An exchange-traded fund that hands US large-cap stock selection to an artificial-intelligence model has been trading since February, and Kevin Max said on air that it has significantly outpaced its benchmarks.
The case against handing a portfolio to a network of AI agents usually comes from people who do not build them. Irina Bevza runs quantitative strategies for a living and uses AI every day to write code, and she still says the machine is least reliable in the conditions that decide a client's year.
"We're still not fully comfortable to outsource investment decision making to a machine."
Bevza wrote the CFA Institute post this conversation is built on, reading the research on multi-agent systems and asking what it does to a portfolio. She is head of quantitative solutions at Fineco Asset Management in Dublin and a research fellow at Trinity Business School.
I listened to the full interview so you can skip it. 32 minutes of audio, 16 minutes of reading.
Here are the 14 takeaways that matter.
👤 Guest: Irina Bevza, head of quantitative solutions at Fineco Asset Management in Dublin, a board member of CFA Society Ireland and a research fellow at Trinity Business School, who wrote the CFA Institute post on the self-driving portfolio
🎙️ Host: Kevin Max, editor of the Enterprising Investor blog at CFA Institute
📰 Published: 31 August 2026 on the Enterprising Investor feed (CFA Institute)
🟢 Spotify | 🟣 Apple Podcasts | ⏱️ 32 min | ✅ Time saved: 16 min
Key Takeaways
A self-driving portfolio is quantitative investing with an orchestration layer on top, not a new idea What is new is the software connecting research, risk and construction, work people used to do by hand
The blocker is not the models, it is the state of the data
AI is faster than a portfolio manager, not better than one yet
What a machine cannot supply is ownership of the decision Nobody has settled who answers to the client when an agent loses the money
One agent misreading a headline can move the whole portfolio the wrong way A human chain is slow enough that the follow-up story arrives and corrects it
Agents built from similar prompts agree with each other and call it consensus
The self-improving meta agent is weakest in the markets that matter most A system trained on calm markets has no time to relearn once volatility arrives
A large language model's audit trail is longer than the decision it explains
Juniors will take ownership earlier because AI has taken the work they used to learn on
The head of quantitative solutions would buy the AI fund and still not put a pension in it A long-short strategy is a hard sell to retail investors while the index keeps rising
Trust needs governance, auditability and a name attached to the loss, and none of the three is settled
The human portfolio manager is still there in 10 years, with a different skill set
1. Old quant idea, new layer
Max opened by defining the thing: a network of AI agents that runs continuously, generates capital market assumptions, builds portfolios, challenges its own conclusions, monitors outcomes and rewrites itself when it gets things wrong. He then put the architecture to Bevza as a software version of an entire investment firm — researchers, portfolio constructors, risk managers, critics, and its own investment committee.
Bevza said the concept is not new, and traced it to systematic investing. "So the idea there behind it is just you encode investment decisions into rules and models."
The new part is the layer above the models, not the models. "However, AI adds an additional layer of orchestration and I think it's great and interesting." Connecting analysis, research and risk management to each other used to be human work, she said, and that is the piece AI now contributes to
She said she does not see it as a completely new concept, because the underlying idea has been around for a while
2. Data prep is the blocker
Asked what the architecture says about where AI is headed, Bevza stopped the question and reframed it around what a firm would need before any of it works.
The gap at an average investment firm is data, not intelligence. A huge amount of work on data governance, data cleaning and data preparation still has to be done before AI can function efficiently, she said, and the investment industry is not yet fully capable of it
What AI already does well is speed. Where a firm has it enabled, she said, the processing is very fast
Her conclusion is a scope limit rather than a rejection. For now, she said, AI belongs on localized tasks rather than on automating an investment firm end to end
3. Faster, not better, yet
Max pointed out that the retail version already exists — he said he can go online and buy into a self-driving portfolio today for $30 a share — and asked why, given the drawbacks, anyone would not do this.
Bevza would not accept the premise that the machine is already better. "So I cannot say that AI is doing a lot of processes better than humans yet."
