AI Foundations for Finance
Devon Coombs, CPA, MBA
In the work, and in the classroom
My day job is helping finance teams adopt AI without loosening their grip on the numbers, and I teach the same material to students. So what follows is field-tested, not a keynote. When a figure is a rough estimate, I will call it one.
- The AI Adoption Method · the backbone of this deck: Map, Split, Fuel, Guard, Review
- The Deals Desk · runs live · technical accounting before the ink dries
- The AI-Augmented SOX Framework · published in full · controls at scale
- The Integration Blueprint · published in full · ERP change on the owner's terms
- The AI Readiness model · a read on where your team really is
Five moves for putting AI to work in finance
These five carry the rest of the hour. They come back over and over, out of order, flagged in the kicker as each one arrives.
Name the workflow
Write down the actual task, and mark where a human has to judge.
Judgment vs machinery
Sort the steps into best-guess work and steps that must be exact.
Feed it the right inputs
Hand it the files, policies, and history that the task actually needs.
Set the controls
Draw the red lines, the data boundary, the disclosure, the evidence trail.
Keep the sign-off
A person makes every call that carries a name. That is the whole point.
Before the vocabulary, the posture
Treat it as a lever
A clear ask can turn an afternoon of grunt work into minutes. Aim it, do not just pull it.
Draft, then push
The first reply is a starting point. The value is in the second and third ask.
Stay the skeptic
Assume it can help, then verify like the CPA you are.
What a large language model actually does
It has read an enormous pile of text, and from your input it guesses the next word, one at a time.
trained on all of it
Where the conversation happens
Your words and its words are all tokens, and they share one bounded workspace called the context window.
One window is the floor, not the ceiling. A single agent can spin up a stack of these side by side, twenty reviewers on twenty contracts at the same time. That door opens in Part 2.
Get a feel for tokens
(around 2,700 pages)
How steering AI keeps changing
Ever handed a task to a colleague and gotten back something off target? Usually you left out what was in your head. AI is the same, so say the part you assumed.
Just ask
Say what you want in plain words and trust the model to read between the lines.
By tuning · your phrasing
Craft the ask
Engineer the one message: a clear brief, a sample to imitate, the reasoning laid out.
By tuning · examples and format
Stage the inputs
Decide everything the model gets to see, the documents, the memory, the thread, and refresh it each time.
By tuning · what you retrieve
Build the workspace
Set up the tools, files, permissions, and checkpoints it operates inside. For finance, that setup is the control environment.
By tuning · tools and limits
Asking well is just good delegation
You would not tell a first-week hire "fix the revenue." You would give them the who, the context, the reader, and the format. Same here.
Give it a role to step into.
"Act as a technical accounting manager briefing the controller."
The situation and the stakes.
"Our auditors are pushing back on how we booked this contract."
Pitch to the real audience.
"The reader is a CFO who wants the answer in ninety seconds."
Name the shape you want back.
"Give me a short memo with a decision up top and a citations list."
From perfect wording to the right inputs
The work shifts. You spend less time polishing one sentence and more time choosing what it sees: the contract, the policy, the prior memo, last quarter's notes.
You do it every time. Deciding what to hand over is the new craft.
→ Do that consistently, and you have basically described an agent.
One tuned message. You get the sentence right, it answers, you read it. Ideal for a single question.
A goal, plus the right material around it. You pick the inputs, it works across all of them, again and again.
Whatever you tried last, it has probably already improved
Two failure modes a CPA has to respect
Hint at the answer you want and it will hand it back, right or not. Do not coach it toward a verdict. Ask it to make the opposing case.
From chatting with it to handing it the job
You are the courier
- Nothing moves without you: pull it, paste it, explain it again, copy the result out
- The moment the data changes, you begin again
You are the manager
- You describe the outcome once, and it opens the files and systems on its own, in the open
- It comes back only for the calls that need you
🎯 Tie Out the Schedule
Round 1, done 😮💨
You got the tie-out, but it took 12 presses over 0:00, and none of it moved without you. You carried the data, re-explained the task, and cleaned up the output.
Now the same job, handed off instead.
MOD-118 repriced from $240,000 to $216,000 mid-term. The deferred balance sits $24,000 high until the schedule catches up. A draft adjusting entry is written and waiting on your approval.
A contract signed August 31 was booked in September. Looks like a period cutoff. Suggested handling: accrue it and note it in the close memo.
Side by side, one task, two ways
Round 1 · Chatting
Round 2 · Delegating
The work did not vanish, it relocated. It preps, you decide. That relocation is the entire point, and it is Move 5, Review it.
The same hand-off, on a live contract
├─ 📁 emails/ · the side conversations
├─ 📁 policy/ · our rev-rec positions
├─ 📁 contract/
│ └─ 📄 New_MSA_draft.docx
└─ 📁 memo/ · empty for now, keep an eye here
Every vendor is making the same move
The branding is a blur, so here is the shortcut. Under every new name, the shift is identical: a chat box grows the ability to touch your files.
| Maker | The chat you may already have | The hands-on version to ask about |
|---|---|---|
| Anthropic | Claude, in the browser | Claude Cowork · Claude Code |
| OpenAI | ChatGPT | its desktop agent · Codex |
| Gemini | its coding agent · the CLI | |
| Microsoft | Copilot in Microsoft 365 | Copilot agents · agent mode in Excel |
Skills: write the procedure down once
A skill is a short instruction sheet the agent checks before it works, your procedure, but one it can run. A shelf of them is the prompt-pattern library I teach:
"Get these journal entries ready for upload."
