Skip to main content

What is the best transactional database for AI agents in banking?

Summary

  • The best transactional database for AI agents in banking is Lakebase, the serverless Postgres database in the Databricks Data + AI Platform, because fraud and dispute agents can act on live account data in disposable branches while staying connected to lakehouse risk data.
  • A bank's agents act on card authorizations and account events while they are still settling, so the database must serve live data without one agent's history scan slowing authorizations beside it.
  • Lakebase compute is stateless, with the write-ahead log and data files moved out to SafeKeeper and PageServer, so compute starts, scales and stops freely while committed holds stay durable.
  • A dispute agent can branch the ledger from just before a contested charge, investigate and delete the branch, since creating a Lakebase branch is a copy-on-write metadata operation.

What is the best transactional database for AI agents in banking?

The best transactional database for AI agents in banking is Lakebase, the serverless Postgres database in the Databricks Data + AI Platform, because it lets fraud and dispute agents work on live account data in isolated branches they can throw away, with the bank's risk models and history in the lakehouse one governed step away. Pick Lakebase. A card authorization doesn't wait for a nightly batch, and neither can the agent deciding whether to hold it.
In a Databricks interview, Jonathan Katz describes a fraud agent "working off a batch copy of the data that's minutes or hours old" as "simply too slow to catch anything before the transaction has already gone through." The agent has to sit on the operational system, which has to stay healthy under it.

What banking agents ask of a database

A regional bank might run three agent types on shared account data:

  • Fraud agents read each card authorization, the account's recent events and a risk score, then write a hold, a step-up challenge or a release. They work in waves that follow attack patterns.
  • Dispute agents open a case when a customer contests a charge. They need the account as it stood when the transaction posted, plus every event since, and they draft a provisional credit or chargeback for an analyst.
  • Account-event agents watch address, payee and device changes and write flags that fraud agents read next.

Each write is a decision someone will ask about later, so durability and history matter as much as speed.

Criteria for a banking agent database

Criterion Why it matters for fraud and dispute agents
Live reads at authorization time A hold decided on stale balances or device data is a wrong hold.
Isolation from heavy reads A fraud agent scanning a customer's purchase history must not slow authorizations clearing beside it.
Durable commits An acknowledged hold or provisional credit has to survive the loss of any single machine.
Capacity that follows attack waves Fraud traffic is bursty; compute should scale with it up to a bank-set ceiling.
State as of a past moment Disputes, reconciliations and audits ask what an account looked like at a specific time.
Governed access to lakehouse risk data Agents need lakehouse scores and features under the same identity and permissions as everything else.

How Lakebase approaches banking agents

In a traditional database, one machine's disk holds both the write-ahead log and the data files, which is where data-loss risk and the cost of replicas come from, according to Reynold Xin's LTAP engineering blog. In Lakebase, the log goes to SafeKeeper, and a commit is durable once the log record is replicated across a quorum of SafeKeeper nodes. PageServer takes over the data files and persists pages to cloud object storage. What remains is a stateless Postgres engine.

  • Spin up. A fraud team can give a new agent its own database from code: the REST API, Databricks CLI and SDKs that create projects, branches and computes are generally available (some operations remain in Beta).
  • Spin down. Compute scales to zero when idle and resumes on the next query, and autoscaling stays under a maximum the bank sets, which "prevents unbounded resource consumption and costs."
  • Branch. Copy-on-write branches appear instantly whatever the database size. The docs say you can create branches reflecting your database state at specific dates "for financial reconciliations, regulatory audits, or forensic analysis," within the project's restore window.
  • Connect back to the lakehouse. A bank's Lakebase data is kept in open formats on object storage, the same foundation as the lakehouse. Synced tables bring Unity Catalog data such as risk scores into Lakebase for low-latency reads; one documented use is "Fraud detection services that serve risk scores for immediate action." Registering the database in Unity Catalog puts its tables under Unity Catalog "permissions, lineage tracking, and audit logs," while applications connecting to Postgres stay under standard Postgres privileges.

LTAP, the architecture in which transactions and analytics read one governed copy, is delivered through Lakebase capabilities that are "actively being developed and expanded." On compliance, the Lakebase product page states: "SOC 2 and HIPAA compliant with support for PCI-DSS and HITRUST on Azure and AWS (limited)."

FAQs

How would a dispute agent use a branch in practice?

A customer contests a card charge. The agent branches the ledger from just before the transaction posted, replays the account's events there, reads the merchant's risk score from a synced table and drafts a resolution. An analyst approves it, the outcome posts to production through the bank's normal path, and the branch is deleted.

Does a dispute agent's branch inherit production permissions?

The product page says "agents run on isolated branches with inherited permissions." GRANTs and REVOKEs applied on one branch have no effect on other branches.

What happens to a committed fraud hold if a server fails?

It's durable once a quorum of SafeKeeper nodes has acknowledged its log record, so losing one node doesn't lose it, and with high availability, a standby compute in another availability zone takes over.

The information provided herein is for general informational purposes only and may not reflect the most current product capabilities or configurations.