AI is commoditising single-stock research, so the edge is moving to long-term theme analysis

In July, Daniel Gladiš of Vltava Fund told his shareholders that the greatest danger of AI in investing is that it will make investors think alike. We agree with him. AI is turning single-stock research on public information into a commodity. The investors who stand to gain most are those running concentrated portfolios on long-term themes, because working out how a theme plays through an industry, and which companies capture it. This post sets out why, drawing on conversations with portfolio managers over the past few weeks, and how Finn supports that work.

AI has made analysis cheap for everyone

Reading filings and transcripts, building a first model, comparing a company with its peers, and writing the bull and bear case are all things AI tools now do well enough for a first pass, at a cost close to zero. Finn does this, as do many other tools. Adoption has been rapid, but because everyone does it, it rarely leads to better returns.

At large buy-side firms, 70% now use AI in the front office. Among asset managers, 8% name improved returns as a benefit.

The 70% comes from SimCorp’s January 2026 survey of 200 buy-side firms with more than $10 billion each, up from about 10% a year earlier. The 8% comes from Mercer, which surveyed 131 asset managers in early 2026. More than half of them have AI in at least one strategy’s investment process, but 6% use it for decision-making, and 69% named operational efficiency as a benefit. That is what you would expect from tools built to read and summarise faster, since faster reading reaches every investor at the same time.

When everyone has the same analyst, views converge

Gladiš puts it this way: if professional investors use the same tools, trained on similar data and built to solve similar problems, the likely result is a convergence of opinions. He expects original thinking to become rarer once everyone has their own AI assistant, and the value of independent judgment to rise for that reason. Michael Steinhardt called what he was looking for a variant perception, meaning a well-founded view that differs meaningfully from what the market believes. Analysis that starts from the same public information and the same questions is becoming a poor source of one, because everyone else can now produce it too.

What stays with the investor are three things that do not converge:

• Information others do not have, such as a broker note, a conversation with a supplier or a site visit.

• A mandate others do not have, which usually means a longer horizon and the ability to sit through a drawdown while the thesis still holds.

• The investor’s own record of reasoning, meaning what they believed, what they decided and why, and how it turned out.

All three pay off only if they are used deliberately. A valuable research report left in an inbox gets read once. A long horizon helps only if, in the middle of a drawdown, the investor knows whether the thesis still holds. A record of reasoning teaches nothing if the reasons were never written down.

Themes are where this matters most. A theme carries more claims than a single stock, takes longer to play out, and its evidence is spread across more companies and sources. It is also where a long horizon gets tested, because holding through a drawdown is only an advantage if the investor can tell whether the theme still holds.

Where investors actually lose money

The research on how professionals make decisions points to holding and selling as the weak spot. Akepanidtaworn, Di Mascio, Imas and Schmidt looked at 783 institutional portfolios averaging $573 million, and 4.4 million trades.

Buying decisions showed skill. Selling decisions did worse than selling at random, at a cost of about one percentage point a year.

We went through what that looked like in one book in The only alpha that compounds. Essentia Analytics, which sells this kind of analysis to fund managers, reports that its clients were losing about 94 basis points a year to behavioural patterns, most of it from holding losers too long. Terrance Odean, after decades of studying why investors hold on to losers, advises writing down the selling rule before the position is opened.

Investors who work from long-term themes face this over a longer period and across more companies. A theme can turn out to be right and the investment can still lose money if the investor holds companies that do not capture it or pays too much for them. Morningstar’s guide to thematic funds sets out three conditions that have to hold at once:

1. a theme that is real

2. the stocks that actually capture it

3. a price that has not already discounted it

Over a holding period of five or ten years, each of the three can change more than once, and the investor has to notice when it does.

In his April letter Gladiš borrowed an idea from running, that the most likely next injury is the same as the last one. In investing, he wrote, the most likely next loss comes from the same weakness in the process that caused the previous one, and unless that weakness is identified and corrected it will cause the next loss too. An investor can only find that weakness if the reasons behind past decisions were written down when the decisions were made.

Analysis that starts from your own view

Most AI tools for investors start from a ticker and the public record, and when asked the same question about the same company, they give every investor much the same answer. We think the better use of AI starts from your own view and the material only you have, and builds the research around that. Two investors looking at the same theme then end up with different analysis, because they started from different views and different information. That analysis also has to stay current for the two or three years a theme takes to play out, and that is the part most investors never find the time for.

The work starts with you writing down three things about a company or a theme:

• what you believe

• what would prove each belief wrong

• roughly when you expect to know

Then the evidence arrives in fragments: a quarterly filing, a comment on a call, a broker note, a price move, a budget request in February. Each fragment bears on a specific belief, and none of them arrives labelled that way. So you read it as news, file it, and once a quarter try to reconstruct in your head whether the thesis still holds. That reconstruction is where good theses get dropped too early and broken ones get held too long, because nobody has the time to do it for thirty names at once.

