Building Knowledge Graphs for Financial Institutions

AI models can now solve complex math problems and reason through a full diligence process. They work best when the information they need sits in one place, which is rarely the case in a financial institution. Questions like which borrowers had a covenant loosened in their last amendment, which portfolio companies share a supplier, or which loans are impacted by a default elsewhere draw on agreements, holdings data, and other files that rarely reference one another. 

AI finds information in one of two ways. It searches for the passages that read most like the question, or it reads through organized folders of data. Either way, connecting data points that never reference each other starts from scratch on every question, and there is no way to confirm the answer is complete. Even the financial institutions furthest along in adopting AI are running into this limit.

So today, we are launching knowledge graphs in Hebbia. Each firm builds its own system of record for how its structured and unstructured data relates, including the connections several degrees out that no single source may have written down. Because every connection is recorded with the source it came from, different agents and teams reason from the same facts. That shared context not only leads to better decisions but also makes them cheaper to reach. In early evaluations, answers using the graph cost 80% less and were more accurate than a leading frontier model working without one.

In an industry where firms compete on how quickly they act on novel insights, the advantage goes to the ones that connect what they know precisely, and do it first.

Why building for finance is unique

Firms across the deal-making ecosystem, from banks and law firms to investors and advisory teams, need to make these kinds of connections every day. A graph that works for them has to address five specific things:

1. Encoding large volumes of information. A fund’s data rooms, purchase agreements, and portfolio reports can run to millions of pages, well past the point where most AI pipelines lose accuracy or grow too expensive. The graph needs to process every document once and store what it finds, so later questions draw on the whole archive.
2. Interpreting deal terms. A credit agreement sets economic terms, like pricing, and non-economic terms, like EBITDA add-backs. What each one permits often depends on definitions elsewhere in the agreement and on later amendments. The graph has to follow those references and record each term as a fact it can compare across deals.
3. Matching names to the right entity. A sponsor, its fund, and its deal vehicle can each appear under a different name, and private investment vehicles often have no public record. The graph needs to tie each name to the specific legal entity behind it, without merging companies that only look alike.
4. Tracking how facts change over time. Compliance certificates, amendments, and waivers arrive on different schedules, and each can change a fact an earlier document includes. The graph has to keep every version with its effective date, so it knows which applies today and updates only what changed.
5. Controlling access to information. Many professionals may see information only on the deals they work on, because of information barriers or client confidentiality, and once facts are linked, a restricted one can still shape an answer. The graph needs to enforce access on every entity and relationship, and say when access limits an answer instead of filling the gap with a guess.

Hebbia has spent years building information retrieval at scale, the same technology that runs Matrix today, and working inside large financial institutions to learn how they store data and connect their systems. And there is no migration needed. Hebbia connects to your data where it already lives, in SharePoint, Box, Snowflake, SQL tables, or other systems.

More importantly, getting those five things right takes more than software. It takes the judgment industry professionals develop over years of working through deals. Our in-house experts bring that perspective to every graph so firms can tailor how it maps the relationships that matter most to their business. Credit teams asked us for an entity master, banking teams for a deal graph, and legal teams for a contract registry, and a knowledge graph built on that expertise answers all three.

Tracing deal exposure across a firm

Consider the week Silicon Valley Bank (SVB) failed. On March 8, 2023, SVB said it had sold $21 billion of securities at a $1.8 billion loss, depositors tried to pull $42 billion the next day, and on March 10 regulators shut the bank. Every risk desk then had to ask what it was exposed to through SVB, including positions no one would think to check. Regional bank stocks and companies with deposits there were repriced within hours.

The exposure that still mattered sat a few steps out. For example, a fund holding USDC was exposed because Circle, the stablecoin’s issuer, kept $3.3 billion of its reserves at SVB, and by Saturday USDC traded at 87 cents. The fund’s holdings never mention SVB. The digital assets team knew where Circle banks, but on a day when every desk was dealing with its own crisis, chasing that context across the firm wasn’t anyone’s first priority.

A knowledge graph built from the firm’s files would already hold that link, with the source behind it. Because an agent can follow every matching relationship the graph holds, it can start at SVB and trace exposure two or three steps out, to the issuers with reserves there and the funds that hold them. The answer is only as complete as the sources behind the graph, but it no longer depends on who happened to be asked.

The same link points to an opportunity. A fund that knew exactly how much of USDC’s backing sat at SVB could judge whether 87 cents was a fair price, and USDC was back near a dollar within days of regulators guaranteeing SVB’s deposits on March 12. Before, the firm was protecting itself from SVB. With the graph, it can also look for where SVB’s collapse created openings, and act while others are still tracing their exposure.

Building on what the firm already knows

Every firm will have the same AI models. No one else has your firm’s context, from the deals it has done to the counterparties it knows and the clauses that turned out to matter. A knowledge graph turns that context into a shared record every team and agent can trace, so the next time a market moves, your firm sees both its exposure and its opportunities first.

Bring us a question your firm struggles to answer today, and we’ll show you how the graph traces it across your data.