How Agents Cut Hour-Long Data Checks to 5 Minutes - Amol Kapoor & Clifford Ressel, Nori Agentic
Loading slides…
Read the talk
How a cloud agent cut an hour of security master checking to five minutes
At a Fortune 200 asset manager, every changed cell triggered an investigation. Amol Kapoor and Clifford Ressel of Nori Agentic explain why private deals make those checks difficult and how an agent prepares the evidence for the person who owns the decision.
From a talk by Amol Kapoor and Clifford Ressel
At a glance
Ideas worth remembering
Model the relationships among legal entities, share classes and exchange listings before joining incoming feeds. Two correct records can appear to conflict when mapped at the wrong level.
Use the existing day-over-day comparison to trigger one investigation per changed cell. Route the explanation, citations and keep-or-discard recommendation to that cell's owner.
Write source locations, trust rules and review criteria as analyst instructions. Scope the agent's credentials to the records it needs, and measure the human time required to reach a decision.
More of the market sits outside the public exchanges
Amol Kapoor grew up picturing finance as a trading floor and a ringing bell. Preparing for AIE New York, he was struck by how much investment now takes place outside that picture. He cited $240 billion in secondaries volume in 2025, up 50 percent from 2024, and a forecast of $32 trillion in alternative assets under management by 2030. For an engineer at an asset manager, the shift raises a practical question. What records arrive when the firm buys one of these assets?
Kapoor offered several reasons for the shift. Investment in AI extends to the land, energy, data centers and GPU clusters that support it, as well as to secondary shares in private companies such as OpenAI and Anthropic. Founders of high-growth companies can raise a Series F, retain more control and avoid some of the paperwork of going public. Investors, meanwhile, have an incentive to look elsewhere for an edge when well-funded firms already compete for returns in public markets.
Drawing on an observation by Matt Levine, Kapoor described how that movement can reinforce itself. Suppose half the market were in private companies. An investor could buy every public stock and still miss the returns on that half of the economy. As private investment grows, even investors whose aim is simply to participate in the wider market have a reason to follow. Their back offices then have to account for what they bought.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Where do you look up the rights to a plot of land?
As a startup founder, Kapoor already finds it frustrating to track how many shares he owns and what his strike price is. Now consider a firm buying land next to a hydro plant. An analyst needs to know both what the firm paid and what rights it acquired. The purchase terms might be in a PDF in Box, an Outlook email or a Zoom call transcript. Kapoor took the example all the way to a napkin signed in haste at a bar overlooking Columbus Circle.
Those records have to yield information the firm can use to manage, record and audit its investments. Kapoor, Nori Agentic’s CEO, came to the problem after about 12 years in AI engineering, including four and a half at Google. His co-founder and CTO, Clifford Ressel, had worked in back-office finance. Ressel began with public equities to explain what a security master, or SecMaster, already does well and which parts of that design a private deal puts under strain.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
The security master gives risk, accounting and compliance a common starting point
An asset manager needs to know what exists in its investment or tradable universe. The security master is the reference database for the securities and other instruments that represent those assets. Maintaining it brings together engineering, operations and data work. Once an instrument is represented there, different groups can use the same record to model risk, account for profit and loss, report holdings or meet regulatory requirements.
Consider two incoming sources that disagree about a security’s value. The feeds might report different values for the same instrument, or each might be correct about a different instrument that the firm’s code has mapped to the same record. Ressel wants that conflict flagged before either value enters the security master. The team can then trace the discrepancy back to the feeds and check both the reported values and the mapping before downstream analyses rely on the record.
Ressel called this an organizational primitive because it gives the firm one place to enforce standards, make discrepancies visible and check quality. A medallion data lakehouse with many tables can serve that function, as can a single denormalized table. The organizational benefit is the same in either design. The people producing data have somewhere to resolve a mapping once, and each team consuming it can start from the resulting record.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Mapping Google from legal entity to share class to exchange
A firm's internal view has to accommodate counterparties that speak different languages, settle in different currencies and operate under different regulatory regimes. Public markets offer several points around which to organize that work. Vendors provide identifiers used for market data and transactions, including Bloomberg FIGIs and RICs. Regulatory and numbering systems supply identifiers such as ISINs and CUSIPs. With the relevant mappings in place, an asset manager can connect an outside source to its own positions for risk modeling, holdings reports, alternative-data joins or financial models.
Ressel used Google to show why the identifier has to match the question. GOOG and GOOGL distinguish two share classes traded on the same exchange. A dataset about the legal entity, however, describes the company as a whole. If a holdings table is indexed by ISIN, the analyst needs a mapping from the security to its legal entity to join that data. The same company-level record can then serve a holding in either share class.
The mapping has to work in the other direction, too. The manager might hold just one share class but need to trade it on several exchanges for different funds or investors. Ressel illustrated the result with a denormalized table containing several rows for what an investor might casually call one Google equity. Once those relationships are represented, code can place an incoming record even when the source supplies only a subset of identifiers or describes the asset at a different level of detail.
Ressel compared the experience to using a date-time library. A good implementation makes the operation look simple. Modifying it, or working in a language that lacks one, exposes the edge cases its authors had to resolve. A well-mapped security master offers coding agents the same advantage. They can write code against the firm's established definitions. The difficulty changes when the incoming evidence lacks a regular identifier to map in the first place.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
For a private loan, the agreement defines the instrument
Imagine lending money to a company building a power plant near proposed data centers, or acquiring nearby land for future development. The parties enter the deal without a shared set of identifiers for everything they might negotiate. No single regulator prescribes the full structure of the information that will emerge. Nor can the analyst call one vendor API whose data model accounts for every possible arrangement.
