On September 24, 2026, thousands of people around the world joined the first 24-Hour Global Financial Modeling Conference to hear from more than 40 experts on the future of the profession.
AI can now build financial models, and it gets better every month. But a model is only useful if people can trust it. AI can get the answer wrong and sound just as certain as when it’s right.
That’s why you need a person who knows how the model should work, can spot when something is off, and will stand behind the result. That takes strong fundamentals and good judgment. It’s also what makes others trust the model, and the person who built it.
The experts did not downplay what AI can do. Carolina Lago, a financial modeling educator with 25 years in FP&A, put it bluntly: “AI can build financial models. Don’t let anyone tell you otherwise.” Several experts noted that AI only became genuinely practical for financial modeling around February 2026, and it has improved dramatically since.
The Google Sheets team gave an example. They tested their AI assistant, Gemini, on past cases from the Financial Modeling World Cup, a timed competition in which top modelers build a complete model from scratch. On one of the hardest 2024 cases, which typically takes an expert 45 minutes to an hour, Gemini built a fully linked 12-month model in about eight minutes and scored every available point. Across 139 past cases, it was accurate about 90% of the time.
Brian Jones of Microsoft’s Excel team said people who used to ask the office Excel expert how to write a formula now ask a different question: “Can I trust these results?” Sloka Balasubramanyan of EY India made the same point. Nobody wonders anymore whether AI will produce an answer. They wonder whether they can trust it.
Speed to build is interesting.
Speed to trust is gold.
Ian Bennett of PwC Australia pointed out that building the model may be only about 20% of a modeling project. The rest is understanding the problem, designing the model, reviewing it, and handing it over. Cutting build time in half saves about 10% of the project. The bigger prize is speed to trust: how quickly you get to a model people can rely on.
Building is about 20% of a modeling project.
AI tools work by predicting the most likely next step. Usually that works well. When it doesn’t, the result does not look like an error. As Carolina Lago put it, “The worst part is not that it’s wrong. The worst part is that when it’s wrong, it looks right.”
Lance Rubin, a fractional CFO who tests every new AI model before relying on it for client work, shared a test. He and fellow modeler Craig Hatmaker each built the same software-company model by hand. Their answers for the key output, the investor’s internal rate of return (IRR), were close: about 23%. Lance then ran the same assignment 13 times across several AI tools.
The two hand-built models landed at about 23%. Drag the blue marker to your guess.
{{ guessVerdict }} The AI-built models produced IRRs from 13% to 158%.
Several of those models looked excellent. They were well laid out, they balanced, and they had dashboards. Lance’s verdict: “The best-looking model was confidently wrong.” Even the strongest AI result, guided by hundreds of detailed instructions Lance had written, came in further from the correct IRR than either human-built model. He also caught AI tools quietly forcing the balance sheet to balance with a plug figure, and one tool even asked for the answer file so it could work backward.
98% correct is still 100% wrong.
A model that is almost right can still lead to a badly wrong decision. A balance sheet that balances proves the arithmetic adds up. It does not prove the answer is correct.
Even the Google result deserves a second look. Ninety percent accuracy on competition cases is remarkable, but real work comes without an answer key. If one model in ten is wrong, someone still has to know which one. The Google team added a further limit: AI does not know a company’s contracts or policies on its own. “You, the analyst, remain the driver.”
At 90% accuracy, one model in ten misses. There is no answer key.
{{ pickVerdict }} {{ pickNote }}
If AI can do the calculations, what is left for the modeler? Expert after expert gave the same answer: deciding what the model is for, understanding how the business really works, and knowing whether an odd number is an error or the real story.
Defining the problem statement remains the most critical part of building the model, and that’s not something you can delegate to AI. Not yet, probably not ever.
Carolina Lago showed what this looks like in practice. She ran an AI workflow she designed to build a three-statement model, with built-in pauses where the AI must stop and check with her. Early on, the AI flagged something odd in the company’s history: cost of sales was 128% of revenue in 2021 and 116% in 2022, then dropped to 62% in 2023. The company lost money on every sale for two years, then suddenly became solidly profitable.
{{ cosNote }}
The numbers alone could not explain why. Carolina could: those two years reflected the company absorbing an acquisition. Had the AI simply averaged the three years, it would have assumed cost of sales of roughly 102% of revenue in every forecast year. That describes a business that loses money on every sale, forever.
