Take a look at any large buy-side trading operation and the architecture is much the same as it was a decade ago: separate teams for equities, FX, and fixed income, each running its own systems, its own data feeds, and its own operational headcount. The asset classes are different, the rationale runs, so the desks must be too.
That logic is now wearing thin. Workflows across the three asset classes have quietly converged to the point where traders are doing functionally identical work – monitoring markets, sourcing liquidity, running pre-trade analysis, executing via algo or RFQ, evaluating the result – with different tools and different teams. COOs are starting to call this what it is: a silo tax. A cumulative cost of duplication that was justifiable when market structures genuinely differed, but is no longer.
A signal that the case for consolidation has hardened came at the recent A-Team Group TradingTech Summit London, where a poll of practitioners returned a 100% expectation of full consolidation into integrated multi-asset trading architectures within the year. Practitioners rarely agree unanimously on questions of operating model. This one is worth paying attention to.
Thirty years of quiet convergence
What makes the silo tax newly indefensible is not a sudden shift in market structure but the accumulated weight of a long one. Between 1995 and 2025, equities, FX, and fixed income each adopted electronic protocols at different speeds and along different paths. Equities led on algorithms – VWAP, implementation shortfall, POV – and on Transaction Cost Analysis. Fixed income built its liquidity model around RFQ and dealer axes. FX gravitated toward streaming prices on multi- and single-dealer platforms.
The post-2008 regulatory wave – Dodd-Frank, MiFID II – forced cross-pollination. RFQ migrated to less-liquid equities. Central clearing extended from listed futures to OTC derivatives. FIX, conceived as an equities messaging standard, became the connective tissue of the entire institutional market. The conventions that travel are the ones that solve a real problem; each of these did, and each took the asset classes a step closer to a common functional playbook.
That playbook now exists. The plumbing is broadly the same across asset classes even where the products are not. Which means the architectural argument for asset-specific desks – that the work itself is too different to consolidate – has dissolved. What remains is organisational inertia and the sunk cost of the systems already in place.
Why AI is ready now (and why that statement needs unpacking)
The temptation, having reached this point, is to declare that AI is finally ready for the trading desk. The more accurate framing is the inverse: the trading desk is finally ready for AI. Generative copilots and large language models have been working their way through the front office for several years, finding traction first in research and idea generation where the infrastructure dependencies are light. Execution has resisted, and for good reason – OMS, EMS, and OEMS platforms are deeply wired into regulated workflows that no generic LLM can replicate or replace.
The complexity functions as a moat. It protects incumbents and prevents unqualified tools from disrupting execution without proper integration. It also explains the split visible in a separate poll at the same event, in which one-third of practitioners expected AI to take over core trading functions within five years and two-thirds said it would never fully happen, citing regulatory constraints and the irreducible value of human expertise.
Both camps are, in their way, right. The wholesale displacement of trading infrastructure is implausible. The deployment of an AI layer on top of that infrastructure is not only feasible but already underway. The interesting question is not whether AI replaces the desk but where it sits in relation to it.
The agentic overlay
The answer taking shape is what might be called the agentic overlay: an AI layer that sits above the unified workflow stack, accepts natural language input from the trader, and translates intent into the structured execution beneath. A trader types “buy $50m USDMXN spot, minimise market impact” and the overlay determines whether to route via RFQ or algorithm, selects the venue, and captures the post-trade record. The trader never fills a structured ticket. The agent does, invisibly, with the precision the underlying infrastructure demands.
This is, in a sense, the closing of a thirty-year loop. Trading began with voice on the floor – intuitive but imprecise. It evolved into highly structured electronic tickets – precise but rigid, demanding that traders adapt to machine logic. Natural-language chat, mediated by AI, restores the intuitive interface without sacrificing the precision underneath. The translation layer is what makes the combination work, and the unified workflow stack is what makes the translation reliable.
It is also why the overlay cannot be retrofitted onto a fragmented foundation. An agent operating across siloed, asset-specific systems inherits every inconsistency in the stack beneath it. Common pre-trade analytics, common TCA, common liquidity sourcing – these are not enhancements once the overlay is in play, they are prerequisites. Firms still running parallel infrastructures will find the agentic overlay misfires, or simply cannot be deployed reliably at all.
What COOs should do next
For execution heads and COOs, the immediate task is auditing what they actually have. A capability matrix – testing whether pre-trade analysis, liquidity sourcing, execution, and post-trade data capture are genuinely consistent across all products – exposes both the operational drag of the silo tax and the gaps that would break an agentic deployment. The two problems share a single answer: a unified, multi-asset workflow layer.
Once that layer is in place, the overlay can be positioned on top, with the degree of human-in-the-loop oversight calibrated to instrument, size, and market conditions. As execution heads grow comfortable with agentic workflows, those parameters loosen. Routine execution moves down the stack. Traders move up it – toward the cross-asset strategy, relationship work, and complex structuring that AI cannot do. The firms positioned to ride that trajectory over the next three to five years are the ones doing the unglamorous infrastructure work now.
The silo tax is the problem. Common infrastructure is the prerequisite. The agentic overlay is the mechanism. None of the three is theoretical, and the timing of all three is now.
Read the full white paper: “The Buy-side Trading Desk of the Future: How Multi-Asset Convergence and the Agentic Overlay Are Reshaping Institutional Trading“, commissioned by LSEG and published by TradingTech Insight.
Subscribe to our newsletter
