Lightkeeper

The Shift Is Already Happening

Agentic AI is no longer a concept that hedge funds are watching from a distance. It is arriving fast, and the gap between funds that are prepared and funds that are not is widening every quarter. Despite the growing attention surrounding autonomous agents, most firms are asking the wrong question. The challenge is not whether large language models are capable enough. It is whether the underlying investment infrastructure is ready to support them.

McKinsey data makes this concrete: while 62% of organizations are already experimenting with AI agents, fewer than 10% have successfully scaled them. That gap should give any fund considering an agentic deployment real pause. The primary obstacle is not model performance. It is legacy infrastructure that cannot support autonomous systems operating at speed. In institutional investing, an AI agent is only as effective as the data, controls, and workflows it operates within.

Traditional AI answers questions.  Agentic AI takes action; planning tasks, executing them, and verifying results with minimal human input. And that distinction changes the risk vs reward equation entirely.

The Operational Problem Agentic AI Is Solving

To understand why agentic AI is gaining traction so quickly, you must understand what it is replacing.

Many hedge fund operations still rely on fragmented workflows built around spreadsheets, siloed systems, and manual processes that were never designed for systematic automation. Analysts spend significant time gathering, validating, and formatting data rather than generating alpha. Compliance teams run periodic manual reviews rather than continuous monitoring. Investor relations teams spend days on requests that AI-driven workflows may eventually address in minutes.

What many firms are discovering is that AI initiatives expose operational weaknesses that already existed beneath the surface. Portfolio, accounting, and risk data often live across disconnected systems with different structures, definitions, and update cycles. Humans can work around those inconsistencies manually. Autonomous systems cannot.

According to Gartner, 52% of organizations cite data quality as the single biggest blocker to deploying agentic AI. That figure is probably higher in investment management, where data flows across dozens of counterparties, each with its own formats, timing, and definitions. The fragmentation is deep, and it predates AI by decades.

What AI Agents Actually Require in Production Environments

Most conversations about agentic AI focus on the model, the interface, or the orchestration framework. The more immediate question is what the agent actually operates on.

In a hedge fund environment, agents need to query positions, compare exposures to limits, reconcile activity across disconnected PMS and OMS platforms, and produce outputs that compliance teams can stand behind.  That means three things have to be true before any of this works: the agent needs a reasoning layer that can interpret financial context, it needs access to data it can trust, and every action needs to leave an auditable trail.

The clearest use cases fall across two domains. On the middle and back-office side: reconciling trades across prime broker feeds, flagging compliance breaches in real time, automating audit trails, and maintaining continuous operational surveillance. On the front-office side: synthesizing earnings calls and filings, monitoring intraday risk exposures, generating attribution summaries, and responding to LP queries at a speed and consistency no human team can sustain.

That architecture only works if the underlying data is clean, standardized, and connected. If it is not, agents do not simply underperform; they scale operational misinformation at machine speed.

Why Infrastructure Readiness Matters More Than You Think

Traditional AI tools fail silently. Agentic AI fails actively. A wrong answer in a chat window is unfortunate but contained. When an autonomous system makes a bad decision, it could trigger downstream actions, initiate workflows, and compound errors across connected systems before anyone has a chance to review anything.

Consider what this looks like in practice. An agent monitoring portfolio exposure pulls position data from two systems running on different update cycles, one refreshing intraday, another delayed by overnight batch processing. The agent reads the inconsistency as a genuine concentration breach and automatically escalates a compliance alert. The operations team scrambles to investigate. The alert turns out to be a data timing issue, not an actual breach. But the workflow has already been initiated; the compliance log has an entry that requires a written response, and the portfolio manager lost an hour to an escalation that should never have happened. Now run that scenario across a dozen agents operating simultaneously.

Funds that have already invested in data normalization, cross-system integration, governance controls, and audit-ready reporting pipelines are in a far stronger position to deploy agentic AI responsibly. Those without that foundation will struggle to move beyond experimentation or find themselves managing operational and regulatory exposure that compounds as automation scales.

The Upside of Getting it Right

For all the complexity of getting there, the upside is worth stating plainly. The single biggest constraint in investment management has never been a shortage of data. It has been the time and resources necessary to turn data into actionable ideas. Time to analyze every position, synthesize every report, stress test every assumption. Agentic AI will not change what good investing looks like, but it will change the multiplier on how much a team can accomplish. A fund that once had bandwidth to run analyst reviews annually can now run them quarterly, and the deep dive on trade efficiency that was back-burnered can run every week. For firms with the infrastructure to support it, that is not an incremental gain. It is a new level of what a lean, high-conviction team can accomplish.