An agent can do that reconstruction every day. It can read each filing, results release and broker note against the beliefs it bears on, record which way the evidence points and where it came from, and keep each belief in one of four states: untested, supported, contested or broken.

A worked example: the space launch theme

The screenshots below are from the version we are testing, on a theme about which companies win the race to make satellite launches cheaper and more frequent.

Beliefs on the space launch theme. Market, filing and industry data in these screenshots are real. The investor’s notes and decisions are example data.

The theme rests on five beliefs, each written as something that has to be true for the thesis to work. The second, that the share of launches on previously flown boosters keeps rising, is supported. In the first quarter of 2026, 39 of SpaceX’s 40 Falcon launches used a flown booster. The first, that Vulcan, New Glenn, Neutron and China’s reusable rockets fly repeat missions without long stand-downs, is contested, with six pieces of evidence against it and one for. Finn drafted these states and the investor accepted them, and the page shows who wrote each line. Every piece of evidence links back to its source and date.

Each belief is measured by a few numbers, and the investor sets a line for each one: the reading that should raise a flag, written in their own words. On Rocket Lab’s operating margin, for example, the line reads “Flag if the margin widens for a quarter while revenue grows.” Finn follows the numbers, alerts the investor when a reading crosses a line, and shows the line next to the new reading.

Key numbers on the same theme, each with the belief it tests, the latest reading, the investor’s line and the date of the next reading.

The theme also has a timeline. It runs from Starship’s first orbital attempt in September 2026 to early 2028, and each of its three phases tests one thing:

1. whether the delayed rockets fly again

2. whether they fly repeatedly and customers deploy on schedule

3. whether a second provider has reached a steady cadence

The dated events that can confirm or break a belief sit on the same timeline, so the investor knows ahead of time when each belief should start to show in the data.

The timeline for the same theme, with the dated events that can confirm or break a belief.

With the evidence tracked belief by belief, the investor can see whether the theme still holds on the day a drawdown tests it.

Tom Slater at Baillie Gifford, who runs Scottish Mortgage, describes the manual version of this discipline. Every holding gets two questions all the time: has the opportunity got bigger or smaller, and has the likelihood of capturing it gone up or down. He also names the pressure:

“You will be tempted to chip it away in the name of risk control.” Tom Slater, Baillie Gifford

A written break condition, meaning the evidence that would prove a belief wrong, and a record of past decisions help an investor notice that pressure while it is happening.

How Finn supports long-term thematic analysis

Finn already does the research. It reads filings and transcripts, follows the companies you care about, and answers questions by email and WhatsApp.

The themes page, sorted by what needs your attention.

Each theme and each company gets a page with the same four tabs:

• Story holds the thesis in your words, the beliefs and their states, who gains and who pays, and the timeline.

• Tracking shows what changed since your last visit, the key numbers with your lines, the upcoming dates and the decisions waiting on you.

• Companies lists the companies in a theme. On a company page this tab is Peers, listing its competitors, customers and look-alikes.

• Research holds the full dossier, the data tables and the filings Finn read.

A company page adds your value estimate and price levels, a card before each results report that sets out what each belief needs to see, and your house rules checked against the company.

Finn companies, board view.

Companies sit on a board with six stages, from Ideas through Watching, Working, Ready and Held to Passed and exited. Moving a company to another stage always needs a reason, and passing on a company or exiting it also needs a category, so the reason behind each decision is written down when you make it. Next we are adding a review for when a position closes, where Finn drafts a comparison of the original thesis with what happened and you write the verdict.

We are running this with a small group of portfolio managers now. If you run a concentrated portfolio and want to join them, request access.

Sources

• Daniel Gladiš, Vltava Fund, “When artificial intelligence becomes abundant in investing, what will become rare?”, letter to shareholders 2/2026 (Jul 2026).

• Daniel Gladiš, Vltava Fund, “Sport and investing (once again)”, letter to shareholders 1/2026 (Apr 2026).

• SimCorp, InvestOps Report 2026 (Jan 2026), survey of 200 buy-side firms conducted by WBR Insights.

• Mercer, survey of 131 asset managers on their use of AI (Feb to Mar 2026).

• Michael Steinhardt, No Bull (2001).

• Akepanidtaworn, Di Mascio, Imas and Schmidt, “Selling Fast and Buying Slow,” Journal of Finance (2023).

• Essentia Analytics client data as reported by Markets Media (vendor-reported figures).

• Terrance Odean, “Are Investors Reluctant to Realize Their Losses?”, Journal of Finance (1998).

• Morningstar, Guide to Thematic Funds (Apr 2022): the three conditions for a successful thematic investment.

• Tom Slater, Baillie Gifford, interview on Motley Fool Money (Sep 2025).

• FinVerBench (arXiv, 2026) on language model accuracy in financial statement verification.

• Howard Marks, The Most Important Thing.