A closer look…

FX Europe revealed an industry that is far more fragmented in its electronification journey than the headlines suggest. There have been some significant steps forward: asset managers have reached high levels of STP, often reporting that more than half of all tickets are sent through RFS or automated rules. In the most liquid pairs some are quoting crossover points around 70 million before they switch from electronic to voice, although in pairs like USD CAD that line can fall closer to 40.

Yet for all that progress, traders at FX Europe 2025 were clear that the decisive force shaping the next decade is probably not going to be execution technology. What they value is the quality of the data that execution relies on. The desks with the strongest datasets will automate confidently. Those without them will keep falling back to chat and voice, which perpetuate the gaps.

One Finance Hive member gave a striking example. Small vanilla tickets create flawless datasets that feed TCA, models and post trade analysis. Larger and more sensitive trades do not. They travel across chat, get priced bilaterally and leave almost no structured footprint. This leaves a blind spot at the exact point where alpha and risk are highest.

This is why voice is not disappearing, it has simply shifted mediums. What used to be phone is now chat. EM, NDFs on volatile days, structured options and broken date swaps all still rely on human negotiation. Until banks can stream reliable pricing for these products, that will not change.  What Bloomberg FXGO offers is a robust electronic framework around these ‘voice’ executions.

AI came up repeatedly but with a realism that cut through the hype. Some firms are experimenting with in-house LLMs, some of which hallucinate so heavily that traders are sceptical of any provider claims until the model can prove its workings. Others praised the progress Bloomberg has made in scraping IB chat, surfacing axes and turning hundreds of unread flashes into something actionable. Most saw this as the strongest current use case for AI: not for execution decisions but for signal extraction.

The key insight that emerged was that traders will not delegate decisions to AI until AI is fed trustworthy data. And desks will not produce trustworthy data until voice and chat trades are electronically wrapped, time stamped and integrated into the same workflow as electronic trades.

When viewed through that lens, electronification is far earlier in its journey than many assume.

The Challenges

The central challenge running through every discussion was the sheer inconsistency and incompleteness of the data that underpins execution. While automated small ticket flow produces near perfect records, larger and more sensitive trades continue to travel across chat, where time stamps are unreliable and competing quotes are rarely captured. This leaves TCA with large blind spots and undermines confidence in smart routing. Traders described this as the most serious barrier not only to automation but to any meaningful attempt at performance measurement.

A second issue is the difficulty of trusting automation when the underlying logic is not transparent. Several buy side desks have been told to automate, sometimes by senior management, yet cannot explain why certain flows are sent RFS rather than algo or why slippage limits are interpreted in particular ways. Without explainability there is a real risk that automation becomes more expensive than voice rather than more efficient.

Another consistent challenge is the fragmentation between currency pairs. The level of electronification is dictated by liquidity, and the crossover threshold shifts constantly. In EURUSD firms may stay electronic up to 70 million, yet in USD to CAD the cut off may be closer to 40. These boundaries determine both execution quality and data quality, and because they move with market conditions many desks struggle to keep their playbooks up to date.

AI also faces structural obstacles. Several firms experimenting with in-house LLMs reported hallucinations, weak comprehension of trading language and inconsistent outputs. This has made traders sceptical of any claims that AI can support execution. As long as models are trained on incomplete or unstructured datasets their use will remain confined to research, insight gathering and chat filtering.

Meanwhile chat itself has become a source of both opportunity and overload. Traders often deal with hundreds of flashing chats in the morning. Important axes risk being buried in noise. Bloomberg’s efforts to scrape and structure this information are promising but still depend on banks formatting messages cleanly and consistently. Even small variations in how axes are written can confuse machine learning systems, but progress continues.

Finally, swaps and broken date pricing continue to hold back full electronification. Despite improvements in streaming, banks still rely heavily on manual pricing for these structures. This forces traders back into chat and deepens the data gap that everyone is trying to solve.

Next Steps: what will move the needle

1. Use electronic wrappers for voice and chat trades to capture time stamps and competing quotes.
2. Map currency and size crossover points and update them regularly as market conditions change.
3. Demand transparency from automation providers so traders understand the logic behind routing decisions.
4. Prioritise AI for areas where it is already reliable such as chat scraping, liquidity diagnostics and anomaly detection.
5. Build datasets for EM, NDF and broken date swaps even if some inputs start manually.
6. Consolidate workflows so all trade types feed one clean dataset for TCA and decision support.
7. Push banks to improve message standardisation so chat scraping and pattern recognition become more accurate.

Source: FX Europe Member Meeting, London, 4 December