# How Agents Cut Hour-Long Data Checks to 5 Minutes - Amol Kapoor & Clifford Ressel, Nori Agentic

Amol Kapoor, Nori Agentic · Clifford Ressel, Nori Agentic

[Talk page](https://ai.engineer/talks/kIlVNGZZr8o-agents-cut-hour-long-data-checks-5) · [Video](https://www.youtube.com/watch?v=kIlVNGZZr8o)

## 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.

### More of the market sits outside the public exchanges

[0:12](https://www.youtube.com/watch?v=kIlVNGZZr8o&t=12s)

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.

### Where do you look up the rights to a plot of land?

[2:43](https://www.youtube.com/watch?v=kIlVNGZZr8o&t=163s)

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.

### The security master gives risk, accounting and compliance a common starting point

[4:19](https://www.youtube.com/watch?v=kIlVNGZZr8o&t=259s)

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.

![The custodian and fund admin both report 12,500 shares for INT-000482, but their highlighted prices differ, 41.02 versus 41.20. The secmaster holds the price as pending.](https://ai.engineer/api/talk-essay-media/e9ed4a5f9b43c7a60f9532442b9ddeb9ebc9d2ab080fa548e5eb7d3f873ce585)

Data quality within the secmaster is the first line of defense

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.

### Mapping Google from legal entity to share class to exchange

[6:06](https://www.youtube.com/watch?v=kIlVNGZZr8o&t=366s)

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.

![An Alphabet security master lists share classes by identifier and trading venue. Green highlights connect an alternate feed containing Alphabet A, GOOGL, and USD to the corresponding fields in the table.](https://ai.engineer/api/talk-essay-media/d71415c7872fb14d87428dbccdede60c3b570b4957488c7079931a00acd51e59)

Public data maps cleanly after centralization enforces the identifiers

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.

### For a private loan, the agreement defines the instrument

[8:58](https://www.youtube.com/watch?v=kIlVNGZZr8o&t=538s)

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.

![A terrain model shows a hydro plant on a river, surrounded by its credit agreement, amendment, collateral schedule, and drawdown notice. A dashed line toward an API is marked “× no feed.”](https://ai.engineer/api/talk-essay-media/044c675b3c10cf40497adcc4dbfc6857604b07ce4db1adb810f0f924a30e4722)

Private assets have no standard identifiers and no data feeds

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.

### The daily comparison worked. Each changed cell still cost an hour.

[10:49](https://www.youtube.com/watch?v=kIlVNGZZr8o&t=649s)

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.

![Yellow points mark 50 changed cells on day 2 in a grid spanning 50,000 assets and 200 columns.](https://ai.engineer/api/talk-essay-media/9868836177800fc771824a4ede2795ac88623f5331564d6d960429a6b37d4c3b)

Day-over-day changes in a 10M-cell security master

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.

### From a detected change to a cited recommendation in Slack

[12:29](https://www.youtube.com/watch?v=kIlVNGZZr8o&t=749s)

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.

![A Slack message recommends keeping the change from Holdco A to Holdco B and shows three citations, Keep and Discard buttons, and approval by J. Rivera in 4 min. A bar chart compares resolution times of ~1 hr before and ~5 min after.](https://ai.engineer/api/talk-essay-media/c6704d572b0dd60037998822743dc3263022d19d5388379d0c7c6bd4fdf15a30)

Human review of Parcel 7’s ownership change

### Give the agent the instructions a new analyst would need

[14:28](https://www.youtube.com/watch?v=kIlVNGZZr8o&t=868s)

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.

![An AGENTS.md file lists four triage rules. The final rule, “Ignore draft PDFs,” is highlighted in green. A dotted path connects the file to an agent and a reviewer.](https://ai.engineer/api/talk-essay-media/bbbd3deda530ed1fa12d413c97dd3639f655914451438bc4ec09d3cf24e343c7)

AGENTS.md instructions for secmaster triage

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.

### Let the data change initiate the work

[15:28](https://www.youtube.com/watch?v=kIlVNGZZr8o&t=928s)

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.

### Three lessons about adopting agents in back-office finance

[17:15](https://www.youtube.com/watch?v=kIlVNGZZr8o&t=1035s)

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.

## Key takeaways

- 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.

## Slides

### Private is the new public — secondaries volume and AI infrastructure

[0:31](https://www.youtube.com/watch?v=kIlVNGZZr8o&t=31s) [Two figures highlight secondaries volume and forecast alternatives AUM. Beside them, an isometric diagram links capital to land, energy, data centers, GPUs and secondaries.](https://assets.ai.engineer/sites/aiecode2025/sha256/d5b730f179fdeb0c2830b96855b27facf0c57b522079528584b57d0550d0cbc0.webp)

Private is the new public

When I was growing up, finance mostly meant public markets. Trading floors, ringing bells, things like that. But that's not really the case anymore. There has been a huge influx of capital investment into private assets.

$240B    secondaries volume in 2025, up 48%
$32T    forecast alternatives AUM by 2030

Some of this has been because AI seems to be the only growth area in the broader economy. So more people are holding onto land, energy, data centers, GPU clusters.

And of course secondaries. So many secondaries. OpenAI and Anthropic secondaries everywhere.

src · Jefferies · Preqin

capital
energy
data centers
land
secondaries
GPUs

### Private is the new public — raising a Series F

[1:52](https://www.youtube.com/watch?v=kIlVNGZZr8o&t=112s) [A vertical line labeled IPO separates three small ringed spheres labeled public from one large green sphere labeled private. Bright trails converge on the private sphere.](https://assets.ai.engineer/sites/aiecode2025/sha256/8863bae9fbe787d3266377c27eb88a6d531105c5c60f85ad888722e1d2267ea2.webp)

Private is the new public

But a lot of this is a regulation story, and a story about seeking alpha.

