When the Florentines broke ground in 1296 for their now world-famous Cathedral (formally the Cathedral of Saint Mary of the Flower), they had no idea how to finish it. That is, they knew the dome of the cathedral was going to be so big that nobody would have any idea how to build it. The plan was that because the construction was going to take decades, along the way, the posterity would smarten up and sort it out. About 100 years after the work began, in 1380, when everything except the dome had been built, the construction work had to come to a stop, because, well, the posterities had not sorted it out. The original design seemed just too ambitious. Not only the dome was supposed to be so big (the largest one ever attempted, bigger than the biggest dome from Ancient Rome. Even today it is still the largest dome built with brick and mortar), it also would have to be an eight-sided one, instead of a semicircle one, because, to make the cathedral truly unique, it had been designed as an irregular octagon with no true center, which means the dome also has to be built on an irregular octagon drum. And because there were no flying buttresses like the gothic one Florence’s rival city, Milan was building, Florentian architects had no idea how the dome would support itself even if they managed to build it. Finally, on top of all these, because the dome was so high, it was not possible to build a scaffolding, as no material strong enough was available then.

After 1380, the half-finished cathedral was just sitting there, in a silent rage, in the center of Florence, without its dome. Whenever it rains, water would pour into it. Every day, people passed by it. Every Sunday, people would come to the site of the halted construction, where the old cathedral, Cathedral of Saint Reparata, was, for the mass. The new cathedral was supposed to replace it but never did, and therefore people had to continue worshiping in the place of the old one.

Must have been awkward. And yet the city had to wait in this awkwardness for decades, alternately being shamed and inspired. Finally, a young goldsmith-turned architect, Filippo Brunelleschi, started to resume the work. By using several ingenious designs, including a double-shell structure, horizontal compression rings that hooped the dome together, and a unique herringbone brick pattern, Brunelleschi managed to deliver the impossible. Nowadays, historians ascribe the beginning of the Renaissance to 1420, the year when Brunelleschi started to build the dome for the Florence Cathedral.

There are other examples of a lofty ambition having to wait for the means to catch up. The first proposal of aeroplanes came in 1810. Pharaoh Senausert III proposed the idea of the Suez Canal in 1800BC. Sometimes, the more ambitious an idea, the longer it has to wait. It would seem that the story of robo-advisors is one like this as well.

A promise in default

Robo-advisors rose from the aftermath of the 2008 financial crisis, as many more people started to seek professional financial advice in the wake of the worst wealth destruction in generations. Economics and regulations however determined that it was not possible to increase the supply of financial advice very quickly. Not by traditional means anyway. Robo-advisors stepped in with a promise to make financial advice affordable and accessible to everybody by using new technologies. The idea was to automate the complete suite of personal finance services: diagnosis, asset allocation, fund picking, investment execution, portfolio monitoring, risk management, and all the necessary reporting and analytics so much that, with no minimum account balance required, any one will be able to afford all these services for 1-2% a year. And the creme de la creme is that all this can be done without compromising the independence and objectivity of the advice, because the new technologies will make it so cheap that a retrocession from the product suppliers will not be necessary.

Except, just as the Florence Cathedral in the 14th century, the critical technologies were not there yet. Like many other ventures from the post-2008 fintech heyday, the grand facade of robo-advisors was backed by a philosophy rather than any particular technology. Specifically, the philosophy that less is more.

Instead of automating the onerous process, robo-advisors just simplified it. The discovery journey to find out the detailed financial situation of the clients was replaced with a questionnaire. Customization was limited to several target portfolios, usually composed of passive funds. Regular in-person meetings were replaced by auto rebalancing. Ad hoc advice was substituted by newsletters. More sophisticated services, such as tax optimization and insurance selection were discarded altogether.

While this did bring down the cost, it also reduced the customer experience to a single dimension. The information available from a robo-advisor and the interaction with it never developed the nuances necessary to build an intellectually meaningful, and subsequently, commercially relevant relationship. In short, robo-advisors failed to bring the needed advice to the customers. Instead, they looked more like a discount fund broker with some fortified visualisation tools.

Defaulting on its promises, robo-advisors gradually digressed and never managed to capture the market share assigned to them. While the initial estimates were for robo-advisors to claim USD1-2 trillion of the wealth management market in five to ten years, almost 20 years later, this target still remains elusive now. Not only robo-advisors never became the sweeping force portended, but as time goes by, their prospects look increasingly bleak too. A robo-advisor charging 25bps with no minimum balance requirement needs a minimum AUM of USDS10billion to break even. In reality, only a handful of robo advisors managed to even get close to that number.

