AI for Financial Modeling: What It Does Well, and What You Must Check

AI is good at the structure of a financial model and unreliable at the numbers inside it. It drafts a tab layout, writes formulas, explains someone else's workbook, and documents assumptions well. It also produces figures that look correct and trace back to nothing.

That split decides how you use it. Treat AI output as a first draft from a fast junior who never shows their work, and the tool becomes useful. Treat it as an answer, and you sign off on numbers you cannot defend.

This guide covers what AI genuinely does well in modeling, where it fails, and how check digits and tick-and-tie controls turn an AI-assisted model into something you can audit and revise.

Can AI Build a Financial Model You Can Use?

AI can build the skeleton of a working model, but not a model you can sign off on without review. It produces tab structure, formula logic, and documentation quickly and accurately enough to save real hours.

The numbers are the problem. A chat model will fill a growth rate, a margin, or a comparable multiple with a plausible figure and no source, and it will not flag that it did so.

So the split is clean. AI owns scaffolding and explanation. You own every input, every output, and the controls that connect them.

• Good: tab structure, formula drafting, assumption documentation, scenario setup
• Good: explaining or restating a model someone else built
• Bad: sourcing figures, benchmarks, or market data
• Bad: guaranteeing links between tabs survive a revision

What AI Does Well in Financial Modeling

AI is strongest on the parts of modeling that are structural rather than numerical. These are the tasks where the right answer is a pattern, not a fact.

Scaffolding a first draft is the clearest win. Describe a three-statement model, a 13-week cash flow, or a restructuring wind-down, and you get a tab layout, row labels, and formula logic in a few minutes.

Explaining an inherited model is the second. Paste a tangled formula and ask what it computes, and you get a readable answer faster than tracing precedents by hand.

• Scaffolding — tab structure, row labels, and section order for a standard model type
• Formula writing — nested INDEX/MATCH, XLOOKUP, date logic, amortisation schedules
• Formula explanation — plain-English translation of a formula you inherited
• Assumptions documentation — writing the assumptions page most models never get
• Sensitivity setup — building the case structure for upside, base, and downside
• Consistency checks — spotting that one tab uses 12 months and another uses 13 periods
• Restatement — converting a model built in someone else's conventions into yours

Where AI Gets Financial Models Wrong

AI fails in financial modeling by being confidently wrong in ways that look right. This is more dangerous than an obvious error, because a plausible number passes a quick skim.

The four failure modes below cause almost all of the damage. Each one is quiet, which is what makes it expensive.

None of these are fixed by a better prompt. They are fixed by controls in the workbook.

• Invented figures — a growth rate, churn assumption, or market size stated with no source
• Silent arithmetic drift — a formula that computes something close to, but not exactly, what you asked
• Broken cross-tab links — a revision that rewrites a driver on one tab and leaves the reference on another
• No audit trail — an output cell with nothing recording where the input came from or who approved it
• Stale logic on revision — the model updates the section you mentioned and leaves a dependent section behind

Check Digits and Tick-and-tie: How to Make an AI-assisted Model Trustworthy

Check digits are self-testing cells that prove the model is internally consistent, and they are the practical answer to "how do I trust this." Each one computes a difference that must equal zero and turns red when it does not.

Ask for them explicitly when you prompt. A model built without them may be correct today and quietly break on the third revision, and nothing in the workbook will tell you.

In the document-extraction work we build for professional-services firms, the recurring failure mode is not a wrong number that shouts. It is a machine-produced number that looks right and has no trace back to its source. Tick-and-tie controls exist to make that failure visible, which is exactly why an AI-assisted model needs more of them than a hand-built one, not fewer.

• Balance check — assets minus liabilities and equity, must equal zero
• Cash tie — closing cash on the cash flow equals closing cash on the balance sheet
• Cross-foot — row totals and column totals of the same block agree
• Source tie — every hardcoded input carries a cell note naming its source document and page
• Roll-forward — opening balance plus movement equals closing balance, every period
• Error sweep — a single cell counting every error value across the workbook
• A control panel tab that shows every check in one place, green or red

What to Trust AI with, and What a Modeller Must Own

Trust AI with anything you can verify by reading it, and own anything you can only verify by testing it. That single rule sorts almost every task in a modeling engagement.

The table below applies it task by task, with the control that makes each one safe.

A Realistic Workflow from Prompt to Reviewed Model

The workflow that works treats the AI as a drafting step inside your existing review process, not as a replacement for it. Seven steps, in order.

The order matters most at step two. Controls specified before the model is built get wired into the structure, and controls requested afterwards get bolted on badly.

Expect three or four rounds. The client wish list grows, and a model with working check digits survives that; a one-shot answer does not.

• 1. Write the spec — model type, periods, output schedules, and the decision it supports
• 2. Name the controls up front — list the exact check digits you want before any building starts
• 3. Generate the skeleton — structure, labels, and formulas, with inputs left blank
• 4. Fill inputs yourself — every hardcoded figure sourced and noted by a human
• 5. Test with known data — run a period you already have an answer for and reconcile
• 6. Review formula by formula on the critical path — revenue, cash, and covenant lines
• 7. Lock and log — protect the calculation cells and keep a change log for the next revision round

What AI Actually Produces in Excel and Google Sheets

You get one of three output types, and they are not equally useful. Knowing which one you are asking for prevents most disappointment.

The first is text you paste in: formulas, row labels, and structure described in chat. This is the most reliable option because you place every piece yourself and see what lands where.

The second is a generated file. Some assistants can write a real Microsoft Excel workbook or produce a script for Google Sheets. Capabilities differ by product and change often, so confirm on the vendor's own documentation rather than assuming.

• Pasted formulas and structure — most reliable, most manual, easiest to review
• A generated .xlsx file — fastest to a working draft, hardest to audit line by line
• Apps Script for Google Sheets — good for repeatable build steps, needs code review
• In-app AI features inside the spreadsheet itself — convenient, still needs the same check digits

Where the Human Must Stay in the Loop

The human owns judgment, sourcing, and sign-off, and no amount of AI capability moves those. A model is an argument about the future, and someone has to be accountable for the argument.

Three jobs never delegate. Choosing which drivers matter, sourcing every input, and deciding what the output means for the client.

Revision rounds are where discipline slips. When the wish list grows late in an engagement, the temptation is to let the AI patch the model quickly, and that is precisely when check digits earn their keep.

• Driver selection — which variables the model is genuinely sensitive to
• Input sourcing — every figure traced to a document, a system, or a named person
• Reasonableness review — does this output make sense given what you know about the business
• Client communication — what the model says, what it assumes, and what it cannot answer
• Professional sign-off — your name, your responsibility, regardless of what drafted it