Recent advancements in artificial intelligence (AI) are beginning to fundamentally reshape the economics of software development. The latest generation of models can now perform a substantial share of technical work at a level comparable to — and in some cases exceeding—that of junior developers, often at a fraction of the cost. While public debate has largely focused on employment implications, the more consequential effects may be structural, emerging within market infrastructure and across the trading ecosystem.
Trading-technology providers and financial institutions—including multi-dealer platforms (MDPs), order management systems (OMS), execution management systems (EMS), and bank e-trading desks—have historically relied on proprietary technology stacks as a primary source of differentiation. As AI reduces both the cost and complexity of building advanced software, however, technology itself is increasingly becoming commoditized. Capabilities that once required large engineering teams, long development cycles, and substantial fixed investment can now be replicated more quickly and at far lower marginal cost. As a result, the long-standing assumption that a differentiated technology stack alone can sustain durable competitive advantage is being challenged. This dynamic may already be reflected in the recent sell-off across technology and SaaS equities market, see Figure 1.
In summary, while many uncertainties remain, one central question stands out: how should fintech and trading firms adapt when technology itself is no longer a sufficient differentiator?
The sections that follow examine recent developments on both the buy side and the sell side, drawing on professional experience and empirical evidence. Through this analysis, I aim to provide a grounded view of how the industry is evolving and, ultimately, to offer a reasoned answer in the concluding section.
Buyside impact
With the emergence of artificial intelligence (AI) and transaction cost analysis (TCA), buy side firms have become markedly more sophisticated in their evaluation of liquidity providers and execution workflow design.
Figure 2 shows that the BestX peer database has accumulated more than USD800 trillion in transactional data. For TCA providers such as BestX, metric definitions are industry-standard and generally taken at face value. The more challenging—and more valuable—task for the buy side is to translate this enormous dataset into a single, actionable decision. To address this need, we developed BestXecutor (Figure3), which allows clients to combine multiple benchmarks — for example, spot spreads, slippage, signaling risk, and hit ratio—to evaluate and select the most suitable algorithm or liquidity provider. This optimization process is central to the life cycle of a trade. As multi-metric workflows gain traction among buy-side users, they are increasingly displacing traditional assessments that relied on a single metric or on relationship-driven decision-making.
A frequently cited concern, raised by several industry practitioners, is the misinterpretation of metrics. Part of this risk has been mitigated by the use of AI: for example, some buy-side traders told the author that they can upload a BestX TCA report into an AI tool and obtain a clear explanation of the results. Overall, the evidence suggests that, with AI support, the buy side is increasingly able to assess algorithmic execution and liquidity-provider performance using data-driven analysis rather than relying primarily on client relationships.
In parallel, rising client sophistication—further enabled by AI—has increased demand for more advanced models and API connectivity. In an earlier article on AI, I briefly discussed the Algo vs RFQ model, which is now live in BestX. The model was initially designed for a small set of clients because it leverages live market data and applies modern statistical and machine-learning techniques to support execution decisions. However, with AI-assisted interpretation and improved usability, adoption has broadened: a wider group of clients can now understand the outputs and incorporate them into their workflows. Additionally, as the cost and complexity of technology stacks continue to decline, demand for API connections has expanded beyond a small “elite” group of technically advanced clients. Capabilities that were previously requested only by a handful of users are now increasingly expected by a much larger client base. While this is positive from a product and adoption perspective, it also implies that vendors such as BestX must continue to scale infrastructure performance and reliability to meet higher usage and integration requirements.
The final—and more fundamental—issue relates to the changing role of traditional asset management. The traditional model is built on scale economics: institutional investors typically achieved better liquidity and pricing by trading in size. However, AI and tokenisation are reshaping this landscape. Retail investors are often viewed as less sophisticated than institutional investors, but AI may reverse that assumption. Institutions face heavier regulation and governance requirements, which can slow innovation, whereas retail participants often have greater flexibility to experiment and adopt new tools quickly. This is evident in the crypto market, which is still largely retail-led. The author has observed a much faster innovation cycle in crypto than in traditional finance: in 2019 the narrative was largely centred on Bitcoin, but within six years the ecosystem has broadened significantly, with developments such as on-chain prediction markets, decentralised lending, and Layer-2 solutions.
Another key point concerns the erosion of the “wholesale” advantage in traditional markets. Crypto-native retail investors can trade fractional Bitcoin or altcoins 24/7 at basis-point-level fees, with settlement in minutes depending on the protocol. Equity markets, however, still reflect legacy constraints—round-lot conventions, market-hour restrictions, and settlement cycles such as T+1 (historically T+2)—that can make participation and execution less efficient. For instance in equity market,round lots have traditionally been defined as trades in increments of 100 shares. As stock prices increased overtime, it became infeasible to trade in round lots, especially for higher-priced stocks. For example, Tesla closed at USD 417 on February 12, 2026 and this means each round lof of Tesla is approximately USD 41,700 notional. This in turn led to a significant increase of odd-lot trading for higher priced securities. Odd-lot executions have historically received weaker transparency and are excluded from the National Best Bid and Offer computation.
The author has observed cases in BestX equity data where odd-lot orders occurred outside the day’s high–low range, consistent with transparency gaps. Recent regulatory initiatives partially address this by redefining round lots based on price —for example, setting the round lot to 40 shares for stocks priced above USD 250 showed in table 1.
According to Figure 4, BestX data shows an immediate post-change improvement in spreads, suggesting a positive impact of the regulatory update. This also highlights a broader economic point. Historically, when processing and technology costs were high, it could be rational to batch smaller trades and process them in bulk—accepting some loss of market efficiency—in exchange for lower operational cost, reinforcing a “wholesale” edge. In an AI-driven environment where marginal technology and processing costs fall, that trade-off weakens: batching becomes less compelling both from a liquidity perspective and from a cost perspective, and the structural basis of wholesale advantage narrows.