Later, the forecast showed margins climbing to 39% and then fading. Every automated check passed, but the curve looked like a bug. As Carolina put it, “It just looks wrong.” She knew it was right: the business was reaching its production capacity. An analyst without that context would likely have smoothed the curve and deleted the very insight the model was built to find.
You cannot orchestrate what you cannot understand.
David Brown, founder of dbrownconsulting in Nigeria, argued that as AI takes on the technical work, a modeler’s value moves to the thinking behind the model. He described six thinking patterns, each tied to a real business story. Their first letters spell WISDOM.
{{ wCur.question }}
{{ wCur.lesson }}
His summary: “AI brings the intelligence. You bring the wisdom.” AI can be a useful thinking partner for each pattern, for example by mapping how a business works or proposing three ways to structure a model. But deciding which approach to take stays with the modeler.
Sebastian Szameitat, an independent modeler in Luxembourg, closed the conference with an example from renewable energy. A solar and battery project must never send more than 100 megawatts to the grid at any moment. AI can confirm the model contains that limit. But the model works in three-month time steps, and a three-month figure cannot prove the project stays under the limit at every instant. The right conclusion is not “compliant.” It is “not yet proven,” and the question stays open.
The danger, Sebastian warned, is not only that AI gets something wrong. It is that AI produces something plausible enough that people stop asking the next question.
If judgment protects a model, fundamentals make judgment possible. Wayne Kennedy, who co-leads KPMG UK’s financial modeling capability, put it simply:
You cannot assess whether the AI is doing what you want it to do unless you understand what it is that you want it to do.
Relying on AI for work you could not do yourself, he said, is like a manager who delegates work they could not coach anyone to do. It is abdicating responsibility.
Alastair Matchett of Financial Edge Training described what large investment banks saw with their 2026 summer interns, who used AI for almost everything. One bank tracked how heavily each intern used AI, and the heaviest users were not necessarily the best performers. Interns who did not understand the work produced generic output anyone could have made. When banks reviewed the prompts, output quality closely tracked prompt quality, and the best prompts came from interns who understood the modeling. Some banks now test new analysts three ways:
On fundamentals.
With AI allowed.
Explain and justify the work to a panel.
Steve Xing, who leads Deloitte UK’s valuation modeling business, shared his team’s golden rule: “If you don’t understand the change, don’t implement it.” AI can generate a schedule, but the modeler must fit it into the model, check it, and be able to explain how the results changed.
Fundamentals also give you a feel for what a sensible answer looks like, which is often how errors get caught. Brian Egger, head of financial modeling at Bloomberg Intelligence, runs simple checks on every model. None of them require advanced techniques. All of them require knowing what normal looks like.
“Better that you ask that question than your boss or a client.”
Brian Egger, Bloomberg Intelligence
Andy Marsden, who leads Kroll’s global modeling team, told a story many will recognize. Reviewing a model, he asked why the business was losing money in every year. The reply: “Is it? I hadn’t noticed.” His fix is to form an expectation before looking at results: how revenue has trended, where margins should head, what returns everyone expected. That expectation becomes the yardstick for spotting problems, including the ones AI introduces.
Trust came up in nearly every session, and experts agreed on where it comes from. Vaughan Grandin, head of financial modeling at Teneo, often works with companies in financial distress, where lenders and boards have stopped believing the forecasts. He named three things that build credibility: understanding the business, being understandable, and making sure the numbers make sense.
The last matters most: “Nothing evaporates credibility as fast as a nonsensical answer.” He also named a new risk to credibility, a client asking, “Did you use AI to build this?” If the modeler cannot show they understand and own every part of the model, the client starts to wonder what they are paying for.
{{ revLine }}
{{ revVerdict }}
Steve Xing and Ian Schnoor, FMI’s Executive Director and the conference host, made the same point. If someone asks what a number means, blaming the data or saying “AI did that” is not an explanation. It is the fastest way to lose their trust.