High growth companies are finding less and less reason to go public, they would rather just raise a Series F. More control, less paperwork.

And if there are massive, well-funded firms taking over and squeezing the juice out of public markets, well, you're going to start looking elsewhere for your edge.

public    IPO    private

### Matt Levine on private assets and the market return

[2:07](https://www.youtube.com/watch?v=kIlVNGZZr8o&t=127s) [Two quotations attributed to Matt Levine appear beside a diagram of a large blue circle labeled public and a smaller green circle labeled private, each containing scattered points.](https://assets.ai.engineer/sites/aiecode2025/sha256/2e7f6f3e4a6d91d2bd082c4b60ac5974fc0a27588e9f577b5ddf2315fd48d3c9.webp)

As usual with all things finance, Matt Levine said it best:

“Private investment firms have become huge [and] the legal distinctions between public and private markets have become less important.”

“If an increasing share of the economy is in private assets, investors can't just index. You can't get ‘the market return’ by buying every public stock if half of the market is in private companies.”

Matt Levine · Money Stuff

public    private

### Dealing with private assets is really hard

[2:41](https://www.youtube.com/watch?v=kIlVNGZZr8o&t=161s) [A term sheet, Box folder, closing email, Zoom transcript and marked napkin sit beside a secmaster row. Each field shows a question mark, with price paid highlighted.](https://assets.ai.engineer/sites/aiecode2025/sha256/269c4352323e7c4e69c5bf308bfe3255b97a679bbc6bfbfa81781d2d0f4bb509.webp)

Dealing with private assets is really hard

How do you manage private assets? How do you keep track of how many shares you own, or what the strike price is? If you invest in a plot of land next to a hydro plant, where is the info about how much you paid and what your rights are?

It's in PDFs. It's in a Box folder, or in an Outlook email. It's in a Zoom call transcript. It's on a napkin with a hastily scribbled signature from a bar overlooking Columbus Circle.

Welcome to the nightmare world that is managing private asset data.

term sheet.pdf

Box · Deals
cap table v7.xlsx
LPA final FINAL.pdf
side letter.docx

RE: RE: FW: closing

zoom transcript
00:41:07    so on the land...
00:41:12    water rights?
00:41:15    let me check

deal!
Columbus Circle

secmaster    row 48,113
asset        ?
shares       ?
strike       ?
price paid   ?
rights       ?
signed by    ?

### A secmaster is the master list of everything a firm can invest in

[4:35](https://www.youtube.com/watch?v=kIlVNGZZr8o&t=275s) [Vendor feeds, exchanges, custodians and deal docs flow through standards into a central secmaster database. Paths lead from the database to risk, trading, accounting and compliance.](https://assets.ai.engineer/sites/aiecode2025/sha256/0737b42b28ba92e0bf98eb4b9d847c29c19bc1396327e8e9d43fc6cc29371f36.webp)

A secmaster is the master list of everything a firm can invest in

Asset managers need to know what exists in their investment universe. Keeping track of this is a critical back office discipline, bringing together engineering and operations. Put simply, it's the master database of all securities.

For the engineers in the room, the secmaster is like any other master database: It is a source of truth. And often the left side of the join for many business critical analyses.

vendor feeds
exchanges
custodians
deal docs

standards
secmaster

risk
trading
accounting
compliance

### Data quality within the secmaster is the first line of defense

[5:08](https://www.youtube.com/watch?v=kIlVNGZZr8o&t=308s) [The custodian and fund admin both report 12,500 shares for INT-000482, but their highlighted prices differ, 41.02 versus 41.20. The secmaster holds the price as pending.](https://assets.ai.engineer/sites/aiecode2025/sha256/e9ed4a5f9b43c7a60f9532442b9ddeb9ebc9d2ab080fa548e5eb7d3f873ce585.webp)

Data quality within the secmaster is the first line of defense

Say a custodian and a fund administrator report different share counts for the same holding. That conflict gets flagged before it reaches the secmaster. Then you go back to the sources to pinpoint the bug in how each one maps the security. When done correctly, upstream problems are fixed before they contaminate every business critical analysis.

In this way, the secmaster is an organizational primitive. It's a way of separating data producers from consumers, a centralized bottleneck where you can enforce standards, visibility, and quickly enable all data consumers.

custodian
INT-000482
maps to    XLON
shares     12,500
price      41.02
maturity   2031-06-30

fund admin
INT-000482
maps to    XNYS
shares     12,500
price      41.20
maturity   2031-06-30

41.02    ≠    41.20
held

secmaster    INT-000482
shares       12,500
price        pending
maturity     2031-06-30

### For public assets, firms use shared identifiers

[6:52](https://www.youtube.com/watch?v=kIlVNGZZr8o&t=412s) [Lines connect the green internal identifier INT-000482 to FIGI, ISIN, and CUSIP. SEDOL and exchange ticker sit around the same orbit, with a separate yellow risk symbol to the right.](https://assets.ai.engineer/sites/aiecode2025/sha256/7e4a081b561df0bdbecdd8aa8c3bf0e62dbd6da409fea996c0e7af30eb9dbb42.webp)

For public assets, firms
use shared identifiers

Depending on where you invest, you need to take this simple concept and match it up with everyone else around the world, who may be speaking different languages and currencies.

For the most part, publicly traded assets have points of centralization: data vendors use identifiers like Bloomberg FIGIs or LSEG Rics, and numbering agencies provide identifiers like ISIN, CUSIP, and SEDOL.

