Artem Fokin on Improving with AI and Expert Calls
Summary
Fokin sees expert-call libraries and AI as the two disruptive innovations that most changed investment research, not merely incremental process improvements. Walker contrasted independent investors with institutions that could commission 20 calls at roughly $1,000 each and deploy large analyst teams. Fokin says libraries lowered expert-call costs and converted a pure variable-cost model into something semi-fixed and semivariable, while AI compresses summarization work. The emerging edge is “human plus machine,” not either one alone.
An expert call should target the load-bearing assumption in the investment thesis. If the thesis rests on product superiority, interview users, switchers, evaluators who declined, and people who never considered it; if distribution is the question, talk to former salespeople and reconstruct the sales motion. Fokin’s reminder is deceptively important: “Sales just don’t happen.”
The best expert work surfaces disagreement rather than manufacturing consensus. Walker’s medical-device example paired a reported 1% defect rate with 10% for older products, yet surgeons argued that the comparison used decades-old studies and ignored subsequent improvements in technique. Fokin similarly heard reactions ranging from liability-conscious enthusiasm to “I don’t really care” when asking doctors about Sofwave’s almost 10 FDA clearances: “Nothing is certain. Different people have different opinions, and that’s okay.”
Former employees are useful witnesses, but their testimony needs context, corroboration, and relentless follow-ups. Fokin asks laid-off employees whether they would return or recommend the company to family, then checks their claims against product quality, customer enthusiasm, management commentary, and other interviews. Investing rarely reaches “beyond a reasonable doubt”; the achievable standard is a “preponderance of evidence.”
Volume matters when the business is complex, but disciplined curation determines whether volume becomes insight. Fokin conducted five bespoke calls on IWG, still felt he had not figured it out, and then read roughly 70 IWG and WeWork calls; the available library has since approached 100. At Crocs, historical interviews helped distinguish a temporary COVID beneficiary from a company that had built its commercial foundation before COVID while the stock traded around 5.5–6 times 2022 earnings.
Fokin currently gives AI a vote in “digging” and part of “analyzing,” but zero authority in “deciding.” As of August 6, he had found many ways to compress existing work but no meaningful task AI enabled that he could never perform before. A question such as how a company expanded from 10 states can now be answered with historical context and sources in seconds instead of a manual filing search; Walker gave three-proxy incentive-compensation analysis as a similar time saver.
AlphaSense’s Grid and Deep Research workflows turn large document sets into navigable research maps, without replacing primary-source reading. A Grid can place 20 expert calls against as many as 12 recurring questions, exposing consensus, red flags, and the interviews worth reading in full; Deep Research can turn a five-page prompt into a sourced 30–40-page primer that “prepares the mind” before the filings and calls. Fokin attributes “90 or maybe even 95%” of what he knows about new AlphaSense features to repeated sessions with his account manager, while product managers provide another source of use cases and feedback.
Deep dive
1. Expert-call libraries and AI changed the economics of research
Walker’s baseline was the resource gap between institutions and independent investors: McKinsey or Bain Capital could commission 20 calls at roughly $1,000 each and deploy 15 analysts across the literature; a solo practitioner could not. Libraries and AI have narrowed both disadvantages.
Fokin’s correction was categorical: these are not merely “process improvements,” but “disruptive innovation.” Expert-call platforms connected creators and readers, lowered costs, converted a pure variable-cost model into a semi-fixed or semivariable subscription model, and—he believes, without claiming access to the data—expanded both usage and the total addressable market.
Alternative datasets such as credit-card data remain “very, very, very, very expensive” and play little role in Fokin’s process; his simpler alternatives include Google Trends. For him, expert-call libraries and AI are the two largest research innovations of the past decade.
His own expert workflow has two modes: “I either listen or I read.” On IWG, five bespoke calls failed to resolve a complicated global flex-space model, so he methodically read roughly 70 calls spanning IWG and WeWork; with the library now nearer 100, old customer and industry observations still answer questions that recur years later.
2. The thesis determines which expert deserves the hour
Walker’s hard case was a conglomerate such as Berkshire Hathaway—or a roll-up such as Constellation Software—where no obvious expert maps cleanly onto the whole thesis. Fokin’s golf analogy supplied the operating rule: different terrain requires different clubs, so first identify the “key ingredients for success” rather than defaulting to one expert type.