When it became clear that the crucial technology could be years away, some robo-advisors decided to go bionic, as human touch seems to be the only way to deliver the desired client experience. All but one of the major robo-advisors now offer some sort of human engagement. But adding human intervention also imposes new costs on the already wobbly financials of the robo-advisors, not to mention a whole suite of other managerial challenges, from human resources to compliance. Some decided to partner up with or seek to be acquired by a bank, or a brokerage firm, in the hope that their existing assets, including client trust and human team can be leveraged. In both cases, the robo-advisors had effectively ceased to be robotic in nature. The last-ditch effort unsurprisingly did little to save the robo-advisory model. Starting in 2015, more and more were forced to exit the market altogether. At this point in time, the robo-advisor idea looked tragically similar to the Florence Cathedral in the year of 1380. It seemed that all that could be done had been done, and what remained to be done, the most critical part, depended on means that were beyond reach. And just like the Cathedral, the hopes in robo-advisors were at last saved by something fortuitous, almost provident: the breakthrough of the large language models (LLM). Almost like a deus ex machina, LLMs overnight enabled chatbots to mimic human conversations, and gave roboadvisors the long-needed human-like interface that is supposed to allow personalisation and trust. Seems that good things had finally come to those who wait, and the novel idea of robo-advisors, despite the initial struggles, had found its Brunelleschi and in a wonderfully symbolic way, can now look forward to its renaissance.

What Large Language Models can do

The LLM and its manifestation of generative artificial intelligence (GAI) has very wide application but is most helpful to robo-advisors in three aspects.

The first, and on the exterior, is to produce naturalistic language, and in doing so, mimic human interaction. The biggest impediment to fully explore what robo-advisors have to offer is its apathy. With LLM, the interaction between the client and the roboadvisor is no longer limited to multiple choices and binary questions. A conversation conducted in naturalistic language, though artificial, allows the human user to be more articulative and inquisitive at the same time. It also allows the robo-advisor to collect and deliver information more effectively, offer more sophisticated advice, and provide better personalized advice. In short, make the correspondence more engaging. And with the rapid development of the LLM and GAI, this can soon develop into a conversation with a virtual interlocutor in human image.

The second, and on the interior side, is its capability to perform qualitative assessment. Since the beginning, robo-advisors have been limited to ETFs as means to construct portfolios for their clients. Actively managed funds, stocks and bonds, and alternative assets are all excluded because the computers are considered not smart enough to fully appreciate the granularity of these investments, probably because of the texture of the analysis required. With the breakthrough of the LLM, not only robo-advisors can analyse, and subsequently provide advice about these assets, it can analyse them more thoroughly and more quickly than any human can. This expansion allows robo-advisors to offer a holistic solution to individual user’s portfolio construction, another feature crucial to building a coherent business model.

While the first generation of robo-advisors relied on the algorithms to generate their investment advice, the LLMs allowed them to complement it with qualitative analysis. Combining the two is like giving a maths geek a liberal arts education, and results in a comprehensive upgrade. When these new capabilities are fully assimilated, a digital-native robo-advisor will have significant advantages vis-à-vis its competitors with a more traditional setup and perhaps finally can deliver the once defaulted promise: make personal financial advice cheaper, quicker and better.

A third, less fundamental enablement is that the GAI based on LLM can produce a virtual human image in real time. Technically, GAI can already produce video conference calls with the users of robo-advisors and make them feel like they are talking to another human, or the image of a human at least. The additional layer of the human likeness helps make the experience more wholesome, though in reality, cost and ethics concerns means that this is not a priority area of GAI application.

LLM as a lifeline, and not a redemption

At this point it is worth caution that while LLM has thrown the robo-advisors a lifeline, the technology itself is not a redemption. It needs to go through specific adaptation to be applied in the financial advisory business. This adaptation will need to account for the rapid development of the technology itself and be subject to regulatory scrutiny, which is also evolving quickly. After all, if advice is a dangerous gift, even from the wise to the wise, as J.R.R. Tolkien noted, wouldn’t it be a peril when mass-marketed as a commodity, by a computer to a human?

