Conversation
Highlights

How F2 Built AI That Private Markets Investors Actually Trust

The private equity bullpen at 3 AM is where reliability matters most. Glassy-eyed analysts audit financial models on their 150th version, hitting F2—the Excel keystroke that traces every calculation back to its source. One miscalculation in an LBO model could kill a half-billion-dollar deal.

Donald Muir lived this reality for years in middle market and late stage private equity buyout, working long weekends doing commercial due diligence and investment committee memo generation for large financial sponsors. Every workflow ran offline through Microsoft suite—Excel, Outlook, brutally inefficient.

In a recent episode of BUILDERS, Donald shared how that firsthand pain became the foundation for F2, the AI platform for private markets investors. But the company’s origin reveals a non-obvious path: building critical internal infrastructure first, then commercializing only when customers explicitly demanded access.

From Direct Lender to Debt Marketplace

After Stanford Business School, Donald founded ARK as a direct lender. “I raised hundreds of millions of dollars of debt alongside my venture rounds. And I lent money to other startups and then I collected it,” he explains. This wasn’t just lending—it was experiencing the due diligence pain points firsthand.

ARK evolved into a debt marketplace connecting hundreds of private credit funds with thousands of tech companies. The business model worked: lenders provided capital to ARK’s banking clients, ARK facilitated transactions. But the debt placement process created a scaling constraint. Each loan application required manual underwriting, data standardization from messy Excel files, and custom screening memo creation for lenders.

With generative AI emerging in late 2022 and early 2023, Donald saw an opportunity to automate what was becoming ARK’s operational bottleneck.

Building AI for Internal Survival, Not External Sales

Donald built AI specifically to automate ARK’s debt marketplace. “We get loan applications from banking clients and I built AI to take that data, standardize the very messy kind of Excel based data set from lower middle market companies and then create custom screening memos to surface those debt investment opportunities to the lenders,” he explains.

The key architectural decision: focus on the hardest engineering problem first. “We were laser focused on the hardest engineering problem first because we needed to do that in order to make the marketplace work at ARK,” Donald says. That problem was reliable financial analysis of unstructured Excel data—making AI that could actually do math without hallucinating.

The technology unlocked massive scale. ARK went “from tens of millions of loan volume at ARK to tens of billions of loan volume thanks to this AI native underwriting system.”

Then came the inflection point.

When Customers Request Your Internal Tools

The institutional lenders on ARK’s marketplace—particularly upmarket players—started making a specific request. “The aha moment for me was when the lenders particularly upmarket on the marketplace, said to me, we want to use your AI for not just the deals you’re sending us in the marketplace, but for all of our deal flow,” Donald recalls.

This wasn’t casual interest. These firms wanted to deploy ARK’s AI across their entire investment pipeline, not just marketplace transactions. But building that product would be non-core for ARK’s debt marketplace and banking business.

Donald made the call to spin out. “I went back to the investors and my team and my customers and saw overwhelming support to start what is now F2.”

The new company launched with structural advantages: former ARK employees who built the original technology, all 50+ of ARK’s investors participating in the first funding round, and several ARK lenders as initial F2 customers. The teams still share office space—100% in-person in San Francisco and New York financial districts—maintaining close operational ties while pursuing independent strategies.

Engineering for Auditability Before Speed

F2’s core technical differentiation addresses the fundamental barrier to AI adoption in high-stakes financial decision-making. “LLMs hallucinate, they get stuff wrong. They can’t do math,” Donald states plainly. “We’ve developed technology that does not hallucinate, that can do math.”

The company developed what Donald describes as “a category leading Excel agent product” that became core to the platform. Over a year later, foundation models still haven’t replicated this capability. “We cracked the code on financial analysis for unstructured XLS data before anyone else. And that continues to be the case as of this meeting today.”

Why does F2 maintain this lead? “I think it’s because we developed it for our own organic internal use case,” Donald explains. Generic AI tools and foundation models aren’t focused on solving this specific pain point. F2 built it because the marketplace required it—creating technology shaped by actual operational necessity rather than market hypotheses.

The company name reinforces the positioning. F2 is “the keystroke you use to audit calculations in an Excel model” when you’re glassy-eyed at 3 AM. “F2 speaks to that core competency. It’s auditability and trust for the financial buyers, for the private credit and private equity associates who are toiling in the bullpen late at night and long weekends.”

Replicating the Super Day Interview Format

Proving AI reliability to institutional investors required rethinking the sales process entirely. Donald reverse-engineered private equity’s talent evaluation methodology.

“Private equity and private credit, talent is the product,” he explains. The stakes are high enough that “associates are being hired, giving offers to join their buy side shop two years before their start date. That’s how competitive the talent market is in private equity and private credit.”