So your secmaster maps from anyone else's view of an asset into your own, which is basically the only way to understand your positions, like modeling risk or reporting holdings.

INT-000482
FIGI
ISIN
CUSIP
SEDOL
dex code
exchange ticker
risk

### Even "Google stock" has many different identifiers

[7:05](https://www.youtube.com/watch?v=kIlVNGZZr8o&t=425s) [A diagram links Google to Alphabet and branches into three share classes. Classes A and C show different ISINs. Class B shows a CUSIP and is labeled “not publicly traded.”](https://assets.ai.engineer/sites/aiecode2025/sha256/871d578a0bdb3ac6921dd8d4757c925256dc321e286655e884b1acbe6cdbec19.webp)

Even "Google stock" has
many different identifiers

Secmasters can contain a lot of information, but mostly they orient around identifiers. Knowing what kind of assets you have is surprisingly tricky. It's a bit like a datetime library: it seems simple, but there are tons of edge cases.

Let's take the simplest example. Your company holds some Google stock. Surely you can just record that stock in your table and you're done, right?

Not so fast. Which Google stock? On which exchange? Does the ticker match the name of the company?

src · OpenFIGI

Google	Alphabet
Class A	ISIN	US02079K3059
Class B	CUSIP	02079K206	not publicly traded
Class C	ISIN	US02079K1079

### Public data maps cleanly after centralization enforces the identifiers

[7:37](https://www.youtube.com/watch?v=kIlVNGZZr8o&t=457s) [An Alphabet security master lists share classes by identifier and trading venue. Green highlights connect an alternate feed containing Alphabet A, GOOGL, and USD to the corresponding fields in the table.](https://assets.ai.engineer/sites/aiecode2025/sha256/d71415c7872fb14d87428dbccdede60c3b570b4957488c7079931a00acd51e59.webp)

Public data maps cleanly
after centralization
enforces the identifiers

Here's what that looks like in practice. Alphabet is one legal entity, three share classes, and a row for every listing, each with its ISIN, CUSIP, exchange, ticker, and currency.

Say some data comes in with only a subset of those columns. Code can still map it, because the columns it does have are standardized by exchanges and regulators.

But once the code can't generalize over the data with no regular identifiers, we're left with a mountain of manual verification and mapping.

secmaster
security	ISIN	CUSIP	venue	ticker	ccy
Alphabet Inc.
Class A	US02079K3059	02079K305	NASDAQ	GOOGL	USD
	US02079K3059	02079K305	LSE	0RIH	USD
	US02079K3059	02079K305	Frankfurt	ABEA	EUR
	US02079K3059	02079K305	Xetra	ABEA	EUR
	US02079K3059	02079K305	Borsa Italiana	1GOOGL	EUR
Class B		02079K206	not publicly traded
Class C	US02079K1079	02079K107	NASDAQ	GOOG	USD
	US02079K1079	02079K107	LSE	0HD6	USD

name	ticker	ccy
Alphabet A	GOOGL	USD
alt data feed

### Private assets have no standard identifiers and no data feeds

[9:28](https://www.youtube.com/watch?v=kIlVNGZZr8o&t=568s) [A terrain model shows a hydro plant on a river, surrounded by its credit agreement, amendment, collateral schedule, and drawdown notice. A dashed line toward an API is marked “× no feed.”](https://assets.ai.engineer/sites/aiecode2025/sha256/044c675b3c10cf40497adcc4dbfc6857604b07ce4db1adb810f0f924a30e4722.webp)

Private assets have no
standard identifiers and
no data feeds

Private assets make this all really hard. When you buy some Google stock, you have a pretty good idea of what you are buying, at what price.

But say you lend money to the company building a power plant near current data center developments. There's no shared identifiers going into the deal. No one is regulating the full structure of information you get, and there's no data vendor with an API that's already thought through the ontology of these entities in particular.

The repayment schedule, the conditions, and the collateral live in a one-off credit agreement and its amendments. It's all non-fungible. The asset is captured entirely in documentation.

collateral schedule
drawdown notice
hydro project
amendment no. 3
hydro plant
credit agreement
× no feed
API

### Day-over-day changes in a 10M-cell security master

[11:09](https://www.youtube.com/watch?v=kIlVNGZZr8o&t=669s) [Yellow points mark 50 changed cells on day 2 in a grid spanning 50,000 assets and 200 columns.](https://assets.ai.engineer/sites/aiecode2025/sha256/9868836177800fc771824a4ede2795ac88623f5331564d6d960429a6b37d4c3b.webp)

I want to ground this in a specific example. We work closely with a Fortune 200 asset manager here in NYC. They recently started owning a bunch of private assets, and it was causing a lot of problems for their data engineering team.

When any of the assets in the secmaster changed, they had to make sure the change was correct. So they had a system that did a day-over-day change analysis of their table.

If you have a table with 50,000 assets and 200 columns, even infrequent changes are going to happen on some daily cadence.

day 2	50 changed
10M cells
50,000 assets
200 columns

### Manual triage of every changed cell

[11:30](https://www.youtube.com/watch?v=kIlVNGZZr8o&t=690s) [Colored cells from a security master pass through a backlog to four reviewers, each labeled “1 hr.” A counter above reads “516 hours this month.”](https://assets.ai.engineer/sites/aiecode2025/sha256/b7e2c077b286a09ce42b40ce88b3bfc1f24b94290e30726ee39aa1b6d687602e.webp)

Every single changed cell was then sent to a team of people to triage. That team would manually go through docs and data sources to determine if the change was good or bad.

Not every row. Not every column. Every cell.

An asset changes price, or changes names, or goes bankrupt, or changes ownership. Every change in the investment universe has on average an hour of work behind it. That's hundreds or thousands of hours a month.