Where Lightkeeper Fits In

At Lightkeeper, we are actively evaluating how agentic AI will fit into the future of investment management workflows and infrastructure. We are approaching the development of agentic capabilities deliberately because we believe these systems only become valuable when the operational foundation underneath them is reliable.

Over the past 15 years, we have focused on solving many of the same challenges that autonomous systems now expose: fragmented investment data, disconnected workflows, inconsistent calculations, and the need for auditability across complex environments. Performance analytics, attribution, risk, and reporting all depend on clean, connected, and trusted data flowing consistently across systems.

That foundation matters because agentic AI does not operate in isolation. If those environments are fragmented or inconsistent, automation does not remove operational risk. It amplifies it.

We believe firms that have already invested in data normalization, governance, integration, and trusted reporting infrastructure will be in a much stronger position to adopt agentic workflows responsibly as the technology matures. In many ways, the operational discipline required for institutional investing is becoming the same discipline required for effective AI adoption.

As the industry moves forward, our focus remains the same as it has always been: helping investment teams operate with greater clarity, consistency, and confidence in the data underlying their decisions.

Human Judgment Remains the Constant

The goal of agentic AI in investment management is not to remove humans from the process. It is to remove the friction that prevents humans from doing their highest-value work.

Effective implementations will streamline reporting, compliance checks, and operational coordination while preserving human oversight where judgment and accountability matter most. Approval gates, review checkpoints, and robust audit trails are not optional. In institutional finance, trust and explainability are table stakes.

This is how Lightkeeper thinks about technology broadly. The goal is to give portfolio managers, analysts, and operations teams better information, faster, with greater confidence in the underlying data. Whatever agentic capabilities emerge across the industry, the firms that succeed will be the ones that treat automation and human judgment as complementary, not competing.

What to Do Now

The firms that benefit most from agentic AI will not be the ones who move fastest to deploy an agent framework. They will be the ones who did the harder foundational work first; integrating their systems, normalizing data across prime brokers and custodians, establishing governance controls, and building reporting infrastructure that every function of the firm can trust.

Put simply: if those foundations are in place, agentic AI becomes a meaningful accelerant. Without them, it becomes a liability.

The question worth asking today is not whether your fund should be thinking about agentic AI. It should be. The real question is whether your data infrastructure is ready to support it. Because the model is the easy part. The data always was, and remains, the hard part.

As agentic systems move into production environments, operational infrastructure will become a competitive differentiator rather than a back-office concern.

 

---

Sources

• McKinsey & Company, AI adoption and scaling research (2025/2026)
• Gartner, enterprise agentic AI adoption and deployment research (2025/2026)

Read More

Read More

Read More

Read More

Named Data Science Professional of the Year at the Waters Women in Technology & Data Awards, Lightkeeper’s Managing Director of Analytics has spent a decade helping investment teams turn historical portfolio data into decisions that actually move the needle.

Before Mary Viviano ever helped a client understand their trading patterns, she was the one trying to figure them out herself.

Early in her career, Mary was doing the kind of portfolio analysis she now builds sophisticated tools for, only she was doing it in Excel, manually, with whatever data she could get her hands on. Later, at another firm, she discovered Lightkeeper and realized how much more was possible when the data was properly structured. She became a user before she ever became an employee.

That experience shapes everything about how she works today.

“I’ve sat in their seat,” she says simply. “I know what they’re trying to solve for.”

The Problem with a Single Number

Mary joined Lightkeeper in 2016 and has spent the decade since helping clients unlock something most investment firms are sitting on without fully realizing it: the story inside their own portfolio history.

The starting point, she explains, is recognizing what a standard P&L number doesn’t tell you.

“I made 300 basis points in Microsoft, but that doesn’t tell you anything. What would you have earned if you’d just invested in the market for that time period? What if you’d doubled down here, gone in more quickly or more slowly, exited differently?”

A single outcome number, she argues, can’t help you improve. To do that, you need to understand how the decision was made; the entry, the adjustments along the way, the exit, and whether each of those steps actually added value.

That’s the foundation of Trade Decision Analytics, the framework Mary developed at Lightkeeper that has become one of the platform’s most significant analytical advances. By decomposing trading activity into distinct phases and measuring how each contributes to long-term portfolio P&L, investment teams can start to identify consistent patterns in their own behavior and use them to make more informed decisions.

“If people understand that they have a pretty consistent bias in how they open positions or close positions, or how they react when certain things happen in the market, they can lean in,” she says. “Maybe: I’m really convicted in this name; I should put the full position on all at once and save myself money. Or: when a stock moves against me by 25% and I double down, that’s usually a bad idea. Understanding that and incorporating it into the investment process, that’s how you maximize returns.”

The Work Behind the Insight

What doesn’t come through in an award citation is how hands-on Mary’s work actually is.