The “super day” format is standardized: “You go to a super day at these funds… literally 18 months before your start date… they give you a laptop and a binder. And the binder has a bunch of docs on a company… and then a computer with no access to the Internet. And your job is to effectively underwrite the business. Come up with an LBO model, an operating model from scratch. You have three hours to complete. You’re on the clock.”

F2 replicates this exactly for AI evaluation. “We now do timed tests under NDA on our customers data. They’re working in a live deal. They’ll send us a data room under NDA for a live deal. They’ll send us sample investment committee memos that they used on different deals. And we will create an AI generated template of that IC memo. We’ll run it against their data room. We’ll work with the F2 agent to clean it up, take it from 80 to 100% and they’ll present it to their team typically same day on a call.”

This approach proves the AI performs at the standard funds use to evaluate human analysts. “That’s the wow factor. That’s how we prove that we’re not a vaporware company. We have real tech, we can deliver real value out of the box at the firm.”

Hiring for Domain Fluency Over Sales Experience

F2’s sales team composition represents a contrarian bet on what drives enterprise adoption in technical markets. “The contrarian bet is there are no salespeople on my sales team. I’ve only hired ex finance industry experts,” Donald reveals.

These aren’t people learning the industry—they lived the exact workflows F2 automates. “These are folks who have sat in the seat, who have lived the pain points that we are solving, who can get on a call with an MD or a principal from a middle market private credit fund and share screen and walk them through their investment memo that they created that same morning and speak to it in the same language as them.”

With “4.5 years of go to market experience selling in the SMB space” from ARK, Donald made the call to teach sales methodology to finance professionals rather than teaching finance expertise to sales professionals. When deals average half a billion dollars flowing through the platform, authenticity and credibility matter more than traditional sales polish.

Transforming Institutional Knowledge Into Competitive Moat

F2’s value proposition extends beyond workflow automation. In Donald’s prior private equity life, producing a screening memo meant “hours parsing through that preliminary data room trying to identify… across thousands of Microsoft share drive folders, precedent transactions to benchmark the deal against and pull down any type of institutional knowledge to inform the underwrite. And then I’d laboriously create a screening memo which would take a dozen hours.”

Today’s workflow achieves 90%+ time savings. But the deeper innovation is contextual intelligence at scale. “You can now pipe in context via API and instantly benchmark all deals that you’ve done historically and drill down by vertical by revenue, sizing by leverage levels, by gross profit margins, by quality of management team to create really crisp benchmarks. Understand what the investment merits were, what the risks and mitigants were, how those vary by industry.”

This transforms isolated point-in-time analysis into context-rich evaluation. “You can get just a much more informed underwrite retaining all of that institutional knowledge and at your fingertips. So not only are you doing existing workflows faster, but they’re actually much better because the AI has complete comprehensive context of the decades of work your firm has done in the past and can use that to inform the net new underwrite on a go forward basis.”

The AI doesn’t just replicate human analysis faster—it accesses comprehensive firm history that individual analysts manually searching SharePoint folders could never efficiently aggregate.

Early Traction and Product Roadmap

F2 publicly launched months ago and is “tracking 10x quarter over quarter growth consistently going into Q1 of next year.” The company reports inflection across multiple metrics: “platform utilization, inflection in weekly active users, inflection in net revenue, growth in retention expansion.”

The initial focus is pre-funding workflows—”due diligence of private markets, investment opportunities. That’s private credit and private equity and commercial banking.” Next comes portfolio monitoring: “You do the deal. Now you got to monitor it on a quarterly basis. We can do that agentically. We can have the first AI native portfolio monitoring solution that’s directly integrated in a unified platform experience with the underwriting product.”

Long-term, Donald envisions infrastructure for deal syndication. “Once we have thousands of enterprise clients leveraging the product, we can start syndicating deals directly on the platform… It’s going from providing the plumbing to do deals more efficiently to actually brokering and syndicating those deals directly on the platform.”

The Repeatable Pattern

F2’s trajectory reveals a specific go-to-market sequence that contradicts conventional startup advice about getting external customer validation early. Instead: solve your own hardest operational problem, build category-leading technology because your business depends on it, then commercialize when customers explicitly request access.

This approach works because organic internal development creates technical depth and defensibility that building for hypothetical external customers rarely produces. When you actually depend on the technology for your own operations, you solve real edge cases and reliability problems that market research doesn’t surface.

The trigger for commercialization was clear and customer-driven: institutional lenders wanted the AI for their entire deal flow, not just marketplace transactions. That signal validated product-market fit before F2 raised external capital or hired a sales team.

For B2B founders, the lesson isn’t to build internal tools hoping they become products. It’s recognizing when your internal solutions solve problems across your entire market—and moving decisively when customers start explicitly requesting access to your infrastructure.