If the claim is that a six-month-old product is materially safer, cheaper, or more efficient, Fokin would concentrate on the product: existing users, customers who switched, evaluators who tried but declined, and prospects who never investigated it. The last two groups can reveal what blocks an apparently “10x better” offering.
If the product is established but growth depends on distribution, the research target changes to former salespeople. Fokin’s Stanford professor reduced business to two problems: “Problem number one, not enough sales. Problem number two, everything else.”
For Sofwave Medical, Fokin asked a former salesperson to role-play a dermatologist visit for roughly 20 minutes: why are you taking time from my patients, and why should I listen? The exercise exposed a sales process that spreadsheets obscure because “it’s very easy to think that sales just happen.”
3. Compliance belongs inside expert selection, not after it
Walker initially described interviewing a company’s dominant customer; Fokin immediately refined that to a former employee of the customer, perhaps someone who left a year earlier. That person might understand the product, procurement process, and relationship without possessing current information that could restrict the investor.
Fokin praised AlphaSense’s conservative compliance process because sourced interviews arrive with a review record. The benefit is not simply access: it is knowing that the expert was screened to avoid an evident legal or compliance problem.
The same discipline applies to every call. The goal is not to extract whatever the expert happens to know, but to obtain thesis-relevant evidence while respecting confidentiality agreements, NDAs, and the boundaries established by the platform.
4. Conflicting customers often reveal more than a clean consensus
Walker’s anonymized implant example looked decisive on paper: an FDA study showed roughly a 1% defect rate for the new device versus 10% for older products, and a defect could require another surgery. He imagined salespeople presenting that as a 10x safety advantage and asking whether a doctor wanted to take on medical-malpractice risk.
Surgeons supplied the missing caveat: the older products’ FDA studies dated to the 1990s, while techniques and processes had improved substantially since then. Their own surgical experience led some to believe the legacy devices were now comparably safe, showing why company messaging, sales execution, and customer belief must be investigated separately.
Fokin found a similarly wide range around Sofwave’s FDA clearances. He believed the number was 9 and rounded it to almost 10 for simplicity. Some doctors valued indication-specific clearance because it reduced perceived liability; others considered it “just a marketing buzz” and were comfortable using devices off-label.
His change with experience was to stop expecting every doctor to return the same answer. “Nothing is certain,” and the point of multiple calls is to map the distribution of opinions, then incorporate that range into the investment judgment.
5. Former employees are witnesses whose credibility must be tested
Walker’s pushback—worth keeping—is that former employees may carry layoff bitterness, while highly successful alumni may remain unrealistically positive even as the business deteriorates. A small sample can contain two satisfied formers and one person for whom a one-star review would be “too high.”
Fokin’s disarming analogy was to former romantic partners: separation does not automatically make their assessment false. More concretely, he asks laid-off employees whether they would return if invited and whether they would recommend the company to a sibling, nephew, or niece; the answers become indirect evidence about culture.
His legal training supplies the standard: investors probably cannot establish facts “beyond a reasonable doubt,” but they can seek a “preponderance of evidence”—a more-likely-than-not case assembled from employees, customers, consultants, management commentary, and observable performance.
Context determines the weight. Because Fokin generally believes unhappy employees and poor cultures do not produce excellent products, a weak product makes a hostile former more credible, while enthusiastic customers and strong products make that testimony less weighty. When someone says the culture was “toxic,” his response is: “Could you give me an example?” Then he asks for another.
6. Screening questions can create value before a call begins
Fokin writes his own three or four screening questions and does not waste one on career history already visible in the biography. He first wants to know what the person actually did day to day, because titles alone can conceal whether the expert touched the thesis-relevant work.
Open-ended questions sometimes generate several revealing sentences before the interview. For a “proverbial needle in a haystack,” he also lists functions—marketing, sales, supply chain, procurement, and others—and asks candidates to rate their knowledge from 1 to 10.
The credible pattern is many ones and twos plus one or two eights or tens. Someone claiming expertise across every function may know little; “marketing, sorry, I have no idea, but supply chain management, nine out of 10” signals both specificity and intellectual honesty.