516 hours this month
secmaster
status
backlog 5
px_close
1 hr
1 hr
1 hr
px_close
1 hr

### What are cloud agents?

[12:36](https://www.youtube.com/watch?v=kIlVNGZZr8o&t=756s) [A diagram places a laptop on the left and a cluster of cloud instances on the right. Between them, a clock represents a scheduled trigger and a secmaster database represents an event trigger.](https://assets.ai.engineer/sites/aiecode2025/sha256/bdc3cdd68ca46fcedaabe69beea3707272e788c96b66ef45f72aa3ea17e09f25.webp)

What are cloud agents?

In simplest terms, a cloud agent is a coding
agent, like Claude Code or Chat GPT, that lives
and runs in a remote environment instead of on
your local machine.

Because it lives in the cloud, we can spin it up
and have it do work on a schedule, or in
response to events -- for example when new
data hits the secmaster.

laptop
cron
cloud
⚡ event
secmaster

### Cloud agent triage for a Parcel 7 ownership change

[13:20](https://www.youtube.com/watch?v=kIlVNGZZr8o&t=800s) [A parcel table highlights Parcel 7’s owner. Beside it, a Nori cloud agent has three evidence sources: Box, Bloomberg, and email. Its findings say Parcel 7 was assigned to Holdco B on Oct 3, with a signed agreement on file.](https://assets.ai.engineer/sites/aiecode2025/sha256/66683595984fad7fda5b429c68e036eb8869d6aa51ba1b27e9a6354631f6e93a.webp)

Cloud agent triage

Whenever there is a change in the investment
universe for this asset manager, a Nori cloud
agent spins up, in this case running a locally
deployed open source model under the hood.

The agent has explicit instructions to figure out
the source of the change. It also has access to a
few useful tools: Box for PDFs, public APIs like
Bloomberg, and even a small segment of email.

The agent then assembles an explanation of the
change, citations for why it happened, and a
recommendation for whether to keep or discard
it. Finally, it pings the owner of the cell in Slack
for review.

asset	type	owner
Parcel 5	land	Holdco C
Parcel 6	farm	Holdco F
Parcel 7	land	Holdco A
Parcel 8	timber	Holdco E
Parcel 9	farm	Holdco C

Parcel 7 · owner
Holdco A → Holdco B   committed

local model
nori cloud agent

Box
assignment agreement.pdf, p.4

Bloomberg
Parcel 7 | Holdco B | 10/03

email
Re: Parcel 7 transfer

cell owner

findings                 Parcel 7 · owner
Assigned to Holdco B, Oct 3
signed agreement on file

play   step

### Human review of Parcel 7’s ownership change

[14:09](https://www.youtube.com/watch?v=kIlVNGZZr8o&t=849s) [A Slack message recommends keeping the change from Holdco A to Holdco B and shows three citations, Keep and Discard buttons, and approval by J. Rivera in 4 min. A bar chart compares resolution times of ~1 hr before and ~5 min after.](https://assets.ai.engineer/sites/aiecode2025/sha256/c6704d572b0dd60037998822743dc3263022d19d5388379d0c7c6bd4fdf15a30.webp)

You still need a human to review changes!

But this dramatically cuts down the response
time, because most of the previous slowness
was people having to manually search through a
bunch of documents.

# secmaster-changes
secmaster-agent   APP   10:42
Owner change needs review

field   Parcel 7 · owner
change   Holdco A → Holdco B

Assigned to Holdco B on Oct 3.
Agreement and Bloomberg both agree.

assignment agreement.pdf p.4   Bloomberg
email · Oct 3

recommend   Keep

Keep   Discard
approved by J. Rivera · 4 min

time to resolve
~1 hr   before
~5 min   after

### AGENTS.md instructions for secmaster triage

[14:41](https://www.youtube.com/watch?v=kIlVNGZZr8o&t=881s) [An AGENTS.md file lists four triage rules. The final rule, “Ignore draft PDFs,” is highlighted in green. A dotted path connects the file to an agent and a reviewer.](https://assets.ai.engineer/sites/aiecode2025/sha256/bbbd3deda530ed1fa12d413c97dd3639f655914451438bc4ec09d3cf24e343c7.webp)

But wait, how does the agent get domain
expertise? How does it know what a 'good
change' looks like?

The same way a new analyst does: from written
instructions about where documents live, which
sources to trust, and how to judge a change.

Agents are pretty good at self-modification!
Changing how an agent behaves is as simple as
editing natural-language English markdown
files. So you use 'generic' agents that are
skinned with configs like SKILLs or AGENT.md
files.

AGENTS.md
# Secmaster triage
- Box deal folder first
- Bloomberg wins on prices
- Flag ownership changes
- Ignore draft PDFs

agent
reviewer

### Cleaner-fish agents watching over the secmaster

[16:02](https://www.youtube.com/watch?v=kIlVNGZZr8o&t=962s) [Small green fish labeled “agents” swim around a stepped pile of dark blocks labeled “secmaster.” A few blocks are highlighted in green or red.](https://assets.ai.engineer/sites/aiecode2025/sha256/08c17299c8341ed457e01bbf178b298211611563cb422e653cbbc83e6b185416.webp)

The agentic secmaster is essentially a database
with a bunch of cleaner-fish automatons that are
dynamically making sure everything is up to
date.

They can quickly spot problems and proactively
surface them.

