MicroCapClub Sep 21, 2026 28m 14m saved
With Michael Fritzell, Founder of Asian Century Stocks
Getting up to speed on an unfamiliar company used to take Michael Fritzell three to five days. It now takes him less than one, and sometimes a few hours. His response is to cover 200 stocks a year instead of 20.
The standard version of that story ends with better returns. Fritzell's does not, and he says so in the first thirty seconds of the conversation.
"I'm not going to necessarily going to make more money, it's just that my focus has shifted a bit."
Fritzell is Swedish, spent 16 years on the buy side across Hong Kong, Singapore and Indonesia, and left in 2021 to write full-time. He now publishes 40-page deep dives on Asian companies almost nobody covers, pays for insider-transaction data across Asia-Pacific, and, on his host's count, reaches more than 20,000 readers. The conversation was recorded live in the MicroCapClub community on 2 September.
The full interview is covered here so you can skip it. 28 minutes of audio, 14 minutes of reading.
Here are the 9 lessons that matter.
Key Takeaways
He thinks AI is making analysts smarter, not dumber — the constraint it removes is having to rely on someone else's writing
Structure is the thing AI cannot supply: "if you don't know what matters, you can get lost"
His benchmark ran 10 analyst prompts through Grok, Claude, Gemini and ChatGPT on a company he already knew cold
Two jobs he will not delegate: calculations and writing — the first because the output is probabilistic, the second because readers stop trusting AI prose on sight
He stopped buying low P/E stocks and started buying a prospective IRR, selling when it falls under a 10% cost of capital
His edge is now insider-transaction data and broken IPOs, on the assumption that idea generation itself gets commoditized
A saved "red flags" prompt runs revenue recognition, depreciation and related-party history before he reads anything himself
Prompting well is almost entirely about supplying context — hand it the annual report rather than trusting recall
Faster research does not mean better returns, it means covering 10x more names to find the same surprise
1. AI Is Making Us Smarter
The episode opens with its own title as a question, and Fritzell disagrees with it immediately. His argument is about what the tools remove rather than what they add.
The constraint they removed was other people's work
In the past, I would have to rely on someone else's writing, whereas now I can dig into any topic as deep as I want.
Michael Fritzell
What makes it work for him is that the thing does not get tired.
An analyst who never stops answering
Because I can continuously ask ChatGPT AI, it's tirelessly answering my questions and I can really get a deeper understanding of any topic. And to me, this is quite revolutionary.
Michael Fritzell
His measured result is that his understanding of businesses has improved, and that it improved faster. Asked in what situations it is wrong to outsource something to a machine, he did not name a task — he named a prerequisite.
The thing the tool cannot give you
I think you need structure, right? And if you don't know what matters, you can get lost.
Michael Fritzell
2. Benchmarking Four Models
Fritzell ran a comparison and published it, in a post he dates to late 2025, with the caveat that the ranking has a short shelf life — the models trade the lead every three or four months.
The test he built
my benchmark test was me signing up for these four biggest and most popular tools Grok Claude Gemini and ChatGPT. And then I did 10 different prompts, which I think might be useful to equity research analysts and investors.
Michael Fritzell
The control was his own knowledge. He ran the prompts against Carabao, a Thai energy-drinks company he already understood, so he could tell whether an answer was any good. He scored on two things: depth — was the answer exhaustive — and brevity, because some of the tools will pad an answer out with text that carries nothing.
Where it landed at the time
at the time, I found that ChatGPT was probably the stronger of these models. Followed by Gemini.
Michael Fritzell
Then Claude shipped an Opus release and the ranking moved.
And where his own usage went
And since then, Claude has taken market share. And I've also shifted part of my usage to Claude.
Michael Fritzell
3. The Saved-Prompt Stack
The productivity claim underneath everything else is a straight time measurement.
Three to five days became a few hours
given that it used to take, you know, 3 or 5 days to get up to speed to a given company. Now, it takes less than 1 day or maybe just a few hours.
Michael Fritzell
His conclusion from that is not that the work gets easier but that the bar moves: analysts will have to turn over far more rocks, keep track of ten times more companies, and still go further than the next person.
The first pass is deliberately fast and generic. If somebody he respects mentions a ticker on Twitter, it goes into a quick model the same day.
The opening question on any new name
typical use case would be for me to use Gemini and just ask how does this company make money in simple terms?
Michael Fritzell
He is looking for the segments, the margin shape, any competitive advantage and the growth potential — enough to decide whether the name is worth his own time.
The second pass is a saved prompt. Claude lets him keep a project with a standing instruction, so he drops in a ticker and gets a fixed output.
The bull versus bear project
I call bull versus bear. I ask Claude to give me the bull cases and the bear cases for any given stock
Michael Fritzell
The answer comes back as roughly five points a side, drawn partly from investor commentary, and his job is to sort which of them are serious and which are trivial or temporary.
4. What He Won't Delegate
The reason for the boundary is a property of the technology rather than a preference.
These are not deterministic systems
they are basically probabilistic computing and the output isn't 100%.
