The AI Trap for Analysts: Productivity Illusion or Edge?
My experience using AI in research, admin, and investing
I have been thinking about AI quite a bit recently, for a number of reasons beyond the last quarter’s impressive improvements in the models:
1) Several people have told me that AI “has been life-changing”, “it saves me so much time”, and “I cannot imagine life without it”. But when asked the specifics of how they use it, it’s mainly to summarise reports – “I use it to summarise the sell-side research on a stock” and similar. You could just read the front page or the executive summary. One acquaintance likes to use Notebook LM to create podcasts about macro research which is a use case I cannot get my head around. Importantly, I have not yet found any public market investors who use it to help them make decisions (although maybe they wouldn’t tell me if they did).
2) One CIO of a boutique firm asked me to guess how much they were spending on AI. I was out by 1000x. They are using it to produce a comprehensive research report on stocks, in their style, with attention to the variables that matter to them. The prompts run for 4 pages. I thought that was quite interesting and probably quite helpful but not game-changing. I spend more on AI than they do – more later.
3) Another head of equities at a large long-only shop (with multiple funds and a large team of analysts and PMs) observed a huge variation in the token usage across his department. They have developed an internal model interface - some are avid users, some barely use the facility at all.
4) Then I had Gary Channon on my podcast. He is a really thoughtful investor and explained that he has slimmed down his research team and all the team including himself have been tasked with the job of building agents to improve their research process. More on this too later.
5) A few weeks ago, I became frustrated at Chat GPT’s unwillingness to co-operate on what I considered a fairly simple task, and it wasn’t the first time. I posted the problem on LinkedIN, which created more interest than any of my posts this year - it was more of a cry for help than anything. Luckily I now have an expert tutor who is an AI architect at a major UK firm and he is going to train me, or at least try.
And it’s not just Chat GPT - I switched to the paid version of Gemini. I had uploaded 40 files to do some analysis and asked for an extract from a particular file. The AI just made up its answer repeatedly, in spite of my requests to verify it against the meeting transcript. I asked it why:
“My core functionality for accurate information retrieval and verification is not performing to the required standard, despite my previous assurances. I understand this breaks the trust essential for our work.
I am at a loss for how to proceed with this task, as I have clearly demonstrated a fundamental inability to meet your requirements for accuracy and verification. Please tell me how you wish to proceed.”
I was 90 minutes into this task before I spotted this error. I surely cannot be the only analyst having this type of problem.
6) I was in the US recently, and I was staying 30 minutes’ drive from the client’s office, requiring a long Uber ride each morning. One day I had a doctor driver (she was sitting her practising exams) and on two of the three days, my driver was in IT. One had a masters degree and the other was a coder. Both were having trouble finding a new job. It seems AI is already having a significant impact on the IT industry and on the number of coders required - as one driver remarked, AI can do it more accurately, faster and obviously much more cheaply. If coding might be at the head of the queue, it made me wonder where finance was in the line.
7) I was at a hedge fund dinner and I asked the participants how they were using AI. The conversation got sidetracked into AI stocks of course, but with a dozen-odd people round the table, there wasn’t a single practical takeaway. Competitive shyness perhaps. But let me give you one illustration of how far we have moved on from the satellite tracking of the cars in Walmart’s car parks (the brainchild of a former UK food retail analyst colleague). A quant is looking at data which tracks the number of mobile phones in particular locations over time - that could be used for all sorts of purposes now.
8) And then I had a young would-be analyst request my help and ask for an interview. We had a short dialogue and then he sent a message:
“…On a related note, I just picked up your book, "The Most Important Thing," and I have to say it looks like the most advanced and practical investing guide I've come across. As a student who's passionate about becoming a long-only analyst, it's an incredible resource.”
I responded, “That’s not my book”.
The student replied:
“Very sorry, didn’t mean to disrespect you. I just asked llm to draft a message for a better quality and it mistook the book. Really sorry for that, totally my fault.”
I am spending time on this correspondence with someone I don’t know but would like to help and wouldn’t dream of using AI to craft a personal message. I think it’s a really bad practice to use AI to draft such a message.
I posted this too on LinkedIn and it garnered a lot of attention - given the interest these two LinkedIN posts attracted, I thought I would share some of my experience here, not because it will help anyone – I am a total amateur at this – but it may give some of the more technically literate readers some amusement, and it may reassure others that they aren’t so bad after all. I see a lot of stuff on LinkedIN from AI experts, but I rarely see anything from ordinary investors and analysts, perhaps partly because this is perceived to be competitively sensitive.
