Liron Mandelbaum

The through-line

Mandelbaum's fixed point is that in private markets the bottleneck is not analysis, it is getting the data in the first place. Seventeen years at Bloomberg taught him to take the raw material for granted, because there "data was publicly available and it had to be publicly available and distributed to everybody simultaneously," which meant everyone's effort went into "analysis calculations on this data making predictions on this data" [2]. The aha moment came when he found the opposite world in fund selection and manager due diligence: "The data component is not taken for granted. In fact, that is the biggest challenge just getting the information in an efficient way" [2]. From that observation everything else follows: the case for structured data over slick interfaces, the case for AI embedded in workflows rather than chat, and the argument that fixing collection is ultimately an argument about who gets allocated to and what end investors earn.

The second, more recent layer is about how AI should be applied to this problem. His view is that generative AI is the first thing in years that genuinely bends the curve on questionnaire overload [1], but only if it is used to build a durable, structured asset rather than to produce answers that evaporate into a chat window.

On the dirt road and the six-lane highway

The image he keeps returning to is infrastructural. In listed markets and portfolio analysis there are "amazing sixlane highways," but "gathering the data about the funds that you invest in is a dirt road," a matter of emails and files going back and forth [2]. He frames this as a scale mismatch rather than a niche annoyance: hundreds of trillions of dollars in assets are allocated through this channel, and allocators are collecting the information as a fiduciary obligation, not a nice-to-have [2]. He also insists the stakes are personal rather than institutional. Pension funds, insurers and banks are ultimately individuals who are "not involved in decision-making of how where our money really goes in many situations. We're just counting on it to be there for us when we need it" [2].

His diagnosis of the allocator's position is that the data problem crowds out the actual job. Teams are stuck "just getting this confirmation," never "ever dreaming of getting to a place where you could analyze it, worry about your calculations" [2]. And he is blunt with clients about inertia: "sometimes the status quo is just the worst thing" [2].

On questionnaires that keep getting longer

He describes a trend he calls scary, running for about six years and worsening: "questionnaires are getting longer and more sophisticated because strategies have to differentiate" [1]. The result is symmetrical damage. Long, in-depth DDQs consume the investor relations team on the manager side and take just as long to review and analyze on the investor side, so that "both sides are getting drowning" [1]. Until generative AI arrived roughly two years ago, he says, "there was uh very little hope" [1]. He is equally clear about the size of the relief available: one pension that took four to five months to gather data from 130 managers now runs the same exercise in four weeks, and the four weeks exist only to give managers time to answer. "The whole collection is now instantaneous" [2]. The consequence is not just speed but cadence. What could previously be done once a year can now be supplemented with smaller quarterly updates, so a strategy drift or a newly identified risk can be acted on "in a realistic timetable, not annually" [2].

On why data comes before the interface

His sharpest argument against the current wave of AI products is that they optimise the wrong layer. A slick chat that lets you connect documents and ask questions is "a great experience, but what you're forgetting is where is all that insight I'm getting going to?" [1]. In due diligence, particularly on the LP side, every piece of research should be an act of data creation: "You should be thinking of creating data with every research attempt that you do," which he calls "the flywheel effect" [1]. Where other tools leave insight stranded in isolated chats, his position is that the output should accumulate into centralised, structured, comparable intelligence [1]. That structure is what makes apples-to-apples comparison possible across asset classes and across managers in a single database [2], and it is what turns unstructured, siloed and stale DDQ responses into something an allocator can actually use [2].

On keeping the human out of the hamster wheel

He accepts human-in-the-loop as necessary and rejects the version of it that most tools deliver. The failure mode is the person "constantly interacting with the AI," which he pictures as "somebody running in a in a on a treadmill or in a hamster wheel" [1]. For due diligence and investor relations, that is simply the wrong use case: "to scale the work, you'll exhaust yourself" [1]. His alternative is to put the AI inside a scalable workflow that automates what it can and returns to the human at defined steps to review answers or tag things, so the human is running the workflow rather than being run by it [1]. The same logic explains why Sidekick is built into what teams already do: "it's not foreign and it saves massive amount of time on critical tasks," freeing people for high-value work and, in his phrase, bringing "back the water line so everybody could see what's going on and participate" [1].

On accuracy through narrowing the question

On hallucination he offers a specific method rather than a reassurance. "The trick is to isolate as much as possible what you're asking": pose one question at a time, hide the other questions, supply only the documents relevant to that question, pull the answer and analyze it [1]. Two years of building and deploying this way has produced "much more discreet answers, more clear answers and um with higher accuracy with essentially almost zero hallucinations" [1]. He is candid that this design has a cost, and accepts it: "That's a lot more expensive, if you will, but it's the best approach" [1]. He also stresses that this is shipped rather than promised, with around 100 LPs and 100 GPs using it actively across investor relations and investing use cases: "it's not a road map. The value is already here today" [1].

