Silicon Valley’s Latest Pitch to Wall Street: Double Down on Your Spreadsheet Problem

Anyone who has worked on Wall Street will readily concede that high finance has a burning workflow problem. That problem, however, has never been the cost of its junior-most resources. The value-destructive logjam exists at the mid-level: crushed from both sides by (justifiably) perfection-demanding seniors above, and an ever-ramping, revolving door of juniors below. When a firm’s effective capacity to take on new deals taps out, it is almost always due to a pinch at this level. The downstream impact? New opportunities may be passed up entirely, or stones left unturned in diligence. Bad outcomes for GPs and LPs alike.

Today, competition for winning high quality deals or mandates across private equity, private credit and investment banking is more intense than ever, which translates to more “at bats” required per completed deal than historically. As the “top-of-the-funnel” has exploded, the quality bar for analysis required at each step hasn’t softened to accommodate this increase (nor should it, with trillions of pensioners’ retirement wealth at stake), nor have deal teams grown in size to absorb the shock. In fact, the horsepower available at the bottom of the pyramid has dulled, in part due to the reverberations of COVID-era training gaps, in part due to entry-level finance jobs losing some luster versus tech and AI, and in part due to work-life balance initiatives that firms aim solely at their junior-most staff. The end result? A mid-level jammed with a higher volume of lower quality work-in-process with a senior bench no more tolerant of sloppiness than they were two decades ago.

Now, a new player has entered Wall Street’s arena with promise of some reprieve. Hundreds of billions of dollars raised towards genuinely transformative technology impacting nearly every end market, with promise for ours too. Large-language models (“LLMs”) have made their way into the daily workflow of every professional who wants to maintain relevance for the next several decades. Truly remarkable stuff with the potential to transform our economy and perhaps even humanity itself. There is simply no faster way to aptly summarize an expert call, Confidential Information Memorandum, or investment memo than running it through an LLM. For language-based tasks, where one randomly selected synonym in place of another makes no difference to the quality of the output, large-language models are the undeniable tool for the job.

AI has also transformed our business at Mosaic. Our engineers are more productive in shipping features customers want, and our user experience has been elevated by AI agents embedded in the platform itself. Users can now spin up a new model just by sending a one-line email to Mosaic, thanks to AI. Agents accessing deterministic tools is a beautiful thing (it is in fact the main driver of perceived performance improvements of the leading chatbots themselves).

What we would never allow LLMs to dictate is the mathematical formulas, relationships, and calculation schedules for any of our fundamental analyses at Mosaic (e.g., LBO, DCF, Operating Model). I can’t think of a worse engineering design decision than opening these up to randomness.

In the last few months, a potentially disastrous extension of transformational AI technology has been proposed by AI hyper-scalers that have never set foot on Wall Street: let LLMs run the spreadsheet supporting your deals.

I suppose if all you have is a hammer, everything looks like a nail.

I can completely understand and appreciate the temptation of this for people who have never felt the weight of owning a model for a multi-billion-dollar buyout. Why spend hours updating your firm’s template when LLMs can write formulas for you? How hard can it be to tell an LLM to change an exit multiple from 10.0x to 9.75x, or save down a few common formulas in a SKILL.md file?

The pervasive problem with analysis on Wall Street isn’t the speed, cost, or even quality of the individual humans producing it. It’s the variance in the subcomponents of deal analysis produced by pairing a non-deterministic resource (human or otherwise) with an analysis tool that has zero guardrails (e.g., a spreadsheet). For exploratory work (e.g., dream up a creative new deal structure for us to discuss), this combo is powerful. For execution of established analyses supporting multi-billion-dollar investment decisions (e.g., LBOs, DCFs, Comps, Precedents, etc.), this combination is a structural problem responsible for much of the wasted effort in our industry. How does the industry solve for that variance today? By dumping it all on the plates of the already overstretched mid-level in the form of never-ending manual reviews to catch the exact same mistakes made spreadsheet to spreadsheet.

The idea of firms selecting a solution that intentionally increases this variance to save cost in the lowest cost cohort of a deal team is comical.

We at Mosaic have always believed that the ideal, north star solution this industry deserves is a 100% precise, deterministic system for deals analysis fitted to the investment philosophies of the firm that deploys it. One that if run a million times would produce the exact same result one million times – and not one time fewer.