This technology is not limited to secmasters! Or
even data cleaning in general. It can be useful
across the firm.

agents
secmaster

### Use flexible agent infrastructure, not one-off agents

[16:39](https://www.youtube.com/watch?v=kIlVNGZZr8o&t=999s) [The top row shows four differently shaped one-off agents. Below, four similar cubes sit on one shared agent platform. The text separates the firm’s instructions, tools, and triggers from the vendor’s platform responsibilities.](https://assets.ai.engineer/sites/aiecode2025/sha256/86e1bb467616e5905ec3fdee71f9a78ae46e87c974a4b77b834b49105b976b9f.webp)

Use flexible agent
infrastructure, not one-off
agents

To make that future a reality, you need to use
flexible infrastructure that can be used for any
task.

Your secmaster agent should just be the same
cloud agent infrastructure as your audit agent or
even your engineering agents.

Each has different instructions, tools, and
triggers. That's owned by you. But the
orchestration, security, integrations...all of that
should just be owned by a vendor.

one-off agents
agent platform

### AI adoption is a collaboration problem

[17:26](https://www.youtube.com/watch?v=kIlVNGZZr8o&t=1046s) [A box labeled “problem” contains blocks labeled “cloud agents” and “skills.” A tilted “triggers” block overlaps its right edge. Icons labeled “buyers” sit below the box, with “engineers” farther right.](https://assets.ai.engineer/sites/aiecode2025/sha256/185f8631b3eb438abf9a354cc670db1181b45a792ab272a391ef5ba64b181c37.webp)

AI adoption is a
collaboration problem

People want to buy solutions; engineers want to
sell building blocks. We have to meet in the
middle.

We found the most success working with brilliant
folks who were able to make the leap from
"cloud agents" to "wait, this can solve a problem
I have today."

problem
cloud agents
skills
triggers

buyers          engineers

### Aim for 95% automation and keep a human in the loop

[17:51](https://www.youtube.com/watch?v=kIlVNGZZr8o&t=1071s) [Two green progress bars reach the same 95% mark. The “100% automation” panel shows 0 shipped. The “95% + human” panel shows 5 shipped.](https://assets.ai.engineer/sites/aiecode2025/sha256/1f1d8972f23402239af9fb4ca5ac4a411f774925efd249739e68cb4e02d55b25.webp)

Aim for 95% automation
and keep a human in the
loop

People want to use AI tools to do 100% of
everything. This doesn't work in software, but it
really doesn't work in finance.

It's OK to have a human in the loop. Lean into it.
There are way more problems you can solve if
you're OK with having one human spend five
minutes looking at something.

100% automation
95%
0
shipped

95% + human
95%
5
shipped

### The cost of accidental complexity has dropped significantly

[18:25](https://www.youtube.com/watch?v=kIlVNGZZr8o&t=1105s) [Beside the text, three large blue-outlined squares rest on stacks of smaller gray and glowing green blocks. The diagram has “agents” at the top and “accidental complexity” midway down.](https://assets.ai.engineer/sites/aiecode2025/sha256/4824bf4ebe259f90b71b7c028e73b8deb6f85f8b8fd5492833492a0920d0bdf4.webp)

The cost of accidental
complexity has dropped
significantly

Fred Brooks used "accidental complexity" for the
difficulty that comes from how a system was
built, as opposed to the problem it solves.

In the back office a lot of the work is like this.
Teams record the same data in different ways
and vendors use different formats, mostly
because that's how things evolved.

Agents are now good at handling this kind of
work, which gives people more time for the
problems that are actually hard.

agents
accidental complexity

## Full transcript

[0:12](https://www.youtube.com/watch?v=kIlVNGZZr8o&t=12s) Hi, everybody. Uh, it's really, really great to be here on stage at AIE New York. This is kind of awesome. Um, today, we're gonna talk a little bit about the Agentic SecMaster. But before I dive into things like introductions, I wanna talk about something that's been on my mind recently, especially as I was preparing for this talk, which is finance has gotten kind of weird, right? Like, when I was growing up, finance mostly meant public markets, you know, trading floors, ringing bells, things like that, and that's not really what happens anymore.

[0:43](https://www.youtube.com/watch?v=kIlVNGZZr8o&t=43s) There has been a huge influx of capital investment into private assets, and some of these numbers are kind of staggering to me. Apparently, last year, there were two hundred and forty billion dollars of secondaries volume in twenty twenty-five, which is up fifty percent from twenty twenty-four. I'm sure in twenty twenty-six it's even more than that. And people are now forecasting that we're gonna do thirty-two trillion dollars of alternative assets under management by twenty thirty, which is only three years away. Now, why is that happening? Why is this the big story in the finance world? Well,

[1:13](https://www.youtube.com/watch?v=kIlVNGZZr8o&t=73s) some of this is because AI seems to increasingly be the only growth area in the broader economy, so more and more people are holding onto and investing into all the things that feed the AI world. That's land, energy, data centers, GPU clusters, and of course, secondaries, right? Everybody is holding onto OpenAI and Anthropic secondaries. They're everywhere. All of these private AI companies, everybody's trying to get ahold of them. But it's not just an AI story. Some of this is also a

[1:42](https://www.youtube.com/watch?v=kIlVNGZZr8o&t=102s) regulation story and a story about seeking alpha. High-growth companies are finding less and less reasons to go public. They would rather just raise a series F, more control for the founders, less paperwork to do with the government. And if there are these massive, well-funded firms that are taking over and squeezing the juice out of public markets, well, you're just gonna start looking for edge somewhere else. As with all things, I think Matt Levine said it best. He actually wrote this a year ago.