The most immediate challenge when applying LLM in financial advice is the problem of hallucination. And in the world of managing people’s financial security, the margin for error is extremely limited. At this point, LLM is still undergoing rapid development, and it seems reasonable to expect most hallucinations to be dealt with in time. But the foundational design of the LLM means that hallucination will be difficult to eradicate. Which means robo-advisory service providers will have to continue investing in a combination of remedies, from strengthened training, to retrieval-augmented generation (RAG), to human oversight to combat it.

Even if hallucination can be neutralised, LLMs have other innate limits. Being a probability-driven model, LLM does not think like a human thinks, and does not construct a sentence as a human does, although it may look infinitely similar. No AI can pass the Turing test yet and despite exciting development, none seems hopeful anytime soon, if ever. A more pessimistic view derived from this is that humans will never trust a machine, that there will always need to be a human to hold the clients’ hands, and that there will always be something, personal finance being one of such things, a human will only talk about with another human, and never a machine.

Among all the trials robo-advisors (and artificial in general) will have to face, this human trust issue is the most existential but also the one with the most coherent solution. While there certainly is a sombre side to the LLM and AI, most of the so-called trust issues are due to the models not being good enough, rather than humans fundamentally incapable of trusting them. What a human user places in a human financial advisor is confidence, not trust, the same as what a human places in her lawyer, accountant, or doctor. What a human user is expected to place in the robo-advisor is trust, in the same way that she trusts her safety belt. Human bonding is not a prerequisite for trust. There are countless examples of a person trusting something completely free of human elements: credit card, ATM, banking app, long call butterfly, credit default swap, central clearing. The list goes on. In fact, trust without intimacy is one of the defining features of modern society and is professed most ruthlessly in the world of technology and in the world of financial services, the junction of which is precisely the theater of robo-advisors. To overcome the trust issue, robo-advisors will only have to create a customer experience distinct from that with a human financial advisor, while working on the foregoing hallucination issues.

Ironically, LLMs themselves also pose a threat to robo-advisors. That is, as the LLMs become better, it can potentially replace robo-advisors altogether. While a singularity moment that presents an omnipotent, omniscient AI may never arrive, an iPhone moment that integrates all specialised models may happen sooner than expected. Already millions of people are asking AI for advice on stock picking or personal finance. For the time being, robo-advisors rely on various regulations to fend off the trespassing of the generic LLMs (the various disclaimers coming with AI responses, self-imposed limits on certain questions, and the licenses needed to charge money for finance-related advice). But unless roboadvisors can assimilate LLM technologies faster than the pace at which LLM upgrades themselves, soon enough the generic catechism will be good enough for most people. The alternatives facing the robo-advisors here are similar to the trust one: it will be strategically foolhardy to try to stop the autophagy of LLM. Instead, building a business model that brings the best out of the new technology is the only possible way out.

Another exogenous uncertainty robo-advisors have to account for are the regulators. Although in the GAI era, regulators serve as part of the railguard for the roboadvisors, they have mostly been a reactionary force since the beginning, by putting various limits on what the robo-advisors can sell. The result is that most sample portfolios found on robo-advisory platforms comprise ETFs only. The heightened market volatility since 2020 reinforced this restrictive attitude. The potential to cause serious legal and regulatory consequences means that regulators’ unabated prudence will likely persist. Robo-advisors need regulators to shield from the inroads made by various LLMs, but more critically they need the nodding from the regulators to flex muscle in the more sophisticated asset classes, an area where their newly acquired capabilities will best manifest and the only way they can build a competitive edge and differentiation in the battle to replace human financial advisors.

It seems that what is ahead of the robo-advisors is a 3D tightrope walk. They need the underlying technologies of LLMs to evolve quickly, but not too quickly. They need the regulators to intervene on their behalf, but not become too involved. They need customers to be open minded, and yet remain able to discriminate.

It took Brunelleschi 16 years to finish the project, even though he had developed all the details of those brilliant designs from the very beginning. The audacity of the work meant that a raft of new tools had to be built, such as new weight lifting machines, to finish the job. The lanterns alone took a decade to be installed. It is perhaps prudent if not reasonable for the LLM-backed robo-advisors to prepare themselves for a similar journey. While the new technologies are invigorating, its raw power has to be harnessed and administered before it can be translated into wide applications. Like every revolution, robo-advisors will have to be both an instant relief and a lasting promise.

For roboadvisors, AI is a lifeline, not a redemption