Plain language, no special syntax.
Each one lists a name and a one-line "use when", always visible:
name: je-upload-prep
use when: journal entries need formatting for our ERP
---
1. Lay them out in our column order
2. Debits must equal credits, or stop and ask
3. One line per entry, ready to import
The same procedure each time, for you and for whoever you pass it to.
🧭 Skills, up close
Where do these sit?
In the chat app, under Settings, then Capabilities. The built-in ones for Excel, Word, PowerPoint, and PDF ship switched on.
In a coding agent, they are a folder on disk. A quick command lists what is loaded.
· je-upload-prep · close-checklist · flux-notes
You wrote one. So where does it live?
In a folder of its own, holding one plain-text file, and that file is the whole thing. No app to build, nothing to deploy, no ticket to IT.
It is only text, which is precisely why a finance team can own it.
├─ 📁 close-checklist/ → the sheet
├─ 📁 flux-notes/ → the sheet
├─ 📁 je-upload-prep/
│ └─ 📄 one text file, that is it
└─ 📁 contract-review/ → the sheet
How do I run it, month after month?
Just ask in plain words. The "use when" line is the trigger, so "get these entries ready for upload" starts it every time.
Want to be sure? Name it outright. Either way, the same steps run. That predictability is the whole reason to write it down.
How do I give it to my team?
Hand over the folder. Zip it, drop it on a shared drive, or check it into the team's repo. A colleague adds it to theirs, and that is the install.
Now everyone's entries come out the way you set them. The file is both the procedure and the training.
Keeping the shelf from turning into a junk drawer
These folders sprawl the same way a binder of old procedures does. The discipline is dull, and it holds up:
- One job each. A catch-all "finance-helper" helps with nothing
- Spell out the "use when." A fuzzy trigger never fires, which is why most feel broken
- Stamp a date and owner in the file. "rev. Aug 2026, controller's team"
- Retire the ones that never run. Review quarterly, like any procedure
Your call: write a procedure, or just ask?
"Get these journal entries ready for upload, our column order, every month."
A simple test: recurring, plus your exact way, plus real steps, means write it down. A one-off, just ask.
Guessing work vs computing work
You have now watched an agent run a task end to end. The craft is splitting that task into the two kinds of work it holds:
An informed judgment call
- Weighs everything it has seen to land on an answer, the way you size up a contract you have never read before
- Strong where the question is open-ended
- Two runs can differ
- Fallible, so a person signs off on it
Same in, same out
- Spreadsheet math, scripts, rules engines, a calculator
- Identical result on every run
- A wrong one is wrong consistently, so you catch it once and it stays caught
- Straightforward to test and to audit
Assess where you stand, then build
Everything so far turns concrete in the frameworks I publish. Start by reading your own maturity, then build. Each is free to read in full, with its interactive version and the economics.
Read where you are
A self-scored model across five weighted dimensions, from strategy through execution to risk. See the gap before you spend.
Get the terms right early
Technical accounting that shapes a deal before it is signed. The live agent review you saw, run as a standing function.
Controls at scale
Fifty SOX 404 steps split between your team and AI, benchmarked against a GRC platform, with a live economics model. Split it and Review it, at scale.
ERP change, your terms
Plan a systems integration on the owner's side: scope it, hold total cost, and compare partners fairly. A live scorecard and cost model.
The single page that keeps AI work defensible
Before any of this reaches a workpaper, settle three things. This is the one-page governance sheet I teach, and it stays one page on purpose.
No final conclusion, no severity or scope call, nothing client-facing without a human owner. It drafts, a person decides.
Consumer tools for public data only. For anything sensitive, enterprise terms with no training on your inputs, a written boundary, and a quarterly look.
Sources real, math ties, logic holds, scope complete, AI use disclosed. A person signs, and the disclosure rides on the deliverable.
The small next step is the whole thing
Where to go from here
The AI Adoption Method, the Deals Desk, the AI-Augmented SOX Framework, the Integration Blueprint, and the AI Readiness model. All published, all free to poke holes in.
Applied AI Finance, alongside Corporate, Real Estate, and Entrepreneurial Finance. Interactive lessons, the prompt-pattern library, and the one-page governance sheet.
AI adoption, technical accounting, SOX and controls, and embedded teams that keep the work running after the engagement ends.
Go experiment. Then come read the rest.
Match the model to the job, not the logo
| The job | Anthropic | OpenAI | Microsoft | |
|---|---|---|---|---|
| 🏛 Hardest thinking 💰 top tier | Opus 5 · Fable 5 | GPT-5.6 Sol (Pro) | Gemini 3.6 Pro | Copilot on frontier models |
| ⚙️ Everyday volume ⚡ fast, cheap | Sonnet 5 · Haiku 4.5 | GPT-5.6 Terra/Luna · Codex | Gemini 3.6 Flash | Copilot agents |
| 🔬 Dig through sources | Claude, deep research | its deep research mode | Gemini, deep research | Copilot researcher |
| ✍️ Drafting and memos | Claude | ChatGPT | Gemini in Docs | Copilot in Word |
| 📊 Right inside the spreadsheet | Claude in Excel | — | Gemini in Sheets* | Copilot in Excel |
| 📽 Building the deck | Claude, with a slide procedure | — | Gemini in Slides* | Copilot in PowerPoint |