Interviewers must also respect functional limits: do not ask a salesperson to explain accounting or a marketer why stock-based compensation is high. If Fokin realizes midway that an expert is weaker than expected, he may use the remaining time for broader questions, but the core questions should match what the expert can reasonably know.
7. Historical interviews can separate temporary tailwinds from durable change
Crocs illustrated expert calls as a guided corporate history. Around summer 2022, Fokin recalled the stock trading near 5.5–6 times 2022 earnings, implying either that “the company is going out of business soon” or that it was badly mispriced.
The central question was whether COVID had created a temporary sales spike that would collapse. Former employees—including some who had left before COVID—described internal changes the company had made before the pandemic.
Fokin’s conclusion, explicitly based on his research and therefore fallible, was that Crocs had built a strong foundation before COVID and then used the stay-at-home, casual-wear tailwind to accelerate. The foundation “was probably not disappearing,” making the historical sequence more informative than the headline COVID exposure.
His analogy was a medieval castle tour led by a local history major: a knowledgeable guide connects changes over time that a visitor would not discover alone. Expert calls can provide that chronology even when their subjects lack current operating information.
8. AI accelerates digging and analysis, but judgment remains human
Fokin organizes the research process into Paul Enright’s three stages: “digging, analyzing, deciding.” As of August 6, AI delivered most of its productivity benefit in digging, some in analysis, and “zero in deciding”—a position he expects could change as both the technology and his own practice evolve.
He also distinguishes doing an old task much faster from doing something previously impossible. He has found many examples of the first but “not figured out use cases for the second one yet,” an unusually candid boundary amid broad AI enthusiasm.
In one research example, an old VIC write-up said a company operated in 10 states. Instead of recording a question and later searching filings, Fokin asked AlphaSense for the current count, the history of state expansion, and any relevant revenue breakdown; within seconds, he had the outline and its sources.
Walker offered compensation analysis as the same compression: comparing incentive structures across three proxy statements once required hours of opening, scrolling, and reconciling disclosures. An AI tool can create the first comparison in roughly 10 seconds, leaving the investor to verify and interpret it.
9. Grid and deep research make document overload navigable
AlphaSense’s Grid is Fokin’s favorite feature: it places documents—expert calls, earnings transcripts, filings, or proxies—against as many as 12 recurring questions. He builds templates targeted at products, company strategy, competition, risk, and other research needs; users could also tailor them to SaaS, industrials, consumer, or other styles.
Across 20 expert calls, most with customers, questions such as value proposition, purchase trigger, sales-cycle length, and evaluated alternatives can expose the full opinion range within minutes. The answers also identify the most thoughtful interviews and the yellow or red flags that deserve full, page-by-page reading.
Fokin may still read the other calls to ensure nothing was missed; Grid prioritizes attention rather than replacing source material. His understanding—offered with a non-technologist’s caveat—is that specialized vertical AI also reduces prompt sensitivity by improving a user’s imperfect request behind the scenes.
Walker uses the same structure to challenge earnings excuses: if management blames a “soft consumer,” he compares three peers’ results and commentary before the follow-up call. If peers report strength, management must explain what is company-specific instead of hiding behind the macro narrative.
10. The best AI workflow prepares the mind before primary-source reading
Fokin’s Deep Research prompts can run roughly five pages and return a sourced 30–40-page output covering customers, segmentation, pricing, history, and other recurring questions. He prints it, reads with a pencil, and marks highlights, stars, and margin notes.
The analogy is hearing an hour-long interview with an author before opening a 300–400-page book: the overview gives new facts somewhere to land. Once “my mind is prepared,” Fokin reads the most important expert calls, filings, earnings transcripts, and conference materials with better comprehension and retention.
He had not yet tested AlphaSense’s newly released AI-generated expert calls, so his view remained deliberately open: he was “really curious,” but earnings season had prevented experimentation. That distinction between awareness and actual evidence runs through his broader AI stance.
Feature discovery is itself a research process. Fokin credits “90 or maybe even 95%” of his AlphaSense knowledge to account manager Amar Capellan, including periodic 30-minute reviews every six to eight months; conversations with product managers reveal less-obvious use cases while giving builders customer feedback. His Kasparov-inspired operating belief remains: “Human plus machine is more powerful than machine and definitely more powerful than human.”