[2:12](https://www.youtube.com/watch?v=kIlVNGZZr8o&t=132s) Private investment firms have become huge, and the legal distinctions between public and private markets have become way less important. If an increasing share of the economy is in private assets, investors aren't able to just index. You can't get the market return by buying every public stock if half of the market is actually in private companies. So you end up with this cycle. There's more investment in private assets, which means more people have to invest in private assets just to make sure that they are getting the returns from the wider market. Now,

[2:43](https://www.youtube.com/watch?v=kIlVNGZZr8o&t=163s) one thing that's really frustrating is dealing with private assets. It's actually really, really hard. How do you manage private assets? How do you keep track of things like how many shares you own or what your strike price is? I mean, I'm a startup founder, and I find that frustrating just for my own little bit of stock. And if you invest in a plot of land next to a hydro plant, things get really weird. Where is all of the information about how much you paid and what your rights are? Well, unfortunately, it's in PDFs. It's in a Box folder or in an Outlook email.

[3:13](https://www.youtube.com/watch?v=kIlVNGZZr8o&t=193s) It's in a Zoom call transcript. Sometimes it's on a napkin with a hastily scribbled signature from a bar overlooking Columbus Circle. Welcome to the nightmare world that is managing private asset data, something that more and more people in this room and in New York in general have been dealing with over the last couple of years. My name is Amol Kapoor. I am the CEO of a company called Nori Agentic. We build cloud agents, and today we're gonna talk about how you can use cloud agents and AI in general to make it easier to manage, record, and audit a company's

[3:43](https://www.youtube.com/watch?v=kIlVNGZZr8o&t=223s) investments, in particular, a company's security master database. Now, my background is actually in AI. I mean, been an AI engineer for about twelve years, uh, four and a half of those years were at Google. So funny enough, I am not the best person to actually talk about what a SecMaster is. But I'm very lucky to be joined by my CTO and co-founder, Cliff, who actually has spent some time in back office finance, here to talk to us a little bit about what a SecMaster is before we dive into how AI can solve those problems.

[4:13](https://www.youtube.com/watch?v=kIlVNGZZr8o&t=253s) Yeah, I got that. Here we go.

[4:19](https://www.youtube.com/watch?v=kIlVNGZZr8o&t=259s) Great. So I think it's helpful before we, before we talk about where this is a hard problem, I think we should get a lay of the land. We can talk about a simpler example with public equities, then we can tackle some of the complexity that comes with private assets. So every asset manager right now currently runs on a SecMaster. So let's define that. Asset managers need to know what exists in their investment universe or in their tradable universe. This is a critical piece of back office, brings together engineering, operations, and data. You can think of this as just the

[4:49](https://www.youtube.com/watch?v=kIlVNGZZr8o&t=289s) master database of all securities or instruments that represent the assets that they're holding. For the engineers, you can think about this like any master database. It's the source of truth and usually the starting point for most of the bus-business critical analyses. In investment, that looks like risk management, P&L accounting, regulatory compliance, all of these functions that are critical for asset managers. So say two different sources of data disagree about particular values. Ideally, that conflict is gonna

[5:19](https://www.youtube.com/watch?v=kIlVNGZZr8o&t=319s) get flagged before it reaches the SecMaster. You can go back to the sources and pinpoint either the divergence in the particular value, or sometimes it's a logic problem, and you can examine how each one maps that particular security. So when done correctly, upstream problems are fixed before they contaminate all the business critical analyses we just talked about. So I like to think about this as an organizational primitive for how you handle data to enable all the consumers of data at your company. This layer separates the data

[5:48](https://www.youtube.com/watch?v=kIlVNGZZr8o&t=348s) producers from the data consumers. So whether you're working on a complex medallion data lake house that has many different tables or it's just a single denormalized source of truth, this is the bottleneck where you can enforce the standards and the visibility and the quality checks.

[6:06](https://www.youtube.com/watch?v=kIlVNGZZr8o&t=366s) So depending on where you invest, you need to take this simple concept and match it up with how everyone else in the world who invests in this particular asset class, potentially speaking different languages, languages, settling in different currencies, and operating in different regulatory regimes.

[6:25](https://www.youtube.com/watch?v=kIlVNGZZr8o&t=385s) So for public assets, you benefit from a couple different points of centralization. For the needs of exchange and transactions, you have data vendors who will, ah, provide different identifiers. So these might be like Bloomberg Figgies or LSE Ricks. Then there's also regulatory needs, and th- that point of centralization causes numbering agencies to offer a different set of identifiers, like ISINs or CUSIPs. So you're able to use this information to map anyone's view of a particular asset into your own view, and that's the only way you can

[6:54](https://www.youtube.com/watch?v=kIlVNGZZr8o&t=414s) understand your positions, like modeling risk or reporting holdings, joining alternative data, creating financial models.

[7:03](https://www.youtube.com/watch?v=kIlVNGZZr8o&t=423s) So let's take a very simple example. You hold some mega cap publicly traded stock like Google. So already if you know the tickers, is this GOOG or is this Google? Because there are two different share classes traded on the same exchange. Let's say you need to join in some data that exists only at the level of the legal entity. Well, if you're indexing based off of ISIN, now you can't join that anymore. Or maybe you only trade one share class and you only hold one, but you need to trade it across multiple different exchanges because you have different funds for the needs of

[7:33](https://www.youtube.com/watch?v=kIlVNGZZr8o&t=453s) multiple, uh, multiple funds or multiple investors. So already this is getting pretty complex. I like to think about this kind of like a date time library as a comparison for the engineers. It seems very simple when someone solved the problem well, but the moment you're either, you know, working with a new language that hasn't implemented good date time yet, or you just need to make some modifications and peel back how it works inside, it's all edge cases and complex ontologies. So this is what that Google example might look like in practice. Here's a denormalized

[8:02](https://www.youtube.com/watch?v=kIlVNGZZr8o&t=482s) table with many rows for that one equity we want to be able to now, uh, analyze. So some data comes in. Even if it only has a subset of identifiers or exists at a different granularity than your Secmaster, the code can still map it. So here's where it gets interesting. If the code... If the data coming in doesn't have regular identifiers, how do you map it? It's actually pretty clear how AI helps in the first case, right? It's very easy to reason about this data with structure.