On democratizing access to data

His account of how the market got here is that the large consultants and the biggest allocators by AUM built digital collection tools first, because at their scale nothing else worked. That paved a road, but "all that data just went only for their eyes only," and it meant "the bigger the player, the more AUM you had, the easier access you had to data" [2]. Neutralising the collection burden let those players analyse better and decide better, with no level playing field for anyone else. His answer is a pricing and access argument as much as a technology one: build the tool and "price it appropriately for various player types and on their usage," so that any allocator, whatever its size, can get structured data from and about its managers [2].

He is careful about what allocators think they are buying. The purchase decision is about efficiency, and clients believe they already give everyone a fair shot [2]. What surprises them a year or two in is the second-order effect: with months of collection work removed, they can review more managers and more opportunities in greater depth, which is where better decisions and better end-investor outcomes come from [2].

On letting the manager's story shine

The mirror image of the allocator's problem is that managers cannot get their real differentiators across. Track record and strategy description, he argues, "essentially becomes a commodity and it is a commodity"; what actually matters is how a firm operates, how it thinks about strategy, and its talent, and that is precisely what gets lost when it is "buried as question 207" and answered in long free-text paragraphs that no investor can compare [2]. He notes this bites hardest where returns are tight: "when some strategies have very narrow spread and returns, your operations matter a lot," as does how a firm retains talent [2].

The consequence he cares about is selection bias. When allocators are overwhelmed, "the LPs go with the known winners. They make the easy decision. They go with the tier ones, the funds they know," which means newer and even medium-sized funds never get a chance to grow, even where their returns would be superior for the end investor [2]. So the point of better collection is not only risk oversight; it is that managers "want to let their story shine," and if the best story is heard, the pension holder on the other end benefits [2].

On building tools for this industry specifically

Mandelbaum treats the responding side as a first-class customer rather than a data source to be squeezed. Manager-facing capabilities, including reuse of historical answers and AI matching of new questions to previously answered ones, are given away free with the collection request, and he points to industry awards voted by investor relations and client services teams as the evidence that this lands [2]. He also refuses to be purist about workflow: sometimes a team genuinely needs to leave the highway, so "we'll give you a truck, you'll stay in Excel, take that, get the information you need, put it back into the highway" [2].

His argument for a vertical product is that generic RFP tools have existed for decades but "they're designed for hospital vendors to car manufacturer, parts makers responding to RFPs" and do not understand concepts like strategies and funds, or the fact that different parts of an organisation must answer different questions [2]. Managers then asked for the same logic to be applied to requests arriving as long emails and Excel sheets, so those can be parsed and handled as if they were digital requests [2]. Integrating regulatory and database sources such as form ADV as inputs to answers is something he says nobody else was doing [2]. The payoff is reach: a 400-question RFP that was never worth a week of effort at low odds of winning can be largely prefilled, so one investor relations person can pursue far more opportunities [2]. And the time saved goes into craft, because "Responding to an RFP, as we learned from our clients, is an art"; when "90% of the RFP is finished, you can focus on those five questions that will make the difference" [2].

Takeaways

• The private markets bottleneck is collection, not analysis: listed markets have "sixlane highways" while gathering data on the funds you invest in "is a dirt road" [2].
• Questionnaires have grown longer and more sophisticated for about six years as strategies differentiate, leaving both managers and investors "drowning"; generative AI is the first real relief [1].
• Build AI into existing workflows rather than standalone chat, and let the human review and tag at defined steps instead of "constantly interacting with the AI" in a hamster wheel [1].
• Cut hallucinations by isolating each request to one question with only its relevant documents, accepting that this is "a lot more expensive" as the price of accuracy [1].
• Judge AI tools by where the insight lands: research should create structured, centralised data every time, a "flywheel effect," not answers stranded in isolated chats [1].
• Concrete outcome: one pension went from four to five months collecting from 130 managers to a four-week cycle that exists only to give managers time to answer [2].
• Digital collection was historically the preserve of the largest allocators by AUM, with the data kept "for their eyes only"; usage-based pricing is how that access is opened up [2].
• Overloaded LPs default to "the tier ones, the funds they know," so structured comparison is what gives newer and mid-sized managers a hearing on operations and talent rather than commodity track record [2].