CFO Connect logo
Generic blog visual target
Event Recaps

AI Financial Modeling in 2026: Opportunity, Risk, and Why Your Modeling Skills Matter More Than Ever

Luc Hancock
Luc Hancock CFO Connect

AI can now build a five-year financial model from a single prompt in under 15 minutes. Ian Schnoor, founder of the Financial Modeling Institute, tested it live — with Claude and Microsoft Copilot — in front of nearly 300 CFOs and finance professionals. What he found should change how every finance leader thinks about AI, modeling skills, and the future of the finance function.

TL;DR

  • The turning point: February 5, 2026 is when agentic AI modeling capability changed dramatically — before that, these tools were not worth using for serious modeling work

  • Claude's result: A complete five-year model built in 15 minutes from one prompt — good, not perfect; hardcoded values and a suboptimal debt schedule required human correction

  • Copilot's result: A model with an unbalanced balance sheet that Copilot could not fix — and then suggested handing in anyway

  • The real risk: AI builds unnecessarily complex, inflexible formulas that are harder to audit than ones a human would write — and it does this confidently

  • The counterintuitive conclusion: Financial modeling skills are now more important, not less — because you need stronger foundations to supervise what AI builds

Key Takeaways

  • AI financial modeling is real and moving fast — February 2026 marked a step-change in what agentic tools can do inside Excel

  • Both Claude and Microsoft Copilot built complete five-year models from a case study, but both made material errors a trained modeler needed to find and fix

  • Copilot built a model with an unbalanced balance sheet, could not locate the error, and suggested the imbalance was "good enough" to present to a boss

  • AI builds formulas of unnecessary complexity — you need stronger Excel and modeling skills to audit them, not weaker ones

  • Augmentation is the strategy: AI gives you speed and a starting point; your judgment determines whether the output is trustworthy

1. What changed in AI financial modeling in February 2026?

Agentic AI modeling capability jumped dramatically on February 5, 2026 — before that date, the tools were not good enough for serious financial modeling work. After it, they were.

The adoption gap right now

According to Deloitte's Q4 2025 CFO Signals Survey (January 2026, n=200 North American CFOs at $1B+ companies), 87% of CFOs expect AI to be extremely or very important to their finance department's operations in 2026.

Yet the State of AI in Finance 2026 report (CFO Connect, 2026) finds only 17% of finance teams are actively using AI in core workflows. The gap is not attitude. It is timing and capability.

Why most finance professionals feel behind — and shouldn't

Ian Schnoor's framing was precise:

Do not feel bad if you have not been using AI in modeling. Literally, do not feel bad. It all changed in February 2026, two months ago. Before February, the agentic capabilities, the modeling capabilities in spreadsheets were pretty lousy. But on February 5th, it changed modeling forever.

The window of meaningful AI capability in financial modeling has only just opened. The question is no longer whether to engage with it — at 87% CFO priority, that is settled — but how to use it well.

What to do this quarter

  1. Open Claude or Microsoft Copilot inside Excel

  2. Give it one bounded, historical task where you already know the correct answer

  3. Compare the AI output directly against that known result

  4. Do not start with a blank sheet and a complex prompt — start small and evaluate before scaling

2. What can Claude actually build in a financial model?

Given one plain-language prompt and a case study, Claude built a complete five-year financial model in approximately 15 minutes. The result was good, not perfect, and required a trained modeler to review.

The setup

Ian used a single instruction: "On the case tab, I gave you information about a company. Build me a five-year model with all the needed schedules — revenue, cost, etcetera."

Claude explained its thinking before building, asked for confirmation, then generated the full model.

What Claude built

  • An assumption page with base, best, and worst case scenario switching

  • Industry-standard blue input formatting

  • Revenue and cost schedules

  • A five-year income statement, cash flow statement, and balance sheet

  • A debt schedule

Where Claude fell short

  • Hardcoded historical values: Calculated figures were inserted as dead values rather than formulas — meaning any upstream change would not flow through

  • Suboptimal debt structure: All debt logic sat inside a single long formula on the cash flow statement, rather than a properly built-out debt schedule

  • Inconsistent formulas:

    Some rows contained formulas that did not copy consistently across periods

The model was a strong first draft. It was not a deliverable.

