How to actually build AI for finance

“AI” has become the most overused label in enterprise software. Every week a new startup announces it has cracked the code by wrapping ChatGPT in a glossy UI or by automating some Excel workflow. It makes for slick demos and snappy press releases, but it doesn’t actually change how financial firms make decisions. These tools are surface-level convenience features pretending to be intelligence. They don’t embed themselves in the hard machinery of capital allocation. They don’t compound.

So, how do you build meaningful AI tools for finance professionals?

Stage 1: Build the ontology

Every investor operates with a mental model of the world. They break it down into attributes that become the scaffolding for judgment. If your software doesn’t have this ontology in its DNA, it’s blind. The first step in building AI tools for finance professionals is not training a model. It’s building the system of entities and relationships that defines the investment universe.

Imagine a private equity team evaluating a growth buyout in a mid-market software company. The data room is a swamp of PDFs, Excel models, and scattered slides. The first question isn’t “Can AI summarize this?” The question is “What’s the ontology that makes this mess interpretable?”

The firm’s ontology defines key entities - company, product line, cap table, working capital facilities, supplier network - and the relationships that matter between them. When an analyst uploads a data room into the firm’s diligence platform, the system doesn’t just spit out a PDF summary. Instead, it parses every line item into the firm’s ontology, stitching them into a structured graph.

Now, the team can benchmark margin expansion against peers, stress-test downside scenarios, and trace how every assumption connects to a decision variable. But more importantly, the ontology reflects the firm’s own worldview. If a fund has its own sector classification, such as carving out “frontier healthcare” or “critical infrastructure software” as proprietary categories, that is what the system encodes. If the firm measures risk through a custom metric such as founder dependency or supply-chain fragility, that too becomes part of the ontology. The ontology doesn’t just unify information; it becomes your language for making sense of investment opportunities. Without it, you’re stapling autocomplete onto chaos.

Stage 2: Encode the frameworks

Pattern recognition is what separates the great shot callers from the rest of the pack. Investors learn by doing deals, winning some and losing others, and building frameworks that guide judgment. They are not oracles. They are trained operators who know what good looks like and what bad smells like.

With its ontology in place, the firm can encode the frameworks it has refined over dozens of deals: resilient recurring revenue, disciplined leverage, and management teams with operational grit. They’ve also learned the scars of failed bets - overpaying for growth, underestimating competition, and ignoring integration risks.

Encoding those frameworks into AI means more than training on generic benchmarks. It means digesting the firm’s own memos, analysis, and outcomes, and distilling the factors that have made certain investments successful and others less so.

When the system analyzes a deal, it doesn’t just spit out “EBITDA margin looks poor.” It flags that margin expansion sits one standard deviation below the firm’s historical winners, that management’s integration plan resembles past failures, and that the leverage profile would have triggered covenant breaches similar to deals that went sideways.

The AI here is not naive. It’s a digital twin of the firm’s institutional memory, compact and searchable, and is always available to every partner and associate.

Stage 3: Re-optimization

The real power of AI isn’t when it just reflects past frameworks. It is when it improves upon them. This is the stage of re-optimization. The AI doesn’t just speed up today’s workflow; it increases the probability of tomorrow’s success.

A good system doesn’t conclude its analysis once the deal closes. It looks at the anti-portfolio - the competitor SaaS business the firm passed on that later raised at a 5x markup - and asks what signals were missed. It runs projections on the acquired company’s revenue trajectories under different market scenarios. It tests whether similar peers who took on higher leverage outperformed or collapsed.

Over time, the AI measures the firm’s frameworks against real outcomes, sharpening them with every cycle. This process mirrors reinforcement learning with human feedback (RLHF) in large language models. Just as LLMs improve when human evaluators reward good outputs and penalize bad ones, investment AI improves when real-world deal outcomes form a feedback loop that provides reinforcement. Passing on a deal that later outperforms is negative feedback. Spotting risks that end up materializing is positive feedback. The system uses these signals to tune its internal decision-weighting, effectively “reward modeling” the frameworks of a firm’s best investors.

Re-optimization is how AI compounds. It doesn’t just accelerate workflows—it continuously learns from outcomes, encoding not only what the firm believes but what reality has proven. Each cycle makes the ontology sharper, the frameworks smarter, and the decisions more precise.

The Opto Approach

At Opto, we don’t believe in building wrappers or workflow hacks. We embed the ontology of investing directly into the product, incepting firm-specific frameworks, and continuously re-optimizing with real-world feedback. That’s why our products are not chatbots perched on top of Excel, but a system of record and intelligent tools for private markets investing.

See applicable disclosures at https://www.optoinvest.com/disclaimers.