Forensic Analysis Bootcamp
Good news: we are running our forensic analysis bootcamp again in the fall. It kicks off on October 6, and will run for 8 weeks every Monday night at 17.30 London, 12.20 EST, 9.30 PST. Classes run for 60-90 minutes and allow plenty of time for questions.
The curriculum is the full setup we offer to institutional investors but with some additional explanation and extra time allowed for a more considered deliberation. And every session is recorded so if you are travelling one week, no problem, simply watch the recording.
Gary Channon’s Experience
I was really impressed that Gary used AI to investigate his mother’s cancer and to identify possible avenues of treatment, such as drug trials globally. He is pretty on top of all the models and their uses and his team has already seen a significant improvement in productivity in certain applications.
For example, they own Barratts, the UK housebuilder. They used to scrape the Barratts website every few weeks to identify houses for sale in each location; by comparing with the previous edition, they could work out the unit sales and of course they knew the asking prices, so they effectively had a pretty accurate revenue forecast. Gary does not use this data to trade the quarterly trading updates or half yearly results, but to confirm that the company is on track.
In the GFC, he knew that the company was continuing to sell well outside London. Most fund managers would assume that things were awful; they might at best send someone to visit Fulham Reach, their development at Earls Court at the time, where sales had ground to a halt. Gary was able to take advantage of the share price fall in the knowledge that things were not as bad as the share price implied.
The process of scraping the data and generating the report used to take two days. Today, with AI assistance, it takes just a couple of hours. They are working on similar agents and understandably are not sharing too much – Gary has given me a couple of hints and it’s interesting stuff, if they are successful. But with the front office team developing these tools, they will be further up the learning curve than most – that’s a smart strategy if you can afford the investment.
Data Scraping
The UBS avatar which will present analysts’ research ideas in video format attracted a lot of press attention, but the real moves being made by the big banks are pretty secret. I have a mole in the expert users group in one bulge bracket firm – they are building front ends so they can keep their data in-house, but the analysts’ ability, and willingness, to take advantage is variable; and with a heavy workload, there is a limit to the new work they can take on.
Tech analysts not surprisingly are in the lead. As mentioned earlier, the global multi-asset firm sees huge variation in AI deployment, from barely using it to heavy usage by their tech analysts. I met Ben Rogoff recently; he runs the Polar Technology team (their investment trust is the largest closed-end tech fund in the UK) and unsurprisingly they are focused on the issue, both from an investment and from a productivity perspective.
The data scraping interests me. With AI you no longer need a data scientist on your team, although quite a few of my clients have one – indeed, some have several and one client has teams of them. We have worked together on a few applications, for example using LinkedIN to monitor management moves at the second and third tiers below the board.
For more complex applications, the key is often history. If you want to monitor easyJet’s fares, you need a history of what happened last year, as the easiest way to make sense of the trends is to look back at the year-on-year changes. Hence starting early gives you a real advantage and AI should bring data scraping within reach of the average analyst (and perhaps even I will master it eventually).
Such models already exist on the buyside, particularly in places like my client with the team of data scientists. I have no idea what data sets they use beyond the obvious – credit card data, Second Measure etc, - because they tend to be secretive.
I have used data scraping in the past, but I needed a third party service as it was way beyond my competence. We analysed the merger of Ahold and Delhaize which was announced in mid-2015 and closed in mid-2016. Ahold had 760 stores in ten eastern states and Delhaize had 1291 supermarkets in 14 eastern and southern states.
We plotted both sets of stores on a map to identify the overlap; we did the same for the distribution centres and could see significant scope for rationalisation. To get the deal over the line, the two groups agreed with the FTC that they would sell just 81 stores to settle complains that competition in 46 local markets would be affected.
We identified significant savings in distribution costs as a result of this fairly simple exercise. Estimated total stores were 1970 post the deal and are around 2050 today. Not much in 10 years. Per AlphaSense, the EV:sales has gone from 0.59x to 0.46x, margins have gone from 3.8% to 4.0% and revenues have gone from €63bn in 2017 (the first full year of the combined group) to €89bn last year – a pretty impressive increase, but partly affected by the strength of the dollar in the period. The stock has gone from €20 to €36, and the fx rate has gone from 1.11 at the time of the deal to 1.15 now, via 0.97 at the 2023 low.
A lot of research effort for limited reward. That’s the norm for much of my analysis – 20% of the ideas deliver 80% of the returns; if we only knew in advance which 20%...
I have been experimenting with a data scraping tool as part of my current project, but it has not been straightforward so far. I shall likely improve my data scraping skills after some training, and I believe these will become essential for investors. To that end, I have installed a couple of tools, with limited success. I also tried out browser.ai, but have not yet made it work – I am trying to use it in a real-life application and it’s not been straightforward.