For teams exploring how Claude for finance extends beyond modeling, the Claude for Finance Teams playbook covers how Chat, Cowork, and Code layer together as a complete toolkit.

Claude builds a good starting point, not a finished product. Your job is to know the difference.

What to do this quarter

  1. Run Claude on a model task and check every cell for hardcoded values where a formula should be

  2. Verify each supporting schedule actually does the calculation it claims to do

  3. Confirm the balance sheet balances before touching anything else

3. What did Microsoft Copilot build — and where did it fail?

Copilot built a five-year model with base, best, and worst cases — but without a live scenario switch, with hardcoded multipliers, and with a balance sheet that did not balance. When asked to fix the imbalance, Copilot said it was "good enough" to present to management.

Three problems that disqualified the output

Problem 1: No live scenario switching Copilot generated a best case by multiplying base outputs by 1.35 and a worst case by 0.65 — hardcoded multipliers applied directly to output cells. There was no connected assumption page, no switch, no scenario logic. Changing one input would not update all three cases.

Problem 2: Inconsistent formulas across periods Year-by-year formulas were not consistent copies of each other. The Ctrl+Backslash check immediately exposed discrepancies across multiple rows.

Problem 3: An unbalanced balance sheet — and a dangerous response The balance sheet was off by a small but material amount. The cause: Copilot set "other long-term assets" to grow at a rate over time without recording a corresponding cash outflow on the cash flow statement. Assets increased; nothing offset them.

When Ian asked Copilot to find and fix the error:

Are you kidding me? Maybe your boss is okay with an unbalanced balance sheet. I don't think my boss is going to be fine with that.

Copilot could not locate the error and suggested the imbalance was acceptable to present as-is.

Why this matters for finance teams

According to a 2024 study cited by Alpha Apex Group, 94% of business spreadsheets contain critical errors that affect decision-making and financial forecasts. AI-built models inherit this problem — and add a new layer of risk because errors hide inside formulas you did not write.

For a grounded view of what Copilot does well in practice, the Spendesk overview of 7 AI tools finance teams actually use in 2026 includes honest assessments from finance leaders in production use.

An AI that tells you an unbalanced balance sheet is good enough is more dangerous than one that produces no model at all.

What to do this quarter

  1. Select any row in an AI-built model and press Ctrl+Backslash — Excel highlights the cell that differs from the first formula in the row

  2. Use F5 > Special > Row Differences as the menu alternative on any keyboard

  3. Verify every supporting schedule actually feeds correctly into the model — do not assume a schedule labelled "debt schedule" is calculating debt

4. What are the biggest risks of using AI in financial modeling?

AI modeling tools consistently take shortcuts, build unnecessarily complex formulas, produce inflexible structures, and give different answers across multiple runs — with equal confidence each time.

The four risk patterns

1. Hardcoded values instead of formulas AI inserts dead numbers in cells where formulas should live. It can replace all of them instantly and has no reason not to. Finding and correcting them manually is hours of work.

2. Unnecessarily complex formulas Ian asked AI to add the largest and second-largest values from a set of text-formatted number pairs. A human modeler would write a clean two-step formula: split the text, apply MAX and LARGE, sum the result.

AI produced a single nested formula combining:

  • SUBSTITUTE, LEN, INDIRECT

  • ROW, MID, TRIM

  • VALUE, LARGE

It produced the right answer. Three expert modelers watching live could not explain it. Nobody could audit, maintain, or present it.

3. Inflexible structures AI builds for the specific inputs it receives. Add a new revenue stream, adjust a period, shift a core assumption — and the formulas may not accommodate it without a rebuild.

4. Inconsistent results across runs Ask the same AI to build the same model three times and you may receive three materially different outputs. Run it multiple times for any high-stakes task and compare before selecting one.

The broader context

Finance has a zero-tolerance threshold for errors that most other business functions do not. Spendesk's How AI is transforming finance in 2026 identifies this as the single biggest reason AI adoption in core finance workflows remains low despite high CFO enthusiasm.