Her distinction is between speed and quality. It does things faster and can be rational, she said, but in many cases the technology is still enabling and collaborating with the human rather than replacing the human
She said the big question is what happens to portfolio management overall, and asked for a detour before answering it, saying she needed to think about how to answer it
4. What's missing is ownership
Max put the governance question directly: committees meet once a month or once a quarter, markets move daily if not continuously, and he asked whether what is called good governance is sometimes just institutional latency.
Bevza agreed that committees are slow, and said so from the inside. "And working in those committees and being present of those committees it is frustrating."
The delay is in the preparation, which is the part AI can take. What holds a committee up, she said, is several stakeholders preparing analysis and deliberating before they reach a conclusion. "Where AI could be really efficient is to bring all this information to the committee faster so committee can work more efficiently."
She said a system of agents can be built to function like a committee, communicating and discussing without stopping
The thing the agent committee does not have is ownership. "We're still not fully comfortable to outsource investment decision making to a machine." Her test is what a person does with their own money: even someone who buys the retail product, she said, would most likely still put a pension in a traditional fund
Regulation is the other half of the answer. From the point of view of investor protection, she said, we are not there yet and will not be for a while
Her forecast is a faster version of the same structure: with proper data infrastructure the time to market shortens, roles evolve, and people work with the system rather than being replaced by it
5. Rivals will speed up too
Max played devil's advocate. If he were KKR raising money from pension funds, he said, he would go to the ones that can decide on a rolling basis rather than the ones that schedule three months out and then take another month — market forces doing what good governance has not.
Bevza's answer was that the slow institutions do not stay slow. She said the three-month scheduling goes away, "so those people will become faster and you will be debating between instant decisions that you probably don't need and just the faster decisions"
Max conceded the point, and the exchange ended with both of them having played devil's advocate at each other
6. Compared to a perfect human
Max listed the human failure modes an AI is supposed to fix — groupthink, confirmation bias, career risk, recency bias, personal politics — and asked whether AI is being held to a standard of reliability that human portfolio managers cannot match themselves.
Bevza said the comparison itself is the bias. People measure AI against a perfect human, she said: the best investor in the world, summoned in a single line of prompt
Human biases are at least mapped. There is an extensive body of research on them, she said, and current products have been adjusted to them
The machine's biases are not. "AI will also have its own biases." She put the sources at prompt engineering and at the model being overly positive, which is the failure she flagged in her paper
The system she wants is complementary, not substitutional. AI offsets the human biases — she said taking emotion out of the question has been the role of quantitative investing all along — and the human watches for the biases coming out of the model The target keeps moving: a new large language model arrives every couple of weeks with different fine-tuning, and new skills develop around it
She does not expect the problem to be solved. "So we will never eliminate the biases"
7. One bad read cascades
The first AI failure mode from her paper is the cascade. Bevza built it out of a news story.
The mechanism starts with one agent reading a headline wrong. A tariff story misinterpreted by the news agent turns something positive into something negative
The error then travels to the agents that trade. It reaches the portfolio manager agent or the risk agent, she said, and they take the opposite side of the portfolio — while a follow-up story would have cleared the original up. "So that's where you have a cascade problem." A single data point arriving downstream can end in a material loss on a position that was misunderstood, and she said the scale of it can run from small to the whole system
The human version of the same chain is protected by its own slowness. Because people take longer over decisions, she said, there is time for the news to come back, be interpreted again by news agencies, and be corrected by more senior managers on the way
8. Similar prompts, one opinion
Bevza said the second failure mode is a human one before it is a machine one. "False consensus is related to have a lot of similarity and I think it's actually very human as well."
Identical settings produce agreement rather than analysis. If the agents are built from very similar models and prompts, she said, you have a system that agrees with itself all the time and produces no diversification of opinion
She said one of the papers she referenced found the same effect in the models themselves, with some large language models too agreeable in their settings, so that when one talks to another they reinforce each other's biases Her comparison for it was people being too polite
Max's fix was a prompt. "Maybe these models should all start with the question, should I short sell this? And if not, why?" Bevza said it was a good start
9. Meta agents fail in a crisis
Max asked about what he called the ghost in the machine: the meta agent that reviews the system's mistakes and rewrites its instructions.
