In regulated AI, the ontology may matter more than the chatbot. Databricks' Genie One for funds shows why
Databricks describes agents that reconcile positions and prepare NAV overnight, grounded in a governed business ontology with a trace behind every number. The design lesson outlasts the product.
Listen to this article · 6 min
AI-generated narration of the full article.
In fund administration, a wrong number is not an embarrassment. It is a mispriced subscription, a fee charged on the wrong base, an investor letter that has to be corrected and a conversation with the regulator. That is why most asset managers have watched the generative AI wave with interest and kept it far away from NAV.
Databricks published a reference picture on October 2 of what it would look like to bring agents into that work. It is a vendor blog, not a customer case, and it should be read that way. But the architecture it describes is the right one for any high-consequence workflow, and it puts the emphasis in an unusual place: not on the assistant, but on the meaning of the data.
What Databricks describes
Genie One is presented as a “data-smart AI coworker” for asset management finance teams. The design has three layers:
- A governed business ontology (Genie Ontology) that captures what things mean in fund operations: definitions, strategies, fee drivers, and how they relate. Answers are grounded in that ontology rather than in whatever documents happen to be retrievable.
- Backend agents (Genie Agents) that run overnight: ingesting custodian files through Lakeflow, reconciling positions, pricing them against market data feeds and producing a preliminary NAV before the morning review. Performance fee tracking is part of the same flow.
- Traceable answers. Every NAV figure is meant to be traceable back to its pricing sources, accrual entries and custodian comparisons, with Databricks describing the goal as “audit-ready proof for every valuation and fee number.”
The post places this inside the obligations asset managers live with: SEC liquidity risk management requirements, Form PF reporting, GIPS-verified performance and, for European managers, AIFMD.
What the post does not include is a named customer, production metrics or a formal availability statement for Genie One. Treat it as a reference architecture and a product direction, not as evidence that someone has run NAV this way at scale.
Why the ontology is the control
- Ingest custodian files
- Reconcile positions
- Price against market feeds
- Deterministic checks and tolerances
- Preliminary NAV with trace
- Controller reviews and signs
- Overnight, grounded in the ontology: Ingest custodian files · Reconcile positions · Price against market feeds · Deterministic checks and tolerances · Preliminary NAV with trace
- Morning: Controller reviews and signs
Agent workDeterministic control
Most enterprise AI failures in finance are not model failures. They are definition failures. “Net assets” means one thing in the fund accounting system, another in the risk system and a third in the investor report. An assistant that retrieves documents will happily answer with whichever definition it found first. An assistant grounded in a governed ontology answers with the one the business agreed on, and can show which one it used.
That makes the ontology a control, not documentation:
- It constrains what the agent can mean. A fee calculation refers to a defined fee driver, not to a phrase in a PDF.
- It makes answers checkable. If the definition is explicit, a reviewer can verify the answer against it.
- It survives model changes. Swap the model and the definitions stay; that matters when models get retired on someone else’s calendar.
This is the same principle we highlighted in Snowflake’s AI accounts payable design: keep the deterministic parts deterministic. In fund finance, pricing rules, tolerances and reconciliation logic should be code and reference data, with agents doing the gathering, matching and explaining around them.
What it means for regulated teams
The audit trail is the product. In a U.S. fund complex, liquidity classification under Rule 22e-4 and Form PF filings both depend on numbers someone must be able to defend. If an agent touched those numbers, the trace has to show what it read, what it matched and what a person approved.
Semantics are a shared asset. A business ontology for fund operations is useful far beyond one assistant: reporting, risk, investor relations and every future agent. We argued in our piece on the enterprise AI harness that business definitions belong in the layer you own once, across vendors. Whether you build it in Databricks, Microsoft Fabric or elsewhere, own it.
Humans sign; agents prepare. The overnight pattern works because the morning review remains. An agent that produces a preliminary NAV for a controller to review is a productivity tool. An agent that publishes NAV is a control failure waiting to happen.
What to do now
- Start with the definitions, not the assistant: list the 30 to 50 terms that drive valuation and fees and get finance, risk and operations to agree on them.
- Separate deterministic logic from agent work in every workflow you automate. Pricing rules and tolerances stay as code.
- Design the trace before the interface. For every number an agent produces, decide what evidence a reviewer will need to see.
- Keep the sign-off where it is. Agents move the preparation overnight; the controller still approves in the morning.
- Ask vendors for proof, not posture: production references, error rates on reconciliation breaks, and how the trace is retained.
The bottom line
The interesting part of Genie One is not that an AI can answer questions about a fund. It is the claim that the answer should be grounded in agreed business meaning and backed by a trace to source. That is what regulated finance requires of any system, AI or not. Teams that invest in the ontology and the trace first will find the assistant is the easy part.