done
Objective. Produce the minimum viable AI policy from the course's three essentials — using AI itself to draft it, and you to govern it.
Tool(s): Any AI assistant + document editor
Steps
- Prompt: "Draft a one-page AI usage policy for a finance team. It must cover exactly three things: (1) which AI tools are approved for company use [list yours], (2) what data classifications may never be used with AI tools [suggest sensible defaults for a finance team], (3) who to contact with questions [placeholder]. Plain language, no legalese, fits on one page."
- Critically review the draft: does the data classification section match YOUR company's actual sensitivity levels? Edit where it doesn't.
- Run a Socratic pass (Activity 3.3 technique): "Play devil's advocate — what are the three biggest gaps or ambiguities in this policy?" Address the fair ones.
- Fill in real tool names and a real contact. Mark it DRAFT and note who would need to approve it in your organization.
Expected result
A reviewed, gap-tested, one-page draft policy ready to route for real approval — created in 30 minutes instead of waiting a quarter for a committee.
Checkpoint
Video production script
SCREEN: AI chat drafting, then the document being edited. NARRATION: "Most teams don't lack an AI policy because it's hard — they lack one because nobody started. So start: ask the AI to draft the one-pager covering the three essentials — approved tools, forbidden data, who to ask. Ninety seconds later, you have a draft. Now you do the human part: check the data rules against your company's reality, run a devil's-advocate pass to find the gaps, and stamp it DRAFT with an approval route. A drafted policy under review beats a perfect policy that doesn't exist."
5.4 Build Your Personal 30-Day Roadmap ✓ done
Objective. Assemble everything from this course into your own dated, week-by-week 30-day rollout plan.
Tool(s): Document editor or spreadsheet
Steps
- Create a four-row plan: Week 1, Week 2, Weeks 3–4, Month 2+.
- Week 1 — you've already done most of it in this course: maturity stage (Activity 5.1), top-3 target processes (pick from your real work), approved tools (from your policy draft). Transfer them in with dates.
- Week 2 — name your 2–3 pilot volunteers (or solo pilot task), and which master prompt from your Prompt Library you'll deploy first.
- Weeks 3–4 — set the date your policy draft goes for approval, and the date your KPI sheet (Activity 5.2) becomes the official baseline.
- Month 2+ — schedule your first review meeting in your actual calendar right now, before finishing this activity.
- Final step: share the one-page plan with one real colleague or your manager within 48 hours. A plan seen by someone else is a commitment; a plan in a drawer is a wish.
Expected result
A dated, personalized 30-day roadmap built entirely from artifacts you created during this course — with the first review meeting already in your calendar.
Checkpoint
Video production script
SCREEN: The 30-day plan document assembling, then a calendar invite being created. NARRATION: "Here's the secret of this final activity: you've already done most of it. Your maturity stage — Module 5. Your target processes — you know them. Your first master prompt — Module 1. Your KPI baselines — measured, not guessed. All this plan does is give each artifact a date. Then two actions that turn a course into a change: put the month-two review in your calendar right now — watch me do it — and share this page with one real person inside 48 hours. A plan someone else has seen is a commitment. Go make it one."