Michael Fritzell
With ordinary software you know what a given input produces. With generative models you do not, and they sometimes hallucinate.
The two exclusions
for that reason, I would not use them for calculations. I would also be careful with writing.
Michael Fritzell
Inside the boundary, he is enthusiastic and specific. Brainstorming is the use he rates highest — asking which companies in the Korean defense industry have the most exposure to consumables, and questions shaped like that. Search is the second.
Obscure data is where he thinks the tools are underrated
I use specifically ChatGPT to find very obscure data.
Michael Fritzell
His example is finding a detail inside the transcript of a Japanese-language video, which he says a search engine would never have surfaced. Summaries are the third use: he opens a sidebar next to a web page, asks what is on it, and decides from the answer whether the page is worth half an hour.
And a fourth, where the model's character is the point
Claude, I think is quite good at proofreading, finding issues, finding errors. It's a lot more critical.
Michael Fritzell
He pays for all of the tools, and picks between them by temperament: ChatGPT to go deep on a subject, Claude to attack a draft for factual inaccuracies.
5. Writing by Hand
Writing is the exclusion he feels strongest about, and the reason is commercial rather than technical.
Readers stop trusting a page the moment they recognize the register
I think people the moment they see AI-generated writing, they stop trusting that it's considered and high quality.
Michael Fritzell
He adds that the oversupply is part of it — there is simply so much of it now that readers have learned to stop.
What a reader is actually buying
People want to see someone's voice and they want to hear that you have actually spent some time thinking about the topic. So, writing is something I definitely do not use these tools for.
Michael Fritzell
The host's contribution here was that everyone has built the same reflex very quickly.
On how fast the detector formed
how quickly we all sort of build a automatic intuitive sense for what's AI slop, quote unquote, and what isn't
MicroCapClub
He also raised a tweet of Fritzell's from a couple of months earlier showing an AI checker scoring his writing as entirely his own. Fritzell allowed that everyone may end up writing with these tools eventually, but only once a model has been trained on the writer themselves, and said the same caution applies to images.
6. From Low P/E to IRR
The change AI made to his business, he says, was less dramatic than the change it made to his process. On the business side it taught him skills fast: he asks a model for the theory first — how a newsletter converts free subscribers to paid — then asks it to apply that theory to his own situation.
He was frank about the result. Subscriber growth has not been strong, and he has been trying different methods. What it does give him is a counterpart.
Investing is where he says the improvement is real, and the specific change is what he underwrites. He used to buy low price-to-earnings multiples. He now buys a prospective internal rate of return.
The reason is structural to the region.
Why a cheap Asian stock stays cheap
The issue with stocks in Asia in particular is capital allocation. Almost always.
Michael Fritzell
Cash piles up on the balance sheet, return on equity collapses and the stock becomes a value trap — or the management team takes resources out through related-party transactions.
The arithmetic that killed the low-multiple approach
if there's no growth and nobody cares about the stock for 10 years, your IRR will be pretty terrible.
Michael Fritzell
A seven-times multiple that should be worth fourteen is a double only if somebody eventually cares. So he now wants a company growing around 10% a year, where the floor return is roughly that growth rate, and any re-rating of the multiple adds another 15% to 20% on top.
The rule, on both sides of the trade
I now try to buy stocks at 20% for you know a few prospective IRR. And once they get below 10% which is kind of like a typical cost of capital, I try to sell the stock or look for better ideas.
Michael Fritzell
He admits he does not always follow it. The host's point was that it gives him a selling discipline and not just a buying one, and Fritzell agreed that the alternative is a known failure mode.
The behavioral case for a sell rule
There is a real risk that you fall in love with the stock.
Michael Fritzell
The mirror-image risk, he added, is selling anything that goes up quickly. Reassessing the position and updating the forecast is what keeps the decision forward-looking rather than anchored on the price.
7. Ideas AI Can't Commoditize
Asked where AI helps most in idea generation, Fritzell's answer was that idea generation is exactly the part he expects to be commoditized, so he has moved away from it. His edge now is data he pays a lot for.
Where he looks first
I really emphasized insider transaction data.
Michael Fritzell
He buys expensive insider-transaction data across Asia-Pacific and writes up the biggest transactions at least once a month, frequently investing in them too. The appeal is the asymmetry of attention.
Nobody is watching and somebody inside is buying
There aren't many active investors in these stocks, yet insiders are seeing something.
Michael Fritzell
That is where a model earns its place — not finding the idea, but explaining it.
The follow-up question he puts to a model
You can even ask Claude or ChatGPT, why might insiders be buying? And you'll get surprisingly great answers from them.
Michael Fritzell
His example of an answer that paid off was a new factory two months from completion, which is what turned up behind one of his recent Indonesian deep dives.
His second source is the broken IPO: a company that listed two or three years ago and has fallen since.
Why the failed listings are a growth-stock hunting ground
newly IPO companies are often run by first-generation entrepreneurs who are still keen to grow the business.
Michael Fritzell
The method is mechanical — pull the list of IPOs from one or two years ago, then run the business models and growth rates through Gemini or ChatGPT. And the third source is other people.