Bevza said her blog is based on a paper about the self-driving portfolio, and that the meta agent comes from there. She called the design elegant — one agent overseeing the others and adjusting their prompts on the evidence of what went wrong
The objection is that markets are not a precise science. They are a social science, she said, full of unpredictable events, and in a period of market stress there is not enough information or data to draw on
A system tuned on quiet markets has the wrong training set for the moment it is needed. "So if you have a system that has been trained a lot on their calm markets, it might have very difficult time adjusting it."
The self-correction loop runs out of time exactly when it matters. "It would try to reorganize the agent system but wouldn't really know how, but it would not even have time to relearn because the volatility is so high."
She said the same pressure applies to systematic strategies generally, and to people, who face the same push to cut a position under stress
10. Auditing a debate is hard
Max raised the problem he says keeps arriving on his desk: whether the black box can be audited, and who is accountable to clients and to regulators — noting that governance in Europe and the US are not the same thing.
Bevza said the accountability question is open and will be answered slowly, and expects a lot of the direction to come from regulation
She said the black box label is inherited from the machine-learning wave, when a neural network made decisions nobody could account for
A large language model is a step forward on that, because it writes its reasoning down. She named ChatGPT and Claude as the everyday example: the model sets out its thinking for the user to read
The improvement creates the audit problem it solves. A quant, she said, usually operates in one sentence, while "you work with like LLM who likes to produce a Bible for decision making" — and a multi-agent system debating a decision produces a transcript of the argument. "So imagine going and auditing it after it as a human."
Max's two analogies for the same discomfort were his own. "In some ways I feel like on the personal level I'm working with a crazy person." He described asking a model about cars, being told they are all great, and then drilling into one and being told he should never buy it The second was a broken family heirloom, four children looking at each other and nobody accountable — where the heirloom, he said, is the portfolio
11. Ownership comes sooner
Bevza's answer to the accountability problem is a change in the shape of a career.
The work that trains a junior is the work AI is taking. She listed slide-making, quick analysis, consolidating databases and creating charts as what she did in her first years, and as the jobs nobody wants that turn into a learning curve. "So that's what AI will take away."
Her conclusion is that responsibility arrives earlier rather than later. "But the positive thing in it is that junior person will be more trained to take over the ownership much faster" Learning to audit and validate a decision is what she thinks that produces, and she called it good for critical thinking
She said she is very positive about the developments
12. Playing money, not pension
Max set a study against a product. He said he had recently read a piece by a group of PhDs and CFAs from UC Berkeley, Stanford and IBM Research, which reported that the first large-scale study of AI agents in production found successful deployments to be simple, tightly constrained and continuously supervised, and that agents today are not autonomous. That, he said, was a year ago.
Against it he set a live fund. An exchange-traded fund with the ticker AINT, an all-AI-driven, dollar-neutral portfolio of S&P large caps, has been trading since February and has significantly outpaced its benchmarks, he said, although he added that it is hard to say what the benchmark for it should be
Max's framing of the gap between the research and the product was that the technology is past the point of debate. "The genie is already granting wishes to investors."
Asked whether she would put her own money and her pension into it, Bevza drew the line at the amount. "I'm thinking about actually playing money right now until some of these issues have been resolved we're talking about." What happens after that, she said, is anyone's guess
The reason given in the exchange was track record, not technology. The fund has traded only since February, and the point returned to earlier in the conversation: a strategy has to be watched through several stress periods before its performance means anything
The harder problem for the product is the strategy, not the AI. A long-short portfolio is confusing to explain to retail investors, she said, who are more comfortable with long-only exposure to the S&P — and it is a difficult thing to sell in a market where the index keeps rising
She would buy it herself. "I saw the performance, it's fantastic." She said she would buy it partly to support innovation in the space, and that the problem is that she would not trust a pension to it
Max's summary of where that leaves the fund was that it is lunch money for now
13. What trust would require
Bevza's conditions are institutional rather than technical. She said the technology is moving fast and knows more than she does in places, and that what has to catch up is governance, autonomy and auditability — the pillars that would let a system act independently and be trusted
The first missing piece is a name attached to the loss. "I don't know who's going to take responsibility if something's happening to my investment."