[8:33](https://www.youtube.com/watch?v=kIlVNGZZr8o&t=513s) The coding agents are increasingly good at writing code, and if you have this source of kind of gold standard ontology about the business critical data for your company, AI can enable new use cases very easily. But with unstructured, potentially not identified data sources, you have a huge burden of manual verification and manual mapping. So this is where the happy path breaks down when private assets come into the Secmaster.

[8:58](https://www.youtube.com/watch?v=kIlVNGZZr8o&t=538s) So when you're buying the Google stock, you know what you're paying, you know what you're getting, and going into the transaction, everyone has a shared language of different identifiers that they can speak. Let's imagine you're gonna lend some money to a company building a power plant near some data center developments, or you're investing in land near the power plant for potential future, uh, development itself. There's no identifiers everyone shares going into that deal. There's no one that regulates the entire structure of information you might get out the other end, and there's no one data

[9:28](https://www.youtube.com/watch?v=kIlVNGZZr8o&t=568s) vendor that has a single API where they've thought through the ontology of these entities in particular. Everything from the repayment schedule to the collateralization, all of this lives in one-off agreements or amendments. It's not fungible. The asset is captured and defined entirely by the documentation. And this is interesting because now the data scales almost one-to-one with the assets and the burden to scale the d- uh, the Secmaster itself is constantly increasing.

[9:59](https://www.youtube.com/watch?v=kIlVNGZZr8o&t=599s) Great. So I'll pick up back here. Now, Cliff makes a really good point. The private assets are an absolute pain to manage. They just don't fit into the Secmaster all that well. But all that information still needs to actually be collated in some way. It's not like your audit requirements and your cleaning requirements and your standardization requirements disappear. If anything, it's actually a lot worse now because it's just so much less standardized. You can't rely on the people who are doing your risk modeling to know whether or not that data is ground truth. You certainly don't want them to be the ones going through and

[10:29](https://www.youtube.com/watch?v=kIlVNGZZr8o&t=629s) munging through all your documents. So somehow the data that lives in that PDF somewhere, that information needs to make its way back into your Secmaster. And right now, that is just an extremely high labor task. I mean, generally it just takes a lot of people hours and a lot of domain expertise. I'm also gonna really quickly move this mouse out of the way. It's been bothering me. All right. So I wanna ground this in a specific example. We work closely with a Fortune two hundred asset manager here in New York, one of the

[10:59](https://www.youtube.com/watch?v=kIlVNGZZr8o&t=659s) benefits of being New York local. They recently started owning a bunch of private assets, who hasn't? And it was causing a lot of problems for their data engineering teams, the people who were maintaining their Secmaster. Now, when any assets in their Secmaster changed, they had to make sure that change was correct. So they had a system. They did a day-over-day change analysis of their table. If you have a table with fifty thousand assets and each one of those assets has something like two hundred columns, even infrequent changes are going to happen on

[11:29](https://www.youtube.com/watch?v=kIlVNGZZr8o&t=689s) some daily cadence. And every single one of those changed cells was then sent to a team of people to triage because, again, all of this data is super unstructured. It just lives in documents somewhere. So that team was then responsible for going through docs and going through various data sources to determine if any given change was good or bad because upstream data providers may have fat-fingered something or may have gotten something incorrect or they may have just been in conflict. And I wanna highlight that this is not just

[11:59](https://www.youtube.com/watch?v=kIlVNGZZr8o&t=719s) every row, not every column, literally every single cell. So if an asset changed price or if an asset changed its name or if a company went bankrupt or if it changed ownership or you renegotiated a contract or whatever, every single one of those changes in that investment universe would have, on average, an hour of work behind it, an hour of somebody, somebody's time, somebody that you're paying to go and just munge through data. That is hundreds of thousands of hours every single month. Hundreds or thousands, sorry.

[12:29](https://www.youtube.com/watch?v=kIlVNGZZr8o&t=749s) So we're at AIE. How would an AI engineer solve this problem? Two words, cloud agents. What are cloud agents? This is more for the finance folks in the room. In simplest terms, a cloud agent is a coding agent, like Claude Code or ChatGPT, but instead of running on your local machine, it lives and runs in a remote environment instead of on your computer. Because it lives on the cloud, we can do a lot of really interesting things. We can, uh, assert a bunch of audit requirements, we can get auditability and trans- um, transparency, we can get

[12:59](https://www.youtube.com/watch?v=kIlVNGZZr8o&t=779s) observability, and more importantly, we can have it spin up and do work on a schedule without a person being involved. For example, when new data hits the SecMaster. So whenever there is a change in the investment universe for this particular asset manager, a Nori cloud agent will spin up, in this case, running a locally deployed open source model under the hood so all of the data stays on-prem. There's no concerns about information getting out of the underlying environment. The agent here has explicit instructions to figure out

[13:29](https://www.youtube.com/watch?v=kIlVNGZZr8o&t=809s) how the source-- um, figure out the source of this change, and it also has access to a few useful tools, Box for PDFs, public APIs like Bloomberg, and even a small segment of email. The agent then assembles an explanation of the change, gets citations for why that change happened, and then provides a recommendation for whether to keep or discard it, and finally, it pings the owner of whichever cell changed in Slack for review. So you have this agent that is operating basically entirely autonomously,