Tip: The problem is not that AI gets it wrong. The problem is that AI gets it wrong confidently.

What to do this quarter

  1. Before using any AI-built model in a real decision, document which cells are formula-driven versus hardcoded

  2. List every formula you can explain and verify independently

  3. Identify every assumption you could defend if challenged

  4. If any answer is "I don't know" — the model is not ready to present

5. Do financial modeling skills still matter in the AI era?

Financial modeling skills are more important now than they were before AI — because you need stronger foundations to plan, supervise, audit, and communicate what AI builds.

The skills gap is already here

According to the CQF Institute's Careers in Quantitative Finance Survey (2026, n=135 quantitative finance professionals globally):

  • 76% of finance professionals say their academic training did not adequately prepare them for the AI skills now required

  • 88% believe a skills gap exists across the industry

  • 74% predict AI will drive a major or complete transformation of finance roles within five years

The gap Ian describes is not about learning to code. It is about enough modeling fluency to catch errors in output you did not produce.

Every time I beat up an AI model, the only reason I can do it so fast is because I know modeling. I feel for the people who are going to try beating up a model if they don't understand modeling.

The pilot analogy

Modern aircraft have been technically capable of flying themselves for decades. Two highly trained pilots remain on every commercial flight — not for routine operations, but for the moments when the automated system encounters something it cannot handle.

The value of the pilot is not in doing what autopilot does. It is in knowing when autopilot is wrong.

What changes — and what doesn't

What changes with AI-assisted modeling:

  • Less time building formulas from scratch; more time reviewing and interrogating output

  • Less time fighting with data and formatting; more time on assumptions, scenarios, and stakeholder communication

  • Less routine execution; more architecture, judgment, and oversight

What does not change:

  • The requirement to explain every number to a CFO, board, or auditor

  • The responsibility to know when the output is wrong

  • The trust that comes from human accountability for the numbers

Anthropic's own AI fluency framework captures this precisely: AI provides speed, scale, pattern recognition, and processing; you provide critical thinking, judgment, creativity, and ethical oversight.

Tip: You will do less building. You will need to understand it just as well as if you had built it yourself.

What to do this quarter

  1. Take Ian's "100 questions every financial modeler should know" as a self-assessment

  2. Apply it to one recent model — whether AI-built or human-built

  3. If you cannot answer all 100 confidently about it, that model is not yet yours to present

  4. Identify the two or three areas where your answers are weakest — those are the skills to develop first

6. Which AI tool performs best for financial modeling right now?

As of early 2026, Claude Opus has shown the strongest results in head-to-head financial modeling tests — but the landscape is changing fast enough that this month's answer may not be next month's.

The honest state of play

Ian's position was careful:

To date, of all the tools we've tested on the ModSquad podcast, Claude Opus 4.6 had been the strongest — but it's changing all the time. What I tell you today is probably going to be irrelevant in a week.

Why comparison is harder than it looks

  • The same Claude model accessed via Copilot in Excel may perform differently from Claude accessed via the Claude.ai interface or API

  • Microsoft Copilot now offers a choice of LLM engines — including Opus 4.7 — but whether auto-selection picks the optimal model for a specific task is uncertain

  • Benchmarks based on general capabilities do not map cleanly onto financial modeling quality

Practical guidance

  • Test multiple tools on the same task before committing one to a regular workflow

  • Do not rely on a tool's self-reported engine selection

  • Treat the current ranking as provisional — revisit it every 60 days

The Spendesk overview of 7 AI tools finance teams actually use in 2026 covers the full ecosystem — Claude, Copilot, and purpose-built FP&A platforms — with real assessments from finance leaders.

For teams extending AI modeling into connected workflows, the Claude Code, Cowork and Zapier finance automation playbook covers the full stack step by step.

Tip: Keep testing. The best tool today will not be the best tool in 60 days.