He was blunt about the third one
I'm not going to lie, I also steal ideas from others on Twitter. Or on Substack.
Michael Fritzell
He says he knows which individuals are worth reading, and that following them is a way of making sure the time goes somewhere. The host's amendment was that publishing something in public makes it borrowing rather than stealing.
8. A Red-Flags Prompt for Asia
Due diligence in Asia carries the governance problem from section six, so Fritzell built a second saved project for it.
The forensic prompt
I have a prompt for Claude to find red flags in the financial accounts.
Michael Fritzell
It covers revenue-recognition policies and whether they have changed, depreciation policies, related-party transactions past and disclosed, and the track record of the people making the decisions. He opens the project, types the ticker, and gets what he calls a pretty decent report.
The usual effect on his mood
What typically happens is I get really bearish when I see the answers.
Michael Fritzell
None of that replaces the work. He still writes the full piece himself, and the length is the point.
Why the deep dives stayed long
These are 40 pages long.
Michael Fritzell
His view is that every side of a business has to be covered before a conclusion is available, and he keeps a saved prompt for each aspect of a business. He publishes those prompts on his own site, and the host offered to link them.
9. Why 200 Stocks, Not 20
Asked whether an AI model could run the whole investment process ten years out, Fritzell said no, and located the reason one level up from the tools.
The abstraction layer moved; the judgment did not
AI tools can do a lot, but we still need to know what actually matters.
Michael Fritzell
A framework — value, growth, whatever it is — is still what decides the outcome, and a model will not save an investor chasing momentum in a bear market. Finding information quickly and making the right decision are different problems.
His evidence that the second one has not improved is the market itself.
Three years of AI and no more rationality
it's not like people have become more rational.
Michael Fritzell
He pointed at Korean stocks earlier in the year as the recent example of momentum ending badly, and described US tech as a very crowded space. Competition rises for anyone who does not use the tools, which is what produces the shift he opened with.
The whole trade-off in one sentence
Instead of me doing research on 20 stocks per year, I'll do research on 200 stocks. I'm not going to necessarily going to make more money, it's just that my focus has shifted a bit.
Michael Fritzell
The host put the obvious question straight back at him.
The pushback
And if you're not making more money, why not just go back to looking at 20 stocks?
MicroCapClub
Fritzell's answer was that going back is not on the menu, because the competition does not stand still.
Where the variant view has gone
it's going to be really hard to find differentiated views, you know, or variant views on popular stocks. Microcaps might be different because people aren't paying attention to them as much.
Michael Fritzell
The money, on his account, is made by finding a stock where expectations are wrong and then watching them get beaten — a new product, a regulatory change, an expansion into a new region. Something has to surprise. He expects that to get harder, and thinks the universe is large enough to absorb it.
The consolation
10,000 stocks listed in Asia, there are plenty of opportunities.
Michael Fritzell
Bonus Insights
Prompting well is mostly supplying context
the most important thing is to give it context, whether it's a context of what you're looking for or the data that you want it to pull from.
Michael Fritzell
Hand the model the annual report or the source material and the numbers come back right; he says the tool itself barely matters, naming NotebookLM and Claude in the same breath. Specifying the output format is the other time-saver he recommends, and he finds ChatGPT wordy by default.
The theory-then-application prompt pattern
For a business problem he asks for the theory first and only then asks the model to apply it to his own situation, on the view that the structure has to exist in his head before the advice is worth anything.
Working alone is the part AI actually fixed
Writing for himself, he has nobody to shoot ideas back and forth with. He is explicit that the model is a computer and that having something to argue with is still worth having.
The models have personalities, not just scores
The host's framing — that these tools have distinct strengths and weaknesses rather than being better or worse versions of each other — is the one Fritzell's whole workflow runs on: a fast model for the first pass, one model for depth, another for criticism.
Fritzell's bottom line is that AI has collapsed the cost of understanding a company and left the hard part untouched: you still have to know what matters, and the only durable edge is looking where nobody else is.
Products, Companies & Tools Mentioned
ChatGPT, Claude, Gemini and Grok (The four tools in his benchmark; ChatGPT led at the time, Claude took share after an Opus release, and he now splits usage by task)
Carabao (The Thai energy-drinks company he used as the control in the benchmark, because he already knew it well enough to grade the answers)
NotebookLM (Named alongside Claude as proof that the tool matters less than whether you gave it the source material)
Google Chrome (Where he reads: a sidebar opened with a keyboard shortcut summarizes a page and tells him whether it is worth half an hour)
X and Substack (Where he borrows ideas from a short list of people he rates, and where he first hears most new tickers)
MicroCapClub (The private community the conversation was recorded in, whose members have profiled 1,500 companies since 2011)
Books & Resources Mentioned
Asian Century Stocks – Michael Fritzell (His newsletter: 40-page deep dives on Asian companies, a monthly insider-transaction report, and the saved prompts he says are published on the site)
Listen to the full episode
🔴 YouTube | 🔗 Episode page
Watch the full episode:
If this was worth your time, send it to someone closer to the industry than you are.
Get the latest market chatter as it happens: