By 2028, AI will become a leading source of financial advice for nearly 78% of retail investors.
When I first read this Deloitte data, my first instinct was to file it with the other confident predictions that never quite arrive on schedule.
But then I kept reading. Human advisors drop from 35% to 31%. Financial websites currently used by 28% of investors fall to just 9%.
It wasn’t the headline that got me. It was the granularity. Someone had done the math on which advisors disappear first, which channels go obsolete, and by when.
That report was still sitting with me when I walked into a conversation with Pratik Bagaria — one of our fund managers at Dezerv and realised he’d been chasing the same question from a different angle entirely.
His starting point wasn’t a Deloitte report. It was a viral video of Will Smith eating spaghetti. AI-generated, laughably bad, noodles flying sideways, jaw moving at impossible angles. He laughed and moved on.
Eight months later, he watched another AI video. Same kind of task. He couldn’t tell.
That gap, from comically bad to indistinguishable, in under a year, planted a question he couldn’t shake: if AI moved that fast in creative work, what was happening in analytical work? Specifically, in the domain he spent every day in: portfolio management.
So on August 4th, 2025, Pratik ran an experiment. Five AI models — Grok, ChatGPT, Gemini, Perplexity, and Claude and one brief: build a 10-stock portfolio to beat the Nifty 500. Six months in, we have results. And they’re more interesting than either of us expected.
In this edition:
- The experiment: What we asked 5 AI models to do and why
- Month 6 results: Which AI is winning, and is it luck or skill?
- The real story: Each model’s personality revealed in its stock picks
- The prompt problem: Why changing one word breaks everything
- Institutional reality: What AMCs are actually doing with AI today
- The PM as pilot: What AI can and cannot do
- Will AI Replace Wealth Managers? The 80/20 Rule
- The 2026 AI arbitrage
Let’s get into it.
The experiment: Five models, one brief
The prompt was deliberately clean. No hints, no sector tilts, no views on macro. Just a simple mandate:
“Create a portfolio of stocks with the following guidelines.
Guidelines:
- Investment Objective is to outperform Nifty 500 TRI.
- Investment Horizon is 12 Months.
- Invest in exactly 10 stocks, the stocks must be listed on NSE or BSE.
- Assume that I will not undertake any rebalancing of this portfolio.
Provide me the name of the stocks and their respective weights. Also, give me a 200 word rationale for your portfolio”
The Results: Who’s winning?
Halfway through the 12-month experiment, Gemini and Perplexity are outperforming the Nifty 500 TRI, Grok and ChatGPT are underperforming, and Claude sits essentially at market performance.
Here’s the assessment:
But what emerged wasn’t just five stock lists, it was five different personalities making five different bets. Even though these models are trained on overlapping internet data, the portfolios felt built by entirely different fund managers. That’s the result of thousands of design decisions baked into each model that most users never see.
Five Portfolios, Five Personalities
To understand the model’s rationale behind picking certain stocks, let’s analyse whether there is a correlation between a model’s general personality and its investment style.
And here’s where this experiment gets genuinely fascinating: the stock picks were direct expressions of each model’s engineering philosophy.
- Perplexity: Perplexity’s architecture is built around Retrieval-Augmented Generation, it is built for recency. It pulls live data before answering, and the portfolio reflects exactly that — Suzlon’s fresh order book, Solar Industries riding current government incentives, Polycab and Cummins tied to active spending cycles. This is a portfolio of what the market was talking about right now, not what has always worked. It is reactive by design which is either its edge or its risk.
- Gemini: It thinks in frameworks. Every pick, ICICI Bank, Reliance, L&T, Bajaj Finance, is an unambiguous market leader. No mid-caps, no small-cap bets, equal weights across the board. It’s not making strong relative calls; it’s saying these are quality companies, own them. Apollo Hospitals and Siemens sneak in as systematic, theme-based thinking, long-cycle structural trends backed by data. Safe, institutional, and very deliberate.
- Claude: It hedges, acknowledges uncertainty, cross-references rather than asserts. It’s the only model that cited an external source — Motilal Oswal’s top picks, to validate its selections. Rather than backing its own conviction fully, it borrowed credibility from a known authority. The portfolio follows suit: Sun Pharma and Tata Consumer for stability, Bharti Airtel as a compounder rather than a growth bet, and Macrotech at just 7%, included, but barely, as if it wasn’t fully sold on the idea.
