AI Foundations for Finance
Devon Coombs, CPA, MBA
Practitioner first, professor second
I advise finance leaders on putting AI to work without losing control of the numbers, and I teach it. The material here is what I use with clients and in the classroom, not a keynote. Where a claim is illustrative, I say so.
- The AI Adoption Method · Map it, Split it, Fuel it, Guard it, Review it
- The AI-Augmented SOX Framework · 50 steps, team plus AI
- The Integration Blueprint · owner-side ERP planning
- The Deals Desk · pre-signature technical accounting
Five moves for putting AI to work in finance
Everything after this hangs on these five. They recur throughout the deck, not in order, tagged in the kicker as each one comes up.
Find the workflow
Name the real task and where the judgment actually lives.
AI vs deterministic
Decide which steps are best-guess and which must be exact.
Give it context
Curate the files, policies, and data the model can see.
Design the controls
Red lines, data boundaries, disclosure, and evidence.
Own the sign-off
Every judgment call stays with a person. That is the product.
The new superpowers in the age of AI
Curiosity
Try it. Poke it. Ask it why.
Resilience
The first answer is a draft. Iterate.
Belief
Assume it can, then check.
What is a large language model?
Everything it learned, plus what you give it, then it predicts the most likely next word.
trained on all of it
How do you interact?
Everything in and out is tokens, and it all happens inside one working memory.
And it does not stop at one. That was a single context window. An agent can also run several at once, many readers on many contracts in parallel, which is where the shift to agents in Part 2 begins.
Feel the tokens
(about 2,700 pages)
How we talk to AI has evolved
Think of the last time you asked a colleague for something and got back, well, not that. You probably left out context. Same here, so own what you did not say.
Prompting
Type what you want in plain language and hope the model fills the gaps.
Tune · your wording
Prompt Engineering
Shape the single ask: clear instructions, worked examples, step-by-step reasoning.
Tune · examples and format
Context Engineering
Curate everything the model sees: retrieved docs, memory, history, not just the prompt. Iterative, every time.
Tune · retrieval and memory
Harness Engineering
Design the workflow the model runs in: tools, files, permissions, checkpoints. This is where internal controls live.
Tune · tools and guardrails
Prompt engineering is just communicating well
Who is the AI for this task?
"You are an expert CPA explaining this to a client."
The situation, plus what matters.
"Our external auditors are questioning this revenue treatment."
Who is reading the output?
"Write for senior finance leaders at a SaaS company."
The shape you want back.
"Include a table summarizing the key points."
Context engineering: from instructions to goals
Your job changes. You stop wordsmithing the perfect instruction and start setting a goal and curating what the model can see: files, policies, history, memory.
It is iterative. The curation happens every time you decide what to hand over.
→ And that is exactly what sets us up for agents.
One polished turn. You perfect the sentence, the model answers, you read it. Great for a single question.
A goal, plus everything the model should see. You curate the inputs; it works across them, again and again.
The models you use today are likely the worst you will ever use
Two failure modes to respect
Signal the answer you want and you will get it, accuracy optional. Do not lead the witness. Ask it to argue the other side.
From chat to agents
You are the glue
- Every hop routes through you: fetch, paste, re-explain, copy out, fix
- Data changed? You start over
You delegate
- One instruction, and it opens the files and systems itself, showing its work
- It brings you the judgment calls
🎮 Close the Books
Round 1 complete 😮💨
The rec got done, but you made 12 clicks in 0:00, and every step routed through you. You were the clipboard, the context, and the quality control.
Now do the exact same task the 2026 way.
INV-2041, $12,400 in the subledger vs $12,040 in the GL. Difference $360. The pattern suggests a keying transposition at posting. A draft adjusting JE is ready for your review.
3 invoices shipped Aug 30 to 31 posted in September. This looks like a cutoff issue. Proposed treatment: flag for accrual and note in the close memo.
Scoreboard, same task, two ways
Round 1 · Chat
Round 2 · Agent
The work did not disappear, it moved. The agent prepares, you review. That is the whole shift, and it is Move 5, Review it.
The same move, on a real contract
├─ 📁 communications/ · emails
├─ 📁 accounting-policy/ · rev-rec
├─ 📁 contracts/
│ └─ 📄 New_MSA_draft.docx
└─ 📁 outputs/ · empty, watch this
Same shift, every vendor
So many new names. Here is the map. Different badges, same move: a chat window becomes an agent with its hands on your files.
| Vendor | Chat tier, you may already have it | Agent tier, what to ask about |
|---|---|---|
| Anthropic | Claude (web app) | Claude Cowork · Claude Code |
| OpenAI | ChatGPT | Codex · ChatGPT desktop agent |
| Gemini | Antigravity · Gemini CLI | |
| Microsoft | 365 Copilot | Copilot agents · agent mode in Excel |
Skills: teach it your way, once
A skill is a recipe card the agent reads before doing a task, your SOP, but executable. Here is the mechanics:
"Format these JEs for NetSuite."
Just ask normally, no magic words needed.
Every skill's name plus one-line description is always in view:
name: netsuite-je-formatter
description: Use when JEs need formatting for NetSuite import
---
1. Map to the 14 columns
2. Debits equal credits, else stop and ask
3. CSV, one row per line item
Same recipe, every single time, for you and anyone you share it with.
🧭 Skills, hands on
Where are skills?
In the chat app: Settings, then Capabilities, then Skills. The pre-made ones for Excel, Word, PowerPoint, and PDF are already there and on.
In a coding agent: a folder on disk. Type /skills to see what is loaded.
· netsuite-je-formatter · close-checklist · flux-commentary
You made one, where did it go?
Into a folder. One folder per skill, one SKILL.md file inside, and that file is the skill. No app, no deploy, no IT ticket.
It is just text, which is exactly why finance teams can own these.
├─ 📁 close-checklist/ → SKILL.md
├─ 📁 flux-commentary/ → SKILL.md
├─ 📁 netsuite-je-formatter/
│ └─ 📄 SKILL.md, the whole thing
└─ 📁 contract-review/ → SKILL.md
How do I call it, over and over?
Just ask naturally. The description is the trigger, so "format these JEs for NetSuite" fires it every time.
Want to force it? Call it by name: /netsuite-je-formatter. Either way, same recipe, every run. That is the point, consistency you can control.
How do I share it with my team?
Share the folder. Zip it, drop it in a shared drive, or keep it in a team repo. A teammate adds it to their skills folder, done.
Now the whole team's JEs come out your way. The skill file is the SOP and the training in one.
Keeping them organized and versioned
Skills folders get messy the same way SOP binders do. The fix is boring and it works:
- One job per skill. An "accounting-helper" skill is a junk drawer
- The description says WHEN. Vague descriptions never fire, the top reason skills feel broken
- Date and version line inside the file. "v3 · Aug 2026 · owner: controller team"
- Prune what never fires. Review quarterly, like SOPs
Your turn: skill it, or just ask?
"Format these JEs for NetSuite import, same 14 columns, every month."
Rule of thumb: recurring, plus your way, plus has steps, makes a skill. A one-off question, just ask.
Probabilistic vs deterministic
You have watched agents run whole workflows. Now split the workflow into its two kinds of steps:
Best guess from experience
- Draws on everything it learned to predict what comes next, just like you use past experience to judge a new contract
- Brilliant with ambiguity and judgment
- Rarely the same answer twice
- Sometimes wrong, so the output gets reviewed
Same input, same output
- Excel formulas, RPA bots, code, calculators
- Every run identical, forever
- If it is wrong, it is wrong every time, so find it once and fix it forever
- Easier to test, control, and audit
Where this goes next in a finance shop
Everything so far becomes concrete in three published frameworks. Each is free to read in full, with the interactive version and the economics.
Revenue, before signature
A pre-signature technical accounting function that shapes contract terms before they are signed, the agent demo you just saw, made a standing practice.
Controls at platform scale
Fifty SOX 404 steps split across your team and AI, benchmarked against a GRC platform, with a live economics model. Split it and Review it, at scale.
Owner-side systems change
Plan an ERP or systems integration on your terms: scope the work, control total cost, and choose a partner on comparable terms. A live scorecard and cost calculator.
The one page that keeps AI work defensible
Before any of this touches a workpaper, agree on three things. This is the governance one-pager I teach, and it fits on a single page on purpose.
No final sign-off, no severity or scoping call, no client-facing filing without a human owner. The tool drafts, a person concludes.
Public tools for public data only. Enterprise terms, no training on your inputs, for anything sensitive. A documented boundary, reviewed quarterly.
Sources real, math ties, logic sound, scope complete, disclosure attached. The reviewer signs, and the AI-assist is disclosed on the deliverable.
The next small step is the whole thing
Want to go deeper?
The AI Adoption Method, the Deals Desk, the AI-Augmented SOX Framework, the Integration Blueprint, and the AI Readiness self-assessment. All published, all free to pressure-test.
Applied AI Finance, plus Corporate, Real Estate, and Entrepreneurial Finance. Interactive lessons, the prompt-pattern library, and the governance one-pager.
AI adoption and enablement, complex transactions, SOX and controls, and embedded teams that carry the work after the engagement ends.
Go play. Then come find the frameworks.
Which model? Pick the role, then the house.
| Anthropic | OpenAI | Microsoft | ||
|---|---|---|---|---|
| 🏛 Architects 💰 top tier | Opus 5 · Fable 5 | GPT-5.6 Sol (Pro) | Gemini 3.1 Pro | Copilot plus frontier models |
| ⚙️ Executors ⚡ fast/cheap | Sonnet 5 · Haiku 4.5 | GPT-5.6 Terra/Luna · Codex | Gemini 3.6 Flash | Copilot agents |
| 🔬 Researchers | Claude deep research | Deep Research (o-series) | Gemini Deep Research | Copilot Researcher |
| ✍️ Writers | Claude, drafting and memos | GPT-5.6 in ChatGPT | Gemini in Docs | Copilot in Word |
| 📊 Excel-native | Claude in Excel | — | Gemini in Sheets* | Copilot in Excel (agent mode) |
| 📽 PowerPoint-native | Cowork plus pptx skill | — | Gemini in Slides* | Copilot in PowerPoint |