[13:59](https://www.youtube.com/watch?v=kIlVNGZZr8o&t=839s) providing tasks for a human that the human just has to say yes or no to. And of course, if necessary, the human can go and do further research to make sure that the information provided was correct. Now, you still need a human to review those changes. Of course, this is necessary just for, um, accountability among other things. But this kind of work dramatically cuts down the response time because most of the previous slowness was just people having to manually search through a bunch of documents. Agents are really, really, really good at that, and humans are

[14:28](https://www.youtube.com/watch?v=kIlVNGZZr8o&t=868s) pretty bad at that. So if most of the time is just assembling a case, well, you can take something that used to take an hour and make it just five minutes of human work. But wait, it's worth pausing just to ask how does the agent actually get domain expertise? How does it know what a good change looks like? And I'm concerned that some of the folks in the room might think, "Oh man, I have to go and build this like custom agent infrastructure that's gonna do this really, really specific one-off thing." Well, I don't think that's actually how these things should work. Agents get information the same way a new

[14:58](https://www.youtube.com/watch?v=kIlVNGZZr8o&t=898s) analyst does, from written instructions about where documents live, which sources to trust, and how to judge a change, all in English. Agents are pretty good at self-modification these days, so you can change how an agent behaves by just editing natural language English markdown files. You use a generic agent like Claude Code or OpenCode or anything that's open source like Py, and then you skin those agents with configs like skills or agent MD files. And most of the customization actually comes from the environment in which the agent is running in, which of course is just about what

[15:28](https://www.youtube.com/watch?v=kIlVNGZZr8o&t=928s) kind of configs and what kind of credentials you're dropping on the machine. So I wanna take a step back and zoom out a little bit. Everything that we've talked about so far, this agentic SecMaster concept, this is essentially a database with a bunch of cleaner fish automatons that are dynamically making sure that everything is up to date. They can spot problems quickly and proactively surface them, and even solve them in some cases, all autonomously. And one of the things I wanna highlight is that this technology is not limited to SecMasters or even data cleaning in general. It

[15:58](https://www.youtube.com/watch?v=kIlVNGZZr8o&t=958s) can be useful across a firm because it is a generic primitive in which we are having agents do work that was previously loaded onto humans. In my mind, the incredible potential of AI tools in finance is that they can make a huge range of tasks proactive instead of reactive. Right now, every time anything needs to happen and a human has to spend time and energy evaluating whether or not that task is worthwhile, and then sending that work to an agent. If you ask me, the future of enterprise AI is not use a robot on your laptop to solve a problem. It's

[16:28](https://www.youtube.com/watch?v=kIlVNGZZr8o&t=988s) having the robots come to you with tasks, advice, analysis, and recommendations.

[16:35](https://www.youtube.com/watch?v=kIlVNGZZr8o&t=995s) I think that in order to make that future a reality, you need to use flexible infrastructure that can be used for any task and make it really, really easy to modify what kind of tasks it's being used for. Your SecMaster agent shouldn't be a one-off unique agent. It should actually be running on the same cloud agent infrastructure as your audit agent or your code review agent or any of the agents that your engineers are using for their day-to-day job. Each of, uh, each of these agents will have different instructions and tools and triggers, different context that, of course, is owned by you, by the company. But the

[17:04](https://www.youtube.com/watch?v=kIlVNGZZr8o&t=1024s) orchestration, the security, the integrations, all of that should actually just be the same across these different tasks, and really, in my mind, should just be owned by a vendor because everybody is inventing these problem-- or solving these problems over and over again in the same organizations. So what were the things that we learned by building AI for finance? I wanna leave with three lessons that we, that we had that I wanna impart on everybody here. First, if you ask me, AI adoption is primarily a collaboration problem. People want to buy solutions. Engineers want

[17:34](https://www.youtube.com/watch?v=kIlVNGZZr8o&t=1054s) to sell building blocks. We have to find a way to meet in the middle. We found the most success when we work with brilliant folks who are experts in their domain, who are able to make the leap, that intuit- intuitive leap from cloud agents to, "Wait, this actually fits into a problem that I have right now today." Second, I think it's really important to aim for ninety-five percent automation, and it's okay to keep a human in the loop. People wanna use AI to do a hundred percent of everything, and this doesn't work. It doesn't work in software. We've seen that with software factories over and over

[18:04](https://www.youtube.com/watch?v=kIlVNGZZr8o&t=1084s) again. But it really doesn't work in finance. It's okay to have a human in the loop. In fact, I think we should lean into it. There are way more problems that you can solve if you're okay with having one human spend five minutes looking at something than trying to have an agent do something a hundred percent end-to-end every single time. And finally, I wanna point out that the cost of accidental complexity has dropped significantly. Fred Brooks used the phrase accidental complexity to describe the difficulty that comes from how a system was built as opposed to the problem it solves. And I think in the back

[18:34](https://www.youtube.com/watch?v=kIlVNGZZr8o&t=1114s) office, a lot of work is actually shaped like this. Teams record the same data in different ways, and vendors use different formats, mostly because that's just how things evolved organically. I think agents are really, really good at handling this kind of work, this low complexity, high effort work, which actually then allows people to spend more time on the problems that are essentially complex, the problems that are actually hard. So that's all I got, guys. Thank you so much for coming. Uh, before I go, I wanna just quickly highlight two things. One is if you're interested in more talks like this, we run

[19:04](https://www.youtube.com/watch?v=kIlVNGZZr8o&t=1144s) a community called Agentics NYC. It is New York's largest agentic engineering community. We run regular events on building with agents for beginners through experts. And second, if you are interested in cloud agents, software factories, agentic back office, SecMasters, all of these different things, uh, you should reach out. Cliff and I will be in the hallway track immediately after this. Uh, and of course, you can check us out at noriagentic.com. Thank you so much.