Jeff Tan, a CFM in Malaysia who has worked on more than 100 transactions, showed what this looks like when money is on the line. In an M&A sale, the buyer challenges every assumption, because every dollar removed from the valuation is a dollar they don’t pay. If the seller simply says revenue will grow 15%, the buyer will dismiss it. If the seller shows that 4 points come from price increases already in contracts, 3 from existing customers growing, 5 from signed backlog, and 3 from new customers, the buyer has to say exactly which piece they disagree with. His test: if you can’t explain an assumption in one sentence, you probably need to do more work.
{{ bridgeNote }}
It’s not enough for a model to be right. The users need to have confidence that it’s right.
Ian Bennett made the same link on the opening panel: the technical disciplines of modeling “are the only thing that will deliver trust.”
Rob Langrick, Chief Product Advocate at CFA Institute, makes about 24 business trips a year and meets finance professionals all over the world. For two years, he said, the number one topic has been AI, and everyone wants to know how everyone else is using it. In a fast 20-minute session, he shared his ten favorite tips from the road.
{{ tipCur.body }}
His last tip came with a story. Judging the CFA Institute Research Challenge in Dublin, Rob asked seven university teams whether climate risk would change their recommendation on a home builder. Six teams panicked. The seventh answered calmly. They had asked AI for the hardest questions a judge could ask and prepared answers to all of them.
That story captures the conference theme. AI did not answer for that team in the room. It helped them prepare to answer for themselves.
You have to have that foundational knowledge to ask better questions.
Rob was clear that the people who get the most from AI are those with a solid foundation, who can tell when something in a company’s accounts looks off. He does not see AI letting ambitious people go home early. It takes away the grunt work so they can ask more questions and dig deeper, which is where the real insights are found.
Beyond Rob’s tips, the practitioners who spoke shared a consistent set of working habits. None involve trusting AI blindly.
{{ h.body }}
Will AI take modelers’ jobs? The Big Four leaders on the opening panel did not think so. Wayne Kennedy recalled hearing the same prediction when computers, and then Excel, arrived: the work changed and the opportunities grew. Sloka Balasubramanyan compared it to saying computers would make accountants unnecessary. Steve Xing offered a one-line answer:
AI is not taking my job. It’s taking the part of my job that I don’t want to do and find boring in the first place.
What will change is how modelers spend their time. Andy Marsden suggested that modelers may soon spend half or more of a project on the analysis and advice clients actually value. Carolina Lago called it going deeper, not just faster.
Sebastian Szameitat pointed to work modelers have always known they should do but often skip under deadline pressure: clear documentation, assumptions traced to their sources, records of what changed between versions, and real stress tests. AI makes all of these far cheaper to produce, which removes the excuse for skipping them. In his words, “AI does not lower the bar, it raises it.”
The panel on women’s careers in financial modeling looked at the same shift from another angle: as AI takes on more mechanical work, the qualities that endure are good judgment, consistency, and working well with others. And Ian Bennett’s closing thought on the opening panel was an optimistic one: the profession “genuinely has a very strong future.” It will be different, he said, but not as frighteningly different as many fear.
{{ t.body }}
AI is changing how models are built, and it will keep changing quickly. What it has not changed is who is accountable for the answer. The future of the profession is not AI on its own. It is a skilled human modeler using AI: someone with the knowledge to direct it, the judgment to challenge it, and the ability to take responsibility for the result. Strong fundamentals do not compete with AI skills. They make AI skills more useful.
Learn core financial modeling concepts at your own pace. Access all conference videos and presentation materials.
AI can generate a model in seconds, but it can't tell you if the logic holds. The AFM proves you can build trusted financial models. Access learning resources and the full conference video library.
Every theme in this report comes back to the same foundation: knowing how a financial model works well enough to build it yourself. Carolina Lago made the link directly in her session, pointing to the AFM as proof of exactly that foundation.
The Advanced Financial Modeler (AFM) is a program that results in an accreditation. It is designed for people who already build models, or who want to learn to build a three-statement model from scratch. Candidates prove they can build and work through a model themselves, which is the knowledge experts said is needed to direct AI, challenge its output, and take responsibility for the result. For any modeler who wants to use AI with confidence, the AFM is the place to start.
You will never know how to command AI if you don’t know how to do it.
Carolina Lago, Tactic Financial
{{ step1Text }}
The post below opens in LinkedIn already written, and goes on your clipboard too. Edit it here first if you like.
{{ shareStatus }}
{{ shareStep3 }}