When a client needs analysis, she goes into their data manually, works through the statistics she thinks will tell the most useful story, and builds out a deck they can review together. It’s time-intensive, deliberate work, less data engineering, more portfolio coaching.

“The most rewarding part is helping clients see patterns in their own data that they hadn’t noticed before,” she says.

New analytics at Lightkeeper usually start with a client conversation, not a specific feature request, but a question a client is wrestling with. Mary and her colleagues dig into it, develop an approach, and then take it back to clients for feedback.

“We can’t just create new analytics in a vacuum. It has to be based on discussions with clients, making sure it’s going to be useful, that it’s understandable, that clients can access it without it taking 35 minutes to get an answer.”

Often, one question becomes five. “Someone might ask for something specific and when you dig in, it turns into several things, because you realize someone else was asking a similar question, just in a different way. It’s iterative.”

Building for the Long Term

Beyond the analytics themselves, Mary has worked to make Lightkeeper’s knowledge more accessible to clients directly. About two years ago, she and colleague Stephen Scherock developed the Knowledge Center, a searchable resource library where clients can find detailed explanations of analytics and methodologies without needing to pick up the phone or send an email.

It grew out of a practical problem: Mary and Stephen had built up a library of explanatory documents they’d send out whenever a client asked a question. Putting them somewhere clients could find on their own just made sense.

It’s also, she notes, the kind of well-structured content that will matter more as AI plays a larger role in how clients interact with data. “In the age of AI, for that content to be searchable and usable, not just a stat definition but a real explanation of how something works, I think that’s really valuable.”

On AI more broadly, Mary is measured. She sees real potential for it to make her own work more efficient, particularly the manual, server-by-server portfolio analysis she currently does by hand. But she’s clear-eyed about what won’t change.

“People still need to be able to understand it and translate the data to other people. If you’re in a position where you can do that, that’s a real benefit.”

Recognition She Didn’t See Coming

When asked what it felt like to win the Data Science Professional of the Year award at the Waters Women in Technology & Data Awards, Mary is characteristically understated.

“It’s not something I would ever even think about. The fact that colleagues took the time to think about me, to put my name in and do all the work for the nomination, that’s what I’m really appreciative of.”

She’s not someone who seeks the spotlight, she admits. But for those who work with her, the recognition landed exactly right.

“Mary has a unique ability to bridge the gap between sophisticated analytics and real-world investment decisions,” says Dean Schaffer, CEO of Lightkeeper. “She understands the technical complexity of the data, but just as importantly, she understands how investment teams actually work. That combination allows her to turn analytics into insights that genuinely help clients improve their investment process. One of the greatest value-adds that Lightkeeper can provide to our clients is access to and insights from Mary.”

Greg Johnson, Senior Managing Director, Client Solutions of Lightkeeper, puts it this way: “She doesn’t just build analytics, she works closely with clients to understand their challenges and helps them apply the insights in meaningful ways.”

She also has a message for young women considering a career in data science. “If you like math and turning data into actionable insights, I’d definitely suggest it,” she says. “It’s market adjacent, which means it’s constantly changing and exciting. And with AI making such a difference, if you’re in a position to understand it and translate it to other people, that’s a real benefit.”

For Mary, the work and the recognition point to the same thing: data is only as useful as what you do with it.

“The data is already there,” she says. “The real value is helping teams understand what it’s telling them.”

According to Hedgeweek’s Q1 2026 Global Outlook Survey of more than 100 hedge fund managers, 41% now rank AI integration as their biggest priority for the year, surpassing both cost optimization and talent acquisition. Nearly a third report significant AI integration already underway across research and trading.

But a closer look reveals a critical blind spot.

The promise of large language models is straightforward: ask a question in plain English and get an answer in seconds. The challenge is that the questions that matter most to a portfolio manager, positions, attribution, and risk exposures, live in proprietary systems that general-purpose AI tools cannot easily access or verify.

When AI works from disorganized or inconsistent data, the result isn’t just an inconvenience. A plausible but incorrect number in a risk report or investor letter is a material, and potentially career, risk.

This highlights an important distinction discussed in the article: generative AI can produce useful insights, but investment teams still rely on deterministic analytics when it comes to their own portfolio data. For market commentary or macro analysis, approximate answers may be acceptable. For your own book, they are not.

The firms best positioned to benefit from AI may not be the ones deploying the most sophisticated models, but those that have invested first in clean, validated, well-structured data infrastructure.

Read the full Hedgeweek article: Hedge funds rank AI as their number-one priority — but experts say they may be ignoring this blind spot.

 

Most conversations about AI in asset management begin with capability, what the models can do, how quickly they are improving, and where productivity gains might appear. Those discussions are understandable, but they skip the more determinative question: what kind of data environment the technology is being asked to operate in?

