Private assets make security-master data difficult because their terms and identifiers live in documents rather than shared systems.
2
A cloud agent can investigate every changed cell, cite its sources, recommend keeping or discarding the change, and send the owner a Slack review task.
3
The right target is about 95% automation with a human reviewing the result, rather than fully autonomous finance operations.
Summary
Amol Kapoor and Clifford Ressel explain why private assets create a difficult data-management problem for asset managers. A security master is the database that defines an investment universe and feeds risk, P&L, accounting, compliance, and other analysis. Public assets have shared identifiers and data vendors, while private assets are defined by one-off agreements, amendments, PDFs, emails, and call records. At a Fortune 200 asset manager, every changed security-master cell required about an hour of manual research. Nori Agentic uses a cloud agent that runs when data changes, searches approved sources such as Box, Bloomberg APIs, and selected email, explains the change with citations, recommends keeping or discarding it, and asks the cell owner to review it in Slack. The human still makes the final decision, reducing their work to about five minutes. Kapoor closes with lessons about collaboration, 95% automation, and reducing back-office complexity.
Private assets have pushed investment data beyond shared identifiers
Kapoor says finance now holds more private assets, including land, energy, data centers, GPU clusters, and private-company secondaries. The information needed to manage them can be scattered across PDFs, Box folders, Outlook email, Zoom transcripts, or even signed paper. A private investment may include repayment terms, collateral, and ownership rights that exist only in one-off agreements and amendments. There is no universal identifier system or single data vendor that captures the asset's full structure. The data burden therefore grows with the number of assets, while the security master becomes harder to maintain.
A security master is the source of truth for an asset manager
Ressel describes the security master as the master database of the securities or instruments in an asset manager's investment universe. It feeds risk management, P&L accounting, regulatory compliance, holdings reports, and other business-critical analysis. When two sources disagree, the conflict should be resolved before it reaches this database, so an upstream error does not spread into downstream work. Ressel compares the system to an organizational data layer that separates producers from consumers and enforces standards, visibility, and quality checks.
Even public equities require several mappings for one asset
Ressel uses Google's public shares to show why security-master work is already complicated before private assets appear. Google has two share classes, GOOG and GOOGL, and data may arrive at the legal-entity level rather than the individual share-class level. An asset manager may also trade one share class across several exchanges for different funds or investors. Public markets provide identifiers such as Bloomberg FIGIs, LSE RICs, ISINs, and CUSIPs, which help map outside data into the firm's own view. Unstructured private-asset data removes that common language.
Changed security-master cells created a large manual research queue
At a Fortune 200 asset manager, the team compared the security-master table day over day. The table contained 50,000 assets with roughly 200 columns per asset, so even infrequent changes produced a steady stream of changed cells. Each cell was sent to a team that searched documents and other sources to decide whether the update was valid. A change could involve a price, name, ownership, bankruptcy status, or contract. Kapoor says each change used to require about an hour of someone's time, creating hundreds or thousands of hours of work each month.
A cloud agent turns each data change into a cited review task
Nori's cloud agent runs remotely when new data reaches the security master. In the example, it uses a locally deployed open-source model so the underlying data remains on premises. The agent has access to Box for PDFs, public APIs such as Bloomberg, and a limited portion of email. It traces the source of a changed cell, assembles an explanation, adds citations, and recommends keeping or discarding the update. It then sends the cell owner a Slack message. The owner still reviews the recommendation and can research further when needed.
Plain-English instructions provide the agent's domain knowledge
Kapoor says teams do not need to build a separate custom agent for every narrow finance task. The agent can be given written instructions describing where documents live, which sources to trust, and how to judge a change. Those instructions can be edited in natural-language Markdown files. The underlying agent can be a general tool such as Claude Code, Open Code, or Pi, with skills, agent files, credentials, and other configuration supplied by its environment. The specialized behavior comes mainly from the instructions, tools, and access available to that environment.
Enterprise agents should bring work to people when it is ready
Kapoor frames the agentic security master as a database with automated workers that spot problems, surface them, and sometimes resolve them. He extends the same pattern beyond data cleaning. Instead of asking a person to decide whether a task deserves an agent, the system can proactively send people tasks, advice, analysis, and recommendations. He argues that a security-master agent, audit agent, code-review agent, and engineering agents should share cloud infrastructure for orchestration, security, and integrations. Their instructions, tools, triggers, and context can differ while the underlying platform stays common.
The practical target is 95% automation with human accountability
Kapoor's first adoption lesson is that AI projects need collaboration between people who understand a domain and engineers who build reusable systems. His second is to aim for 95% automation and keep a human in the loop. Finance still needs accountability, and a person spending five minutes on a recommendation can solve more problems than a system forced to act autonomously every time. His final lesson concerns accidental complexity: agents can handle low-complexity, high-effort work caused by inconsistent formats and organically evolved processes, leaving people to work on genuinely difficult problems.
"Agents get information the same way a new analyst does, from written instructions about where documents live, which sources to trust, and how to judge a change, all in English."Amol Kapoor14:47
Who should watch
You maintain a security master or other investment database where changes arrive from PDFs, emails, contracts, or unstructured vendor data.
Your finance or operations team spends time checking whether individual data changes are valid and tracing them back to source documents.
You are evaluating agent systems for regulated workflows and need a practical pattern that keeps a human accountable for the final decision.