A system with precision inherent to the method by which it was engineered. Private Equity has been around for far too long for us to still be catching mistakes or doubting math in models presented at Investment Committee. The reason why this problem exists is we’re using the wrong tool for the job – and Silicon Valley is now proposing we double down on it.

Think about your calculator.

How often has it randomly decided that 7 + 9 equals 16 today and 79 tomorrow? Would you ever think to second guess or double check its math?

Never.

That reliability is a direct product of the method by which it was built: leveraging deterministic code, not stochastic or probabilistic processes, for its core calculation engine.

Think of how much of the world’s mindshare has been freed up because we can rely on devices to calculate things correctly. We at Mosaic envision a world where deals can be analyzed with the exact same confidence. Never should there be a question of “is this math working right” – the questions should only be on the theme of “is this business growing at, above or below market? Why? And how does that impact our case?” If our vision became reality, model turns would no longer exist, and so much waste would be eliminated from the system.

The vast majority of people in investing roles on Wall Street understand this distinction immediately and viscerally. The idea of arming junior teams with a technology that produces “correct-looking” random analysis for mid-levels to check and own the accuracy of gives me literal nightmares. I would bet my career (in fact, I have) that this cannot be the long-term solution to our long-standing problem. What concerns me a bit in the short-term, however, is the existence of bandwagon jumpers tempted by the idea of “doing something in AI” and reaching for the most aggressively marketed hammer – and beating their teams with it until they too believe they’re nails.

This week at Mosaic we received a frantic, late-night helpdesk request from a junior user at a large private equity firm. They believed a calculation was off in a model they had sent to their Investment Committee. Their Adjusted Levered Free Cash Flow Yield was presented as negative, despite having positive in-year free cash flow after paying down all cash interest. How could this be?

Simple. Their firm’s agreed upon definition of Adjusted Levered Free Cash Flow Yield – for this deal, and every deal they had presented to Investment Committee through Mosaic – burdens the numerator of their yield metrics for all interest incurred on debt, including paid-in-kind (“PIK”) interest, to reflect the true economic cost incurred throughout the capital stack above them, and make deals comparable regardless of the payment form of interest. The Mosaic team responded with this detail, and the user happily and confidently moved on, having learned something about their firm’s investment philosophy and modeling standards.

Imagine the LLM-in-Excel parallel universe. First off, there would be no one to call. The hyper scalers couldn’t care less about private equity or our ability to compound capital for pensioners. To them, all we’ll ever be is a deep-pocketed end market on a board deck slide titled “AI will wipe out all white collared jobs.”

But back to our chatbot-in-Excel-future example:

In the chatbot, you’d ask your cheery spreadsheet friend, why did you calculate the LFCF Yield wrong? It should be positive!

To which you’d receive the helpful response you’d deserve:

“You’re absolutely right. I’ve updated it to be a hardcoded positive.”

I think we’re on the precipice of a great breakthrough in our industry. Investment firms are finally making meaningful investments in technology and AI to augment their investment processes – in fact, in 2025 alone, five of the top ten global private equity firms selected Mosaic to bring deterministic automation to their investing process, as did two leading investment banks. The best investors I know are making these decisions the same way they approach new deals: diligently, thoughtfully, and unbiased by the noise of marketing fluff (no matter how loud it is).

In the world of LLMs-in-Excel, I often think of Henry Ford’s quote: “had I asked the average consumer what they wanted, they would have asked for faster horses.”

Each private equity firm, private credit firm and investment bank is now faced with a choice in 2026 – do we codify our way of investing into a system that by its design cannot make a math deviation model to model, or hand the reins to randomness in the hopes that “it will get better one day?”

Is this industry going to saddle up faster horses, or see where the automobile can take us?

Am I biased? Absolutely. I’d love to see everyone use Mosaic as the best outcome for this industry that I love.

What is the second-best outcome, you ask? Build a deterministic system yourself, internally, around the process that has made you successful investors or advisors over the past several decades. Codify, and standardize the processes that make you great, such that they exist long after your best investors retire.

Third best? Keep doing what you’re doing, with the people you’re doing it with. Sure, they’re wrong sometimes, but they do more than just this one workflow, and at least you’ll have someone who will own that mistake when it inevitably occurs.

The very, very last resort? Delegate this workflow to randomness in the name of being on trend.