Your AI Isn't the Advantage. Your Judgment Is - Sid Pednekar
Loading slides…
Read the talk
Where Should an Asset Manager Spend Its Best Judgment?
Sid Pednekar follows two hypothetical research teams from earnings transcripts to portfolio decisions to show why the choice of problem, the knowledge colleagues share, and what they learn after deployment determine the value of AI.
From a talk by Sid Pednekar
At a glance
Ideas worth remembering
Trace a material signal from its source to the person making the decision. Locate the delay before funding a faster intermediate step.
Use an explicit investment thesis to guide continuous monitoring, including the search for evidence that could refute it.
Share dataset meanings, gaps, and failed approaches so the next researcher can begin with what colleagues have already established.
Preserve the evidence behind corrections and track where freed capacity goes. Both help determine what to change after deployment.
An analyst can save time before the firm sees a return
Dr. Sid Pednekar approaches enterprise AI from asset management. He has spent much of his career in hedge funds and in investment and academic research, building specialized teams around three objectives: improving efficiency, driving profit and loss, and reducing costs. Those objectives give him different ways to assess the same piece of engineering. A tool can help an analyst finish research sooner. Whether the firm spends less to produce that research, or makes a better investment decision because it arrived sooner, requires a different kind of evidence.
The survey figures Pednekar cites illustrate the gap. About 37% of surveyed enterprises reported a positive contribution from AI to earnings before interest and taxes (EBIT). He puts the high-performing group at about 6% and associates high performance with a roughly 5% contribution to EBIT. In other studies earlier in the year, about 66% reported productivity gains, 40% reported cost reductions, and 20% reported revenue growth. These figures come from different studies. They describe the spread of reported outcomes without establishing how often a particular productivity gain translated into profit.
To examine the decisions behind those outcomes, Pednekar follows a research analyst at each of two hypothetical asset managers, Company A and Company B. Their workflows are thought experiments. He uses them to develop three practices that an individual, a function, or an entire firm can adopt: design the right problem, build shared capabilities, and compound the advantage.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
More answers increase the demands on judgment
Consider the different things a firm can now produce cheaply: software, analytics, recommendations, reports. Each can answer a question or enable an action. An abundance of recommendations still leaves someone deciding which deserve attention. Pednekar defines judgment as choosing what matters and how to act when several answers are plausible. Creativity helps someone identify possibilities worth exploring. Taste helps them recognize good work in a particular context. Weighing tradeoffs lets them choose a course of action.
Those decisions happen throughout a firm. A software engineer chooses what to build, a data scientist chooses what evidence to pursue, and a portfolio manager decides how much weight to give a finding. The C-suite and AI leaders exercise judgment, too, but much of the relevant knowledge sits with the people doing the work. Pednekar’s resource-allocation question is where their best judgment is worth spending. That question applies before anyone chooses a model, a harness, or a copilot.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Trace the late signal before choosing the tool
Company B starts with a copilot and looks through the analyst’s existing work for somewhere to put it. An earnings transcript arrives, the copilot summarizes it, and the analyst receives a shorter document sooner. Once the team discovers that use, it can feed the whole stream of incoming transcripts through it. The benefit is easy to observe. The cost is harder to see. Summarization reduces information, and a detail that disappears in compression may matter to the decision the analyst has to make.
Company A starts with a problem familiar to the people responsible for investments: material signals reach them too late. That wording leaves several explanations open. Perhaps analysts lack the capacity to work through incoming research. Perhaps the underlying data is poor, arrives late, or is unreliable. Perhaps the analyst already has the evidence and the delay occurs when someone must decide what to do with it. Faster reading helps with the first explanation. If the delay occurs after the recommendation is ready, the team needs to look further downstream.
Locating the bottleneck tells Company A where to direct engineering capacity and analyst attention. It also gives the team a way to assess a proposed improvement: follow the material signal far enough to see whether it reaches the decision-maker sooner. Pednekar connects this selectivity to studies in which leaders concentrated their attention on fewer, higher-quality decisions.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Fund the research pipeline deeply enough to change it
Pednekar recognizes the appeal of experimentation and says he has felt the pull of having too many opportunities himself. Company B runs into trouble when it starts asking how many AI pilots it can launch. It builds a document summarization bot, tries a copilot elsewhere, and adds a data retrieval bot. The firm can count each completed pilot and move on to the next initiative without ever asking whether material signals now reach the portfolio manager sooner.
Company A chooses the end-to-end research pipeline as the unit of investment. Engineering, analytics, AI capacity, and the people who understand the investment decision can all work on the same delay. If the bottleneck moves from reviewing evidence to acting on it, that second step still belongs to the project. Pednekar cites research finding that leaders direct more than 80% of their investment to high-value use cases, including transforming key business functions and creating new offerings.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Give continuous monitoring an investment thesis to test
In Company B’s version, the work retains its familiar order. The analyst receives earnings transcripts and research reports, reviews them, updates a financial model, and sends a recommendation to the portfolio manager. AI can accelerate each handoff. Company A could change the analyst’s first action: ask them to define an investment thesis. Now the system has a proposition against which to evaluate incoming information, including evidence that could refute it.
With a thesis in place, monitoring and evidence gathering can run continuously. The analyst investigates exceptions, decides which developments warrant a deeper look, and weighs the evidence for and against the thesis. This gives a concrete destination to some of the time freed by automation. It also changes the analyst’s responsibility. Their judgment shapes what the research process looks for, as well as the recommendation that comes out of it.
Pednekar cites research suggesting that high performers redesign workflows more often. In his example, Company B still asks how quickly an analyst can send a recommendation to the portfolio manager. Company A gives the analyst a different job: decide what evidence would change the investment thesis and which exceptions deserve investigation.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Bring portfolio managers into the build and share what researchers learn
Redesigning the research process takes people who understand both the systems and the investment decision. Engineers are moving closer to business decisions, while business professionals are building tools and automations themselves. Alongside forward-deployed engineers and product managers, Pednekar imagines a forward-deployed portfolio manager. Working with the builders, that manager can explain which evidence could change a position. An engineer working with the analyst can see where that evidence loses context. Pednekar connects this collaboration to a broader shift: he cites studies finding that more large enterprises see their technology teams as a source of business growth.
Pednekar describes the design principle as standardizing the core and personalizing the edges. For data scientists, a shared core could include datasets and their semantic layers, which give the data meaning. It should also carry the experience of using them. Which questions has a dataset answered well? Where are its gaps? What tradeoffs did a predecessor encounter, and which approaches failed? Those findings are part of the work a new scientist would otherwise have to repeat.
Scientists can pursue different research questions using that common foundation. An analyst working with a client may need a different workflow again. The practical test for what belongs in the core is what the next person should be able to take as a starting point.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Follow the agent from the Slack conversation to the ticket
Much enterprise software was designed around one person, or a group of people, interacting with one application. An agent may need to gather context in one application, deliberate, and act in another. The quality of the data and the ability to reuse it become constraints on the whole operation. Pednekar reports that 71% of organizations in a study were modernizing core infrastructure to support AI. He also reports that 23% were investing between 6% and 10% of annual revenue in core system modernization.
He distinguishes three types of team agent, depending on what an organization is ready to support. An expert agent makes subject-matter knowledge accessible to colleagues. A workflow agent carries work across systems. His example is a Slack bot that opens and closes tickets: the conversation supplies context for a change to the state of a ticket elsewhere. A chief-of-staff agent takes a wider view, tracking the state of work across the organization and what still needs doing. Each depends on the firm making the relevant knowledge accessible at the scope of the job.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Keep the evidence that prompted a correction
Once the workflow is in use, an analyst may catch something it missed. Correcting a recommendation benefits its immediate recipient. Pednekar asks how the next analyst will benefit from the discovery. The correction might supply context that future investigations can draw on, or expose a need for a new tool. In either case, preserving the evidence alongside the lesson lets a colleague judge when it applies. A successful investigation offers the same opportunity to make an individual discovery reusable.
Company A can close that loop by returning to the problem it chose. If continuous monitoring frees analyst time, how much of that capacity goes into investigating exceptions? Do material signals reach portfolio managers soon enough to inform their decisions? These are questions the hypothetical firm would need to measure after deployment. Following them connects the original investment to the next redesign and gives the team a reason to keep learning after the pilot succeeds.
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…