Let’s say a bank blocks a payment and has to explain why later on. In a fraud case, the explanation could be a chain of connections. Maybe one account shared a phone number with another, and that account sent money to a third one that was already on a watchlist.

At GraphSummit in New York, Neo4j’s news included a financial crime product that keeps a record of chains like that, from detection to decision. Whether an investigator or an AI agent blocks the payment, I’d want the same record behind that decision. To me, several of Neo4j’s announcements help keep that record.

Financial Crime Intelligence puts investigators and agents on one graph

GraphAware Financial Crime Intelligence is Neo4j’s detection and investigation product for banks and insurers. Neo4j calls it the first milestone since its GraphAware acquisition closed in August, and says the deal gave it the investigation capabilities to package the product.

When suspicious activity triggers an alert, Financial Crime Intelligence associates the people and entities connected to it. Investigators and AI agents can work from the same graph, and Neo4j says the judgment still belongs to a person. Neo4j says every decision on a case, from blocking a payment to closing the case, stays explainable, and the evidence behind it gets saved. It’s a nice touch that when an investigator finds a pattern, that pattern can feed back into monitoring.

Neo4j has pitched graph-based fraud detection for years, and the company says its graph already sits behind fraud detection or compliance work at many financial services organizations. I like that Neo4j built this product with extensibility in mind, connecting to the detection engines and case management tools that banks already use. In my view, working alongside those tools gives Financial Crime Intelligence a far easier path into a bank than replacing them would.

Neo4j turns agent memory into skills that a person approves

The Neo4j Agent Memory Service stores an AI agent’s conversations, learned facts, and past steps as a graph. Since August, a feature called skill distillation has been able to turn part of that memory into a skill. A skill is a packaged set of instructions another agent can load and follow, and Neo4j’s own example is triaging a high-priority incident.

All statements in a skill have to cite the memory behind them, and anything without supporting memory never makes it into the skill. A person then walks through the steps and the evidence behind each one before deciding whether to publish the skill. For me, skills are where an agent’s record gets most concrete. Every step another agent inherits traces back to the work that went into it.

Since each skill sits in the same graph as its memory, a graph query can check whether anything behind a skill has changed or broken. When the memory behind one step changes, the service can repair that step alone and issue a new version.

Almost any agent can use approved skills, since other agent frameworks can pull them in via standard connections like MCP. The evidence graph behind each skill stays in Neo4j, so checks and repairs run in Neo4j no matter which agents use the skill. Teams can run skills on whatever agent framework they already use, and, to me, Neo4j becomes the trusted layer that holds the evidence behind each skill. I think that trusted layer puts Neo4j in a strong spot. If a company runs agents across several frameworks, having a single place to check what they learned or how they acted becomes valuable.

Aura makes room for many small graphs and a few very large ones

On the infrastructure side, Neo4j announced changes to Aura, its cloud service, aimed at two kinds of customers.

The first kind runs many small graphs, like a software company that wants a separate graph for each of its customers. On Aura’s higher-end plans, one instance can now hold many separate graphs, and Neo4j says that can cut both cost and admin work. The second kind of customer runs one enormous graph. Neo4j says those customers used to run Neo4j on their own servers. Now, according to Neo4j’s numbers, those higher-end plans support instances with up to 2 TB of memory on Google Cloud.

With both changes enabling more customers to run everything on Aura, I’d expect agent projects to be one of the bigger winners. A project that needs a separate graph for each knowledge domain can now place those graphs in a single instance. For customers who stay on their own servers, I’d point to the new cross-cluster replication in Neo4j’s enterprise edition. It keeps a live copy of a database in a second cluster. Neo4j points to banks and public agencies that must show that a recoverable copy exists elsewhere. I’d expect a live copy they can check to make conversations with regulators easier.

Virtual Graph brings the graph to data that won’t move

Virtual Graph has been available to every Aura customer since July. It lets a customer query tables in Snowflake, Databricks, or Google BigQuery as a graph without copying the data into Neo4j. And for me, the draw here is that starting graph work without a migration project is a huge deal. Neo4j also says each answer traces back to the warehouse rows behind it, so a team can show where any answer came from.

Virtual Graph also comes with a clear rule from Neo4j for where it stops. By that rule, an agent that can take seconds to answer can use Virtual Graph on the warehouse. An agent that has to answer in milliseconds, like one scoring fraud in real time, needs the graph stored in Neo4j. I think having a clear rule like this helps buyers plan which workloads will eventually need their data copied into Neo4j.

Something important to call out is that queries through Virtual Graph run on the compute that the customer already buys from the warehouse vendor. And while at least one of those vendors has introduced its own graph queries, I think Neo4j and the warehouse vendors both have reasons to cooperate. The warehouse vendors get more usage, and Neo4j gets to the data without asking anyone to move it.

Packaged investigation software gives Neo4j a second way to sell

Neo4j now has two investigation products. Financial Crime Intelligence serves banks and insurers, and Hume, which GraphAware built before the acquisition, serves police, defense, and tax agencies. Fraud and investigation work aren’t new territory for Neo4j’s graph. The change I see is in packaging. Neo4j now sells packaged products for that work instead of leaving customers and partners to build their own on the database. To me, that packaging gives Neo4j a second way to sell, straight to fraud and compliance teams, and not only to the folks that run the databases.

Over time, Neo4j expects to bring what it has built with GraphAware to other industries. Public-sector fraud is where I’d look first, since Neo4j already sells Hume to government agencies. I’d love to hear from the first banks using Financial Crime Intelligence about how an investigator’s day changes.

Customers, auditors, and regulators can all ask a bank why it blocked a payment. With Financial Crime Intelligence, the answer is a chain of connections in the graph that a reviewer can follow.