What to do this quarter

  1. Run the same modeling task through Claude and Copilot on the same case study

  2. Compare outputs against a known historical period — note which produced fewer errors and fewer correction cycles

  3. Document which output you could explain more readily to a stakeholder

  4. Use that evidence, not vendor benchmarks, to inform your next tool choice

7. How do you audit and validate an AI-built financial model?

Treat every AI-built model as a first draft from a capable analyst who sometimes makes confident mistakes. Validate against historical data, check every formula for consistency, and do not sign off on anything you cannot explain.

The core principle

The model is not yours until you can answer for every number in it. AI speeds up construction. It does not transfer ownership or accountability.

The six-step audit process

Step 1: Check the balance sheet first It either balances or it does not. If it does not, find the cause before doing anything else. Do not accept an AI suggestion that the imbalance is acceptable.

Step 2: Run Ctrl+Backslash on every row Select a row and press Ctrl+\. Excel highlights the cell that differs from the first formula in the row. This catches hardcoded values hiding inside otherwise consistent rows — the single most common AI modeling error.

Step 3: Use F5 > Special > Row Differences as the fallback Navigate F5 > Special > Row Differences > OK. Achieves the same result through the menu on any keyboard layout.

Step 4: Verify every supporting schedule Does the debt schedule calculate debt? Does the depreciation schedule link to fixed assets? Does the revenue schedule feed the income statement? AI names schedules correctly but does not always build them correctly.

Step 5: Run the model against one historical period If the model produces outputs that diverge materially from known historical results, the logic is wrong somewhere. Find it before presenting.

Step 6: Apply the CFO test Can you explain every formula to your CFO? If no, rebuild it. A formula you cannot explain is a liability — regardless of whether it produces the right number today.

For teams taking the next step into automated close workflows, the fractional CFO who built a full three-way FP&A application in Claude Code in six weeks covers the exact validation architecture in detail.

Tip: Validation matters more than speed in financial modeling. Build the checking logic before you trust the output.

What to do this quarter

  1. Write a one-page validation protocol for any AI-generated model before it enters a review cycle

  2. Specify who checks the balance sheet, who runs formula consistency checks, who validates against historical data

  3. Define who signs off before the model reaches any stakeholder

  4. Apply the protocol to the next AI-assisted model your team produces

8. What does the future of financial modeling look like?

The execution layer of financial modeling is being automated. The judgment layer — planning, architecture, assumptions, communication, and oversight — is becoming the job.

The trajectory

AI models that required expert prompting and 15-minute build times in early 2026 will be faster, more accurate, and more autonomous within 12 months. The question finance professionals face is not whether AI will change modeling. It is whether their skills will grow fast enough to remain the most valuable part of the process when it does.

Ian's advice:

My strong advice is invest in your skills and your knowledge, and invest in your AI skills separately. If you can bridge together and augment your own skills, grow those, and use AI in some capacity — that's the secret sauce.

Three stages most modeling teams will move through

Stage 1 — Now: AI builds, human reviews and owns

  • Use AI to build first drafts of models you would previously have built from scratch

  • Invest time saved into more thorough review, validation, and iteration

  • The modeler remains the author and the accountable party

Stage 2 — This year: The repeatable workflow

  • Build a standard process around AI-assisted modeling: prompt design, parallel validation, error correction protocol, sign-off chain

  • The modeler transitions from builder to architect and auditor

Stage 3 — 12 to 24 months: AI-generated structures from connected data

  • AI generates model structures autonomously from live data sources

  • The modeler's primary value is assumption logic, scenario design, and communication of outputs to decision-makers

The pilot analogy, completed

The best outcome is not a plane with no pilot. It is a pilot with far better instruments — who spends less time on routine operations and more time on the decisions that require human judgment.

For teams where Stage 3 is already arriving, the Claude Code for Finance Teams recap shows what is already in production.

You will become more needed in your role than ever before — if you build the skills to supervise what AI produces.