For the loan, the repayment schedule and collateralization live in the one-off agreement and its amendments. Those terms define what the firm owns. A record for a different loan to a similar project cannot stand in for them. With Google, mapping a share class pays off again when another feed uses a known identifier. With private investments, Ressel observed, the documentation to understand grows almost one-to-one with the assets. Adding a deal means doing much of the work of reading and mapping again.
Kapoor picked up the consequence for the risk team. Its models still need standardized inputs, and the firm still has to clean and audit the data. Asking each modeler to search PDFs and decide which terms are authoritative would undo the separation the security master was built to provide. Someone with access to the records and enough domain knowledge to interpret them has to get the relevant information back into the master.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
The daily comparison worked. Each changed cell still cost an hour.
Nori worked with a Fortune 200 asset manager in New York that had begun holding more private assets. Its data engineers already had a detection mechanism. They compared the security master table with the previous day's version and sent each changed cell to a triage team. Kapoor illustrated the scale with 50,000 assets and roughly 200 columns, about ten million cells. Even when any particular field changes infrequently, the team can expect a queue every day.
A difference between two snapshots establishes that a value moved. It leaves open why. An asset might have changed price or name, a company might have gone bankrupt, ownership might have transferred, or the parties might have renegotiated a contract. An upstream provider could also have mistyped a value, misunderstood an event or supplied information that conflicted with another source. The comparison flags all of those cases in the same way.
The unit of review matters for the economics. A reviewer had to establish whether to accept each changed cell, even when several belonged to the same asset. That meant locating documents, consulting other sources and working out whether the evidence supported the new value. Kapoor put the average at an hour of someone's time per cell. He corrected his estimate of the aggregate burden to hundreds or thousands of hours a month.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
From a detected change to a cited recommendation in Slack
Kapoor described a cloud agent as a coding agent, comparable to Claude Code or ChatGPT, running in a managed environment away from the analyst's laptop. That gives the operator a place to impose audit requirements and observe the work. It also lets the agent start on a schedule or in response to an event without waiting for someone to open a session. For this customer, the underlying model was open source and deployed locally so the firm's data stayed on premises.
The agent starts when a change is detected in the investment universe. Its assignment is specific to the cell that changed. It must work out the source of that change. The tools available for the investigation include Box for PDFs, data APIs including Bloomberg, and a limited portion of email. An analyst who previously had to move among those systems can now begin with the results of that search.
For each cell, the agent assembles an explanation of what happened, supplies citations for the explanation and recommends keeping or discarding the change. It then contacts the person responsible for that cell in Slack. The routing matters as much as the search. The person with authority to decide receives a case they can inspect, with the evidence and the proposed disposition together.
The owner can accept or reject the recommendation or investigate further if the cited material leaves a question open. Kapoor reported that an investigation that formerly took about an hour now required about five minutes of human work. The comparison measures the reviewer's time. Agent runtime and recommendation accuracy were unreported. The firm retains a person accountable for the decision while moving much of the document search and case preparation off that person's desk.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Give the agent the instructions a new analyst would need
Connecting Box and Bloomberg still leaves a judgment problem. Where does this firm keep the relevant documents? If sources disagree, which does it trust? What makes a change acceptable? Kapoor's approach was to write down the guidance a new analyst would need, in English, in Markdown files. Someone who understands the review process can change the agent's behavior by editing those instructions.
Kapoor described starting with a general-purpose coding agent such as Claude Code or OpenCode and configuring it with skills or agent instruction files. Much of the specialization comes from the environment in which it runs. The configuration supplies the job and the rules, while the credentials on the machine determine which records it can consult. In the asset manager's deployment, for example, email access covers a small segment. Instructions about looking for correspondence operate within that boundary.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Let the data change initiate the work
Kapoor likened the agents around a database to cleaner fish, continually finding issues and surfacing them to someone who can act. The difference for the analyst is visible at the start of the day. In a prompt-driven workflow, someone still has to inspect the differences, decide which ones deserve attention and dispatch the investigations. Here the existing change signal starts the work. The first thing the owner receives is a recommendation ready for review.
Kapoor argued that the same infrastructure could support audit work, code review and engineers' day-to-day tasks. Each job needs its own instructions, tools, triggers and company-owned context. The orchestration that starts a run, the security controls around it and the integrations it uses can be shared. His preference was for a vendor to own that common infrastructure so separate teams at one firm needn't solve the same operational problems repeatedly.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Three lessons about adopting agents in back-office finance
First, Kapoor sees adoption as a collaboration problem. Buyers want a solution to a job they already have, while engineers tend to offer building blocks. Nori found that domain experts could make the connection once they understood what a cloud agent could do. The security master team knew which event opened a case, who owned the decision and where the hour went. That knowledge gave the engineers a specific job to automate.
Second, Kapoor proposed aiming for 95 percent automation and designing for human review. For the security master, he put the change in human effort at about an hour per cell down to five minutes. In his experience, expecting an agent to complete every case unaided limits the set of problems a team will attempt, in software factories as well as in finance. Allowing a person five minutes to examine the evidence makes an investigation worth automating even when a difficult case still needs judgment.
Third, Kapoor drew on Fred Brooks's concept of accidental complexity, the difficulty created by how a system came to be built. Two teams record the same information differently. Vendors transmit it in different formats. Someone spends an hour reconciling those conventions because the systems evolved separately. Kapoor sees that combination of low complexity and high effort as particularly suitable for agents. Once the evidence is assembled, the analyst has time for the genuinely difficult question of what the records establish about the firm's investment.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Full transcript
Loading transcript…