In sum, the buy side is becoming more sophisticated in response to competition from increasingly capable individual investors. At the same time, the foundations of traditional wholesale advantage are gradually eroding—accelerated by AI and tokenisation. This is a key shadow over the AI era for traditional asset managers.
The impact on sellside and platforms
Building on the discussion above, data, AI, and crypto are making institutional buyside clients increasingly sophisticated. While this is partly a natural outcome of the technology cycle, a more fundamental driver is the fear of disintermediation: increasingly capable individual investors—reinforced by crypto’s self-custody model—could erode the traditional role of institutional intermediaries. These changes also ripple through the broader market ecosystem, creating a challenging outlook for the sell side, technology vendors, and MDP platforms. Let’s therefore examine the challenges already observed in the market and propose potential solutions to help participants adapt to this trend.
The emergence of regional banks
Traditionally, regional banks have outsourced FX trading and algorithmic trading to global banks for several reasons: credit, liquidity, and—most importantly—the lack of an in-house technology stack or insufficient economics to justify building one. In this setting, it was rational for regional banks to outsource or white-label FX trading services from global banks, and charge a distribution (or “pipeline”) fee. However, as buy-side sophistication increases, this business model becomes less sustainable. For sophisticated clients who can directly connect to global banks, the value proposition of paying an additional fee to access effectively the same product is weak. To remain competitive, regional banks increasingly need to develop their own trading capabilities—both to differentiate their offering from global banks and to defend client relationships.
As an illustration, I classify the top 10 banks by FX trading volume in 2025 as global banks, and all others as regional banks. Figure 5 shows a gradual increase in market share for regional banks. One driver is a direct strategic response to the challenge outlined above: regional-bank executives are investing in in-house trading capabilities to deliver differentiated execution and liquidity, rather than operating primarily as distributors of global-bank services. At the same time, the emergence of AI has reduced both the complexity and the cost of building and maintaining these technology stacks. As barriers to entry fall, sell-side competition is likely to intensify further.
The Spread War
Economic theory implies that greater buy-side sophistication and more sell-side players lead to spread and fee compression. This is increasingly visible in practice: sellside firms face sustained pressure on spread and fees. In the past, relationship-driven markets allowed banks to earn franchise P&L from sticky client flows. Today, TCA data and AI-driven analytics—together with more capable regional-bank competitors— have raised transparency and intensified competition, pushing spreads lower.
Figure 6 provides direct evidence of long-term FX spread compression. Notably, even when spreads briefly widened during the Liberation Day volatility episode, they rapidly reverted to near historical lows, signalling a more efficient market and a tougher environment for sell-side profitability.
In the AI era, as competition in commoditized products becomes increasingly intense, leveraging TCA data—such as that provided by BestX—can be critical for identifying defensible advantages and building differentiated offerings around them. For example, banks can use sell-side insight tools from TCA providers to diagnose where they are already competitive (e.g., in STIR pricing) and then translate those strengths into targeted client solutions, such as algorithmic execution strategies with optimum forward rolling. More broadly, banks will likely need integrated e-trading teams that combine spot and STIR traders with execution quants. Such cross-desk structures enable the design of unified client solutions and help sustain a competitive edge in an environment of persistent spread and fee compression.
The core function of multi-dealer platforms (MDPs)—and many other tradingtechnology providers—is to connect buy-side and sell-side participants. However, as AI reduces technology costs and shortens development cycles, one of the historical differentiators of these platforms—the proprietary technology stack built over many years—is increasingly being challenged. It is now easier for clients to build their own infrastructure to access bank liquidity and algorithmic execution directly via APIs, rather than routing flow through an MDP.
At the same time, the sell side is facing structurally lower trading margins and higher regulatory and balance-sheet costs. For example, BestX estimates that SACCR can add roughly 0.1 bps cost per trade. Once platform fees are included, receiving flow via an MDP can become uneconomic for some banks. Consistent with this dynamic, Figure 7 shows that algorithmic volume executed via direct APIs has been growing much faster than volume routed through MDPs. This raises an uncomfortable but fundamental question for MDPs: what is the core value proposition they offer to both the buy side and the sell side in an environment where direct connectivity is increasingly feasible?
Answering this question is not straightforward. A plausible path to resilience is for MDPs to evolve from being primarily connectivity venues into integrated, easy-to-use product ecosystems. One advantage that direct API connectivity typically cannot replicate is access to aggregated optin pool data. When clients trade bilaterally with banks, they only observe their own execution history, which provides a limited and potentially biased view of the market. By contrast, an MDP can combine pooled data with AI and trading-technology tooling to deliver an integrated solution—enabling users to leverage shared datasets, generate richer execution insights, and ultimately improve execution outcomes.
performance using data-driven analysis rather than relying primarily on client relationships
Conclusion
Although AI is still in its early stages—arguably in its “infancy”—it has already begun reshaping the economic landscape. On the buy side, traditional asset managers are increasingly challenged by highly sophisticated retail participants emerging from the crypto-native ecosystem. On the sell side, intensifying competition and regulatory scrutiny are compressing spreads and margins.
To conclude this article, I wish to pose a philosophical question: if AI were to reach human-level intelligence—or surpass it—could it be said to possess a soul? This question recalls Roy Batty’s monologue from Blade Runner:
“I’ve seen things you people wouldn’t believe. Attack ships on fire off the shoulder of Orion. I watched C-beams glitter in the dark near the Tannhauser Gate. All those moments will be lost in time, like tears in rain. Time to die.”