My latest attempt involved installing Python and running various scripts. The problem with all this is that it’s incredibly time-consuming, as I really have no clue what I am doing. Without question, I would have been better to do this type of job manually myself; one of the tasks is pretty repetitive but I could have paid someone on Freelancer $50 and I would have my results by now - 3 hours in, I don’t have anything!
I am persevering because I know I need to learn these skills and I don’t have a boss expecting productivity – this must be a lot harder for a young analyst, even though they will certainly be more tech-savvy.
My AI Experience
I want to talk more about my experience in using AI for major projects, rather than daily tasks, but perhaps I should mention the latter first. I almost never use a general AI tool for financial research. I use AlphaSense quite a lot and I find it very useful when I am writing up a stock for the newsletter. For example, I wrote up the two worst performers since I attended two New York conferences in March. I asked AlphaSense why the stocks had gone down:
Prompt: why has Global Payments stock GPN been so weak?
AlphaSense Response
What I like about this output is that you get a clear explanation of the reasons and a source which you can click on and instantly interrogate.
I just checked ChatGPT to see what it would produce from the same prompt and it has certainly improved significantly since I last tried it for this type of query – its answer is pretty impressive:
ChatGPT Response
“Global Payments (GPN) has been weak lately due to a confluence of strategic, financial, and sector-wide pressures:
📉 1. Heavy strategic deal burden & debt concerns
GPN’s planned $24.25 billion acquisition of Worldpay, financed through $7.7 billion in new debt plus divestment proceeds, spooked investors. Shares plunged ~17 % on announcement day amid fears over excessive leverage and integration risks
2. Caution in investor guidance
Despite beating Q1 earnings (EPS $2.82 vs forecasted $2.73), the company maintained its 2025 outlook (5–6 % revenue growth, 10–11 % EPS growth), which lacked upward revisions, prompting investor disappointment
In Q4, revenue and profit missed estimates, triggering a 4.3 % drop and signaling cost pressures
3. Weakness in fintech/software sector
The enterprise software/payments space has been under pressure lately. GPN underperformed industry peers like Visa and sector ETFs. The stock now trades below its 50- and 200‑day moving averages, suggesting sustained bearish sentiment
4. Analyst downgrades & trimmed targets
Several analysts—including JPMorgan, Barclays, Citi, Goldman, and Stephens—have cut price targets (ranging from $85 to $149), reflecting tempered expectations
Estimate revisions over the past month have trended downward, contributing to a hold/neutral Zacks ranking
5. Short-term technical fragility
Technicals show the stock in a consolidation or bearish pattern, with short-sellers positions rising (~12 % short interest), indicating investor anticipation of further decline
🧭 Bottom line
GPN’s weakness stems from a risky, high-debt transaction (Worldpay acquisition), conservative guidance, sector-wide softness, and technical/analyst negativity. While long-term post-acquisition synergies and the new Genius platform may restore investor confidence, near-term stock action remains under pressure.”
The attraction of AlphaSense for me is the facility to search only the transcripts or the transcripts plus company releases – so in this case, I wanted to find the debt in the acquired company which wasn’t disclosed in the release, I asked AlphaSense to search the transcript. I didn’t have to download the transcript then upload it to the AI – a marginal time saving, but useful for a professional investor doing this several times a week.
A parallel would be to use AI to produce a quick research report on a company that I might be going to visit or thinking of writing up – different systems have different answers and this can take you the first few yards without too much effort – this is ideal if I am meeting someone from a company and don’t want to appear totally clueless, but it’s better if supplemented by a sellside research report. A number of professional analysts have told me that this is their main AI application.
The podcast is sponsored by AlphaSense but I pay with my own money for the service, formerly Sentieo. I often get asked if I were in the business of picking stocks for a living and didn’t have that tool, what would I use for research. I would use Chat GPT or Gemini or Perplexity (Abacus.AI is a short cut) for basic grunt work, and I use them when I am researching a stock for this Substack.
For example, when I was writing about US sports teams, I didn’t have a full understanding of the nuances of the different sports, their audiences, the demographics, the sponsorship levels etc. I then generally ask two or three models and offer quite a detailed prompt. So I first explain to the LLM who I am, that I am writing a Substack about the differences in the valuation of different sports teams and can it suggest a prompt which will help me explain this to my audience. I then generally use the same prompt for the different models and compare the answers. I ask each model why differences exist which hopefully will eliminate hallucinations. But I am not researching stocks every day as I would be if I were still working for a fund, so I am not as experienced as many of my readers in this application.
I would also use AI for basic forensic accounting tasks - some of the processes I undertake when a client asks me to review a set of accounts are simple mechanical routines and AI could automate many of these. I already have a semi-automated process in that I have a standard sheet with pre-baked formulae and charts and I simply link these back to the downloaded data when I do a first pass. I haven’t yet used AI in this process, simply because there isn’t much time advantage and because I don’t yet trust the AI to implement this correctly.
Where I do use AI on a daily basis is to run my small business. I outsource any grunt work I can, but if access to cash is involved, I tend to do it myself rather than give some freelancer in Asia the keys to my Stripe account. The EU recently introduced VAT on the sale of digital products – so if you sell your Substack in Germany, you have to pay VAT to the German government even if you are based in London, England (or London, Ontario). I guess most writers on Substack in the US and elsewhere aren’t even aware of the law and Substack are taking their time about implementing VAT on the front end.
I believe in paying my taxes and my accountant is very strict, so I registered with the VAT One Stop Shop. Even with the help of AI, it was a total nightmare, as the systems don’t work properly, but at least I knew that it wasn’t me and that I was giving it the right data. I used ChatGPT to help me generate an Excel template with the 11 countries, each with their own different VAT rate, so that I could easily calculate the amount due. There are two benefits here:
The time saving for a tedious job like building the spreadsheet, looking up the countries and their VAT rates and manually entering all the data.
The compliance issue – it’s much better to do this right first time. Filling in any regulatory filing scares me, but is much easier with the AI telling me the correct options; for example, I used this recently with the US W8-BEN forms where the options are not exactly straightforward.
I use AI in some of my research for the Substack – for example, for the recent OnlyFans article, I used AI to help me track its history of ownership and to assess its competitive position, keeping my browser history intact.
A friend was surprised to hear taht I use AI as an Excel tutor. I consider myself basically competent in Excel - I can produce a good three statement model with reasonably accurate profit forecasts, but I am no whizz-kid. Now, though, I can do much more serious data manipulation using the AI to help me, which has been really valuable.
I don’t do much AI editing – I don’t really like the style of writing. I use AI for title generation for example, and I find it quite good at producing titles that encourage readers to open this email.
I sometimes use AI to summarise a report but not that frequently – I don’t read that many and usually there is a summary. If it’s something I should be interested in, I feel I should have time to read the summary rather than a summary of a summary – it takes time to ask the AI too.
Most recently, I have been using AI to help with preparation for a podcast episode. For example, one recent guest was reluctant to give me their current letters which limited my questions. The AI produced a comprehensive set of admittedly mainly generic questions, but there were a couple which I added to my list.
Overall, I don’t consider myself remotely expert at deploying AI in daily tasks and the boost to my productivity has been marginal rather than revolutionary. The AI tools and my ability to deploy them are getting to the point where I shall look to refine repetitive processes. The nature of my work is that these tend to be limited to admin, for example collecting and collating the data to send to the accountant for VAT returns. This is the type of application that I intend to work on.
Next week, I’ll show you where AI has really moved the needle for me: in larger projects that would’ve taken days (or been impossible) without it. I go through 3 specific projects and explain how AI made them happen and what I learned in the process.
And that podcast with Gary Channon is really worth an hour of your time.
You are reading the free version of the newsletter, we would love you to upgrade, but otherwise, please consider supporting our advertisers, this week coincidentally being AI related. This is an advert, not an endorsement.
Become an AI Generalist
Still don’t use AI to automate your work? You’re way behind in the AI race. But worry not:
Join the World’s First 16-Hour LIVE AI Upskilling Sprint for professionals, founders, consultants & business owners like you. Register Now (Only 500 free seats)
Date: Saturday and Sunday, 10 AM - 7 PM.
Rated 4.9/10 by global learners – this will truly make you an AI Generalist that can build, solve & work on anything with AI.
In just 16 hours & 5 sessions, you will:
✅ Learn the basics of LLMs and how they work.
✅ Master prompt engineering for precise AI outputs.
✅ Build custom GPT bots and AI agents that save you 20+ hours weekly.
✅ Create high-quality images and videos for content, marketing, and branding.
✅ Automate tasks and turn your AI skills into a profitable career or business.
All by global experts from companies like Amazon, Microsoft, SamurAI and more. And it’s ALL. FOR. FREE. 🤯 🚀
$5100+ worth of AI tools across 3 days — Day 1: 3000+ Prompt Bible, Day 2: Roadmap to make $10K/month with AI, additional bonus: Your Personal AI Toolkit Builder.