AI is only as effective as what it is fed. And in many firms, that data diet has been shaped by years of manual processes, informal corrections, and workarounds that made traditional workflows workable, but fragile.

For a long time, that fragility was tolerable.

AI Does Not Like Band-Aids

Historically, most data environments were imperfect but serviceable. Information flowed through a mix of systems, spreadsheets, reconciliations, and institutional knowledge. Reporting cycles allowed time for review, adjustment, and interpretation. Teams knew where the gaps were and how to manage them.

That buffer is becoming more costly.

AI collapses the distance between questions and data, removing much of the time and human intervention. That is a huge benefit, but it also means you lost a buffer that previously absorbed inconsistency. When AI is introduced, answers arrive immediately, whether the foundations are ready or not.

AI doesn’t reconcile ambiguity. It magnifies it. Inconsistent definitions, unclear lineage, or fragile dependencies show up directly and inconsistently in the outputs. What once required experience to detect now appears in front of users, often without warning.

This is why AI can feel unreliable on enterprise data sets. Not because the technology is immature, but because it exposes environments that were never designed to operate without human arbitration.

What AI Reveals About How Firms Actually Operate

Most firms rely on more operational workarounds than they tend to acknowledge. Manual reconciliations embedded in workflows, spreadsheets that standardize data after the fact, and unwritten rules known only to certain teams are common features of day-to-day operations. In many cases, these practices were developed for good reasons. Often, they were the only practical way to keep things moving.

As firms try to capture the benefits of AI, those dependencies become visible. The technology has no awareness of which data source is considered authoritative, which adjustments are provisional, or which exceptions apply only in specific contexts. Logic that exists outside formal systems cannot be inferred.

Consider a simple example: A portfolio manager asks an AI tool for exposure across several strategies ahead of a client conversation. The system returns different numbers, each technically correct, but based on different definitions embedded across legacy systems of market value, notional value, and delta-adjusted exposures. It’s that “exposure” was never defined consistently in the first place.

Moments like this reveal how much reliable output has depended on people bridging gaps rather than systems providing consistency. The shortcuts that worked in the past become constraints on what is possible going forward.

Feeding AI Well Requires More Than Clean Data

When firms talk about preparing for AI, the conversation often defaults to data cleanliness. Records should reconcile. Fields should be populated. Numbers should tie. These things matter, but they’re not enough.

Much of the information firms use every day are computed data elements that are often too diverse to store. A manager may want their returns across an arbitrary data range or an average exposure across time. This requires correct data but also a system that can consistently provide those values without the models resorting to “computation,” which generally involves writing calculations in custom code.

So, reliability is not just about data quality, but data breadth to support the far-ranging inquiries natural language avails itself.

Feeding AI well isn’t about hygiene. It’s about readiness.

Data readiness is an architectural property. It reflects whether data is structured consistently, whether relationships are explicit, and whether the supporting platforms can answer the same question repeatedly without manual adjustment. Firms that focus only on cleanup often discover they’ve addressed symptoms rather than causes. AI simply makes that distinction harder to ignore.

Why Context Determines Whether AI Adds Value

Even well-structured data has limits without context. Metrics don’t exist in a vacuum. Exposure, performance, and risk take on meaning only when interpreted through strategy, mandate, and intent.

Human teams apply this context intuitively for their investment strategies. They know which comparisons make sense and which assumptions apply. AI models generically do not, unless that context is explicitly embedded in the data and workflows they rely on.

Without it, AI can summarize information without producing insight. Outputs may be accurate yet still fail to support decisions. Confidence erodes not because the answers are wrong, but because their relevance is unclear.

Context is what separates AI that reinforces existing workflows from AI that introduces more noise than clarity.

AI will Compound Operational Alpha

Operational alpha has always been about leverage: turning data, systems, and workflows into more valuable outcomes without proportional increases in effort or risk. Firms with strong foundations were able to scale complexity, respond faster, and operate with confidence long before AI entered the conversation.

That hasn’t changed. What has changed is how quickly benefits can be reaped and weaknesses exposed.

AI doesn’t create operational alpha but can compound it. Gaps that once appeared only under stress (growth, customization, heightened scrutiny) now show up earlier and more often. As a result, operational capability becomes a more immediate factor in whether AI becomes an asset or a liability.

Models will continue to improve. That part is largely out of a firm’s control. What firms can control is what they choose to feed these systems.

What to Ask Before Your Next AI Initiative

Before moving forward, operators should be able to answer a few basic questions:

• Would two teams asking the same question get the same answer?
• Are key definitions enforced in systems, or explained after the fact?
• When outputs differ, do teams trust the data or start reconciling?

In the end, feeding AI well isn’t about ambition. It’s about process and systematic capability. And increasingly, it’s the underlying infrastructure, not the model, that will determine whether AI delivers leverage or exposes limits.