- ChatGPT: ChatGPT is designed to be broadly helpful and palatable to the widest possible audience — the answer a smart generalist would give. Thorough, defensible, not particularly edgy. The portfolio is the most balanced of the five: it covers banking, pharma, infrastructure, consumer, and even a small-cap optionality bet. No strong sectoral view — just broad, defensible coverage. Each stock comes with its own specific narrative rather than one overarching thesis.
- Grok bets big. It is built with an explicit anti-establishment personality, designed to challenge consensus, push boundaries, and take risks where other models hedge. It’s the only model that puts Adani Enterprises as its top holding — a stock most advisors handle carefully. It also loaded up on Suzlon and Jio Financial, both high-conviction, high-beta bets. Zero defensive allocation. The portfolio reads like someone who spotted upside others were too timid to take.
What’s striking is that no model broke character, Grok was bold, ChatGPT balanced, Gemini systematic, Perplexity current, Claude hedged. These models weren’t doing independent financial analysis; they were applying their conversational personality to a financial problem.
The two outperformers, Gemini and Perplexity, won by avoiding narrative-driven picks and anchoring to structural data instead — models that chased stories ended up buying at the top.
All of this becomes more consequential as these systems move closer to real financial workflows. Recent developments show that shift clearly. Perplexity Finance now connects directly to FactSet and SEC filings to generate fast, cited summaries, useful for independent advisors needing rapid intelligence at low cost.
Claude, meanwhile, has introduced plugins for financial analysis, equity research, and wealth management. Through its Model Context Protocol, it connects to internal data platforms like Snowflake, bridging proprietary firm data with live market feeds.
But the infrastructure upgrade doesn’t close the governance gap. Firms still need audit trails, compliance filters, and usage frameworks. And hallucination risk doesn’t disappear with specialisation, it simply becomes more polished, and therefore harder to detect.
Character vs. Consistency: The AI Performance Gap
Before we get carried away by the six-month numbers, let me ask the uncomfortable question: are the results good because the models are robust or is this just a coin flip that landed well?
My honest answer: the data is inconclusive, and likely leaning toward mixed. Two models outperforming, two underperforming, that’s probably simple variance. Over a six-month window in a specific market regime, any diversified portfolio has roughly a coin-flip chance of beating the index.
To prove this isn’t a one-off, the models would need a consistent Sharpe Ratio that holds across different market cycles. A bull run favours almost everyone. The real test is what happens when sentiment turns, liquidity dries up, and volatility spikes. We haven’t seen that test yet.
These models are not deterministic, ask the same question tomorrow and you won’t get the same portfolio. There’s no accountability loop, no conviction maintained across time. If the model recommended Reliance at 10% last August and it’s down 15%, it might take a completely different position today without ever reconciling the contradiction.
For now that makes them stochastic brainstormers, not disciplined Portfolio Managers.
The PM as Pilot: A precise map of what AI can and cannot do
By 2026, the Pilot vs. Autopilot dynamic has become standard across every high-stakes profession. AI is the world’s best radiologist, yet it cannot replace a doctor’s empathy when delivering a difficult diagnosis. Algorithms can audit 10,000 documents in seconds, but cannot argue a case where human intuition wins the day.
The modern Portfolio Manager has made the same shift: from calculator to pilot. What defines their value today is everything the autopilot can’t do.
What AI can do:
- Extract Signal from Alternative Data at Scale. Satellite imagery of shipping ports, real-time credit card flows, social media sentiment, AI turns this noise into actionable leading indicators continuously, not just quarterly.
- Contextual Document Auditing. AI flags if Risk Factor wording subtly changed year-on-year, if a footnote contradicts the CEO’s letter, or if management’s vocabulary on earnings calls has shifted evasively over five years.
- Multi-Dimensional Screening. Beyond static filters, AI uses cluster analysis to screen 10,000+ stocks by latent factors, supply chain exposure, patent trajectories, geopolitical risk correlation, in seconds.
- Real-Time Risk Monitoring. AI detects hidden correlations that only emerge under stress and runs instant what-if simulations. The answer comes in seconds, not days.
- Surgical Execution. AI minimises slippage, automates rebalancing, and can already predict roughly 71% of active fund trading decisions based on historical patterns.
What AI cannot do:
- The Field Visit. AI cannot fly to a factory and notice demoralised workers or poorly maintained machinery. It cannot read a CFO’s body language in person. This primary data doesn’t exist in any database.
- Form Relationships. The best insights come from trusted conversations, not public data. No AI can take a retired industry veteran out for coffee and extract the off-the-record insight that builds a truly differentiated thesis.
- Second-Order Thinking. AI is excellent at first-order logic. It fails at the recursive logic of markets: if everyone believes this insight, how will the market overreact, and how do I profit from their collective mistake? That remains uniquely human. The same gap appears in macro shifts. AI records a trade restriction as a revenue variable. A human portfolio manager sees the chain reaction — nationalist politics, regulatory retaliation, supply chain rewiring, capital reallocation over a decade. AI sees the data point. The human sees what the data point sets in motion.
- Strategic Courage. When a portfolio is down 20%, AI cannot have conviction, only an optimisation function. The leap of faith required to back a visionary against the data is not programmable. It is earned.
- Accountability Gap: When an AI recommends a stock that drops 40%, nobody is responsible. No fines, no licence at risk, no fiduciary breach. A human portfolio manager carries real weight, they answer to investors, revisit their thesis, and have reputation and capital on the line. An AI has no skin in the game; it just generates a new recommendation the next day.
Will AI Replace Wealth Managers? The 80/20 Rule
I think about wealth management as 20% math and 80% life design. AI is genuinely superior at math — portfolio construction, rebalancing, tax-loss harvesting — and most of that will be automated within five years. But the 80% is a different problem entirely.
Understanding why a client’s relationship with money was shaped by watching their father lose everything, helping someone hold equity through a 35% drawdown without selling, navigating the emotions of a business exit, none of that is a data problem. It’s a human one.
For retail investors, AI is a powerful research tool that democratises analysis once reserved for institutional desks. But only if the investor has enough financial literacy to know what to do with the output.
The Risk Nobody Is Talking About
What happens when millions of people ask the same AI for a good stock at the same time? If every retail investor queries the same model and gets the same recommendations, the resulting demand creates the very price appreciation that validates the pick — which then gets surfaced to the next wave of users. The AI-driven reflexive loop could be faster and more concentrated than anything we’ve seen.
The second risk is more insidious: institutional bias encoded in the model. If a firm trains its AI on data tilted toward its own products, the model will “neutrally” recommend those products. The bias is invisible, which makes it more dangerous, not less.
The 2026 Model: Three tiers of wealth management
When everyone has the same tool, the differentiator is no longer access to data — it’s the quality of the question you bring to it. The investor who can challenge the AI’s thesis rather than accept it, who brings domain expertise to the interface, retains a meaningful edge. The AI arbitrage in 2026 is not about who has access. It’s about who uses it with the most precise intent and who retains the judgment to disagree with it when it’s wrong.
That edge will look different across different investor profiles. Here’s how I see the landscape settling.
For most readers of this newsletter — CXOs, founders, senior professionals managing meaningful wealth — you sit in the hybrid tier. That’s precisely where the human-AI balance matters most. At Dezerv, we’ve always believed that the best investment process combines the rigour of data with the wisdom of judgment. AI gives us sharper tools. It doesn’t change what we’re building toward: wealth that compounds quietly, through cycles, over decades.
Disclaimer – Investment in the securities market is subject to market risks, read all the related documents carefully before investing. The information provided herein is intended solely for educational purposes. The AI model portfolios discussed are part of an internal experimental study and do not constitute an investment product, investment approach, recommendation, offer, solicitation or advisory service. The stocks, allocations and performance shown are for illustration only and should not be construed as investment advice or a recommendation to buy, sell or hold any security. Readers are advised to consult with their financial advisor before making investment decisions based on the information provided herein.
In the preparation of this document, Dezerv has used information developed in-house and publicly available information believed to be reliable. The information is not a complete disclosure of every material fact and terms and conditions. While reasonable care has been made to present reliable data in this article, Dezerv does not guarantee the accuracy or completeness of the data. The information / data herein alone is not sufficient and shouldn’t be used for the development or implementation of an investment strategy.
Dezerv, along with its directors, employees, or partners or any of its affiliates, shall not be held liable for any loss, damage, or liability arising from the use of this document. Additionally, all trademarks, logos, and brand names mentioned are the property of their respective owners and are used for identification purposes only. The use of these names, trademarks, and logos does not imply endorsement or recommendation.