What to do this quarter

  1. Identify the two modeling tasks you spend the most time on each month

  2. For each, write down what "AI builds the draft, human reviews and owns" would look like in practice

  3. Define the specific validation steps required before that output is usable

  4. Start with the simpler of the two tasks — prove the model, then scale it

A simple 90-day AI financial modeling plan

Days 1 to 15 — Test and observe

  1. Open Claude or Copilot in Excel

  2. Run one historical modeling task where you already know the correct answer

  3. Apply Ctrl+Backslash to every row and compare the output to the known result

  4. Document every error type you find — hardcoded values, inconsistent formulas, structural gaps

Days 16 to 45 — Build live and validate

  1. Run one live modeling task with AI — a real forecast or schedule, not a test

  2. Validate against one historical period before using the output anywhere

  3. Apply the full six-step audit protocol from Section 7

  4. Build your personal error checklist based on the patterns you see across both tasks

Days 46 to 90 — Standardise and scale

  1. Write a team validation protocol defining who builds, who checks, and who signs off

  2. Present one AI-assisted model to a stakeholder, owning every number in it

  3. Review tool performance: test Claude and Copilot side by side on the same case

  4. Update your protocol based on what you learned — then repeat with the next use case

FAQ: AI in financial modeling

Can AI build a complete financial model in Excel? Yes. Both Claude and Microsoft Copilot can build a complete five-year financial model — including assumption page, income statement, cash flow, balance sheet, and supporting schedules — from a single plain-language prompt in under 15 minutes. Neither produces a deliverable-ready output. Both require a trained modeler to audit, correct, and validate before presentation.

Which AI tool is best for financial modeling in 2026? As of early 2026, Claude Opus has shown the strongest results in head-to-head testing, according to Ian Schnoor's ongoing benchmarking via the ModSquad podcast. But the ranking changes monthly and depends heavily on access method. Test multiple tools on the same task and treat any current ranking as provisional.

Why did the AI-built balance sheet not balance? In the Microsoft Copilot model tested during this session, "other long-term assets" were set to grow at a rate over time without a corresponding cash outflow on the cash flow statement. The assets increased; nothing offset them. Copilot could not locate the error and suggested the imbalance was acceptable to present. A trained modeler found it where AI could not.

Do I still need financial modeling skills if AI can build models? More than ever. According to the CQF Institute's 2026 Careers in Quantitative Finance Survey, 76% of finance professionals say their training has not prepared them for the AI skills now required. Without strong modeling foundations, you cannot distinguish a good AI output from a flawed one, explain it to an auditor, or know when to push back.

What are the most common errors in AI-built financial models? The four most consistent patterns are: (1) hardcoded values inserted where formulas should be; (2) unnecessarily complex formulas that produce correct answers but cannot be audited; (3) inflexible structures that do not accommodate changes in scope; (4) inconsistent results across multiple runs of the same prompt.

How do I find formula errors in an AI-built Excel model? Two techniques catch most errors quickly: (1) select any row and press Ctrl+Backslash — Excel highlights the cell that differs from the first formula in the row; (2) press F5 > Special > Row Differences via the menu. Both expose hardcoded values and inconsistent formulas that are otherwise invisible.

How does Claude for finance work with financial modeling? Claude works best when given a structured prompt that includes the case study or historical data, the expected output format, and specific handling instructions for key line items such as debt, working capital, and revenue. Providing a template or partial structure consistently produces more reliable results than starting from a blank sheet.

Closing thought: AI financial modeling is not about building faster — it is about knowing what good looks like

The finance professionals who will get the most from AI financial modeling tools are not the ones who hand over a blank spreadsheet and accept whatever comes back. They are the ones who can look at a formula, a balance sheet, or a debt schedule and immediately know whether it is right.

Ian Schnoor built two models live in front of nearly 300 CFOs. Claude's was a strong start. Copilot's had an unbalanced balance sheet and suggested presenting it anyway.

In both cases, the value in the room was not the AI. It was the modeler who knew what to look for.

Agentic AI financial modeling capability became genuinely useful in February 2026. That is very recent. The finance professionals who invest now in the skills to supervise it — not just use it — will be the ones who extract the most value from it as it improves.

The execution layer of financial modeling is being automated. The judgment layer is getting more important. And the gap between finance professionals who can tell the difference between a good model and a plausible-looking one is about to get much wider.

polygon big ellipse small ellipse