The second is the operating envelope. She said she does not know what parameters such a system would act under, or how it would be controlled
Her evidence is her own use of the technology. She uses AI extensively for coding and is working out how to build agent systems, and said it is already trustworthy in the narrow sense that it checks in with the user and stays out of folders it should not open The same tools, she said, still hallucinate: "but it still hallucinates a lot"
14. Still there in 10 yrs
Max closed with a sentence for her to complete: ten years from now, maybe five, the human portfolio manager will most likely be what.
Her answer was two words. "still there."
The job survives, the skill set does not. The skills will not be the current portfolio manager's, she said, and the way the role acts will be different, but the responsibility stays with a person. "So what I want to say is we're not going to be replaced, right? It's gonna be just different and evolved."
She made it conditional on demand rather than on technology. What happens to the portfolio manager depends on what happens to the investment industry, she said, and she hopes there is no large change in demand
The rest of her case for the industry has nothing to do with AI. She pointed at ETFs, faster trading and faster settlement as what is making the industry more interesting and diversified
Her closing hope was that delegating the boring tasks lets creativity emerge and moves the industry somewhere more interesting
Bonus Insights
Max's worry is about succession, not about the machine. "I worry that five or ten years down the road that we won't have those senior people who still understand and could disintermediate and step in and say, Huh, that doesn't seem right." His reason was that the people coming up will have used AI for everything Bevza said he was underestimating the commitment of the social structures already in place. "I don't think we will see the seniority moving away" Her counter was that AI adds creativity and the ability to run different scenarios, and could give a learning experience to people who would not otherwise get one
Max framed her position throughout as blue-sky optimism, and said at one point that he hoped the sky really is that blue. Bevza answered with the version of his scenario where it works. "Maybe in your scenario all people would have invested the entire pension in the successful robot driven portfolio and retired somewhere on a nice island and enjoying the martinis in the afternoon."
Bevza pointed listeners to her post through the CFA Institute site rather than a link, and Max spelled her surname on air for anyone searching for it
Bevza's bottom line is that the agents can already do the work that surrounds an investment decision, and that until someone can be held responsible for what they decide, the decision itself stays with a person.
Products, Companies & Tools Mentioned
FINQ Dollar Neutral U.S. Large Cap AI-Managed Equity ETF (The ticker AINT, which Max described as an all-AI-driven, dollar-neutral portfolio of S&P large caps that has traded since February and significantly outpaced its benchmarks)
ChatGPT and Claude (Bevza's example of a model that writes its reasoning out for the user, which she said is what makes a large language model more auditable than a neural network — and Max's example of a model that calls every car great until you ask about one)
KKR (Max's example of an allocator that would raise money from pension funds able to decide on a rolling basis rather than ones that schedule three months out)
The S&P 500 (The long-only exposure she said retail investors are more comfortable with, and the universe the AI-run fund trades)
Books & Resources Mentioned
The Self-Driving Portfolio: Promise, Pitfalls, and the Practitioner Gap – Irina Bevza (Her CFA Institute post, which Max said this conversation is built on and told listeners to search for by her surname)
The Self-Driving Portfolio: Agentic Architecture for Institutional Asset Management (The paper she said her blog is based on, and the source of the meta agent that rewrites the other agents' instructions)
A study of AI agents in production (Read out by Max from a piece he attributed to PhDs and CFAs at UC Berkeley, Stanford and IBM Research, reporting that successful deployments are simple, tightly constrained and continuously supervised)
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: