AI is brilliant at summarising – but not yet at discovering.

Want to know how good AI is with financial data these days? Here are the findings from the latest BIS Triennial Survey, unlocked via carefully crafted AI prompts:

• Cross-currency swap growth: Turnover rose 39% to $172 billion/day, with activity driven by hedge funds and PTFs up +430%, making them the fastest-growing participant group.
• Asian XCCY surged: JPY (+145%) and CNY (+160%) led cross-currency swap growth – more funding requirements coming out of Asia.
• Compression exploded: IRD compression volumes tripled to nearly $1Trn/day.

How Do I Improve my Experience with AI?

The AI hype cycle is now at risk of saturating our own “context windows”. Different methods of measuring various abilities consistently claim that the latest models are better than the previous generation. But does that chime with the experience of casual users? People enjoy having an answer engine, an emotional support animal and a restaurant guide in their pocket. But how many people think that their work roles will actually be transformed by the new technology? Surprisingly few in my experience.

Why? Because they haven’t seen prompt engineering in action. When AI is guided and when context, metadata and structure are integrated into workflows, it level’s up your AI experience.

Data vs. Words: The BIS Triennial Test

Conveniently timed for the purposes of this blog has been the publication of the 2025 BIS Triennial Survey – check it out from the BIS here.

Should analysts in 2025 really still be reading PDFs and tables by hand? Shouldn’t AI now be doing the heavy lifting? Let’s see.

What AI Shouldn’t Do

First off, don’t ask your favourite LLM for a generic summary. The BIS already publish one – and AI will just echo it:

• FX Swaps dominate
• USD is the main currency (for FX)
• OIS are the dominant instrument (IRD)
• The EUR is the most traded currency (IRD)

That’s not analysis; it’s a parrot.

What AI Can Do: Sharp Summaries, Fast

People wouldn’t read our blogs if we just summarised the same data in a different way to the BIS. A lot of the time, I need to read the data to even work out what I want to look for. But AI can short-cut our understanding, cutting through hundreds of rows of BIS data quickly, to produce a summary of the IRD market like the below:

• OIS dominance: Turnover in OIS surged to $5.1T/day (65% of the market) in 2025, up sharply from 2022 when total swaps were only $4.5T/day.
• Market rebound: Overall IRD turnover rose 50% to $7.9T/day from $5.2T/day in 2022. Turnover had previously declined from 2022 to 2025.
• Euro ascends: EUR is now the largest currency, reaching $3.0T/day (38.5%), overtaking USD’s long-held top position (which was 44% back in 2022).

That is how AI should summarise: focused on change, not just size.

What AI Misses: The Hidden Niches

Analysts have been writing well-crafted, concise summaries of reports like these for decades. But only for the headline grabbing big numbers. Historically, that has meant that a niche never gets the spotlight it deserves.

AI-driven summaries can now move Cross Currency Swaps out of the shadows:

• Market expansion: Cross-currency swap turnover jumped 39% to $172B/day in 2025 from $124B in 2022, but its share of the overall FX-market slipped to 1.8% (from 2.0% in 2022).
• Asian surge: JPY up 145% to $35B/day and CNY up 160% to $7B/day, highlighting Asia’s rising role.
• Europe splits: EUR turnover +46% to $53B/day, while GBP slipped 3%, highlighting diverging trends within the same region.

And can even provide market commentaries on even smaller niches, such as Compression:

• Compression triples: IRD compression turnover surged 190% to $979B/day in 2025, up from $337B in 2022.
• OIS-driven: OIS compression hit $639B/day, accounting for two-thirds of total activity.
• Swaps dominate: OIS + other swaps made up over 90% of all compression trades.
• FRAs fade: FRA compression remained minor at $69B/day, underscoring the shift to swap-based risk reduction.

Prompt Engineering in Practice

Every one of the bullet points above was generated through prompt-engineered analysis using Gemini and ChatGPT in tandem – Gemini for its massive context window, and ChatGPT for its ability to mirror tone and precision.

An example prompt fragment looked like this:

**Persona & Context:**
You are Chris Barnes, a senior financial markets analyst at ActrixFT. Your tone is sharp, insightful, and data-driven, tailored for an expert audience of finance professionals.
You are analyzing financial data from the BIS.

Add in some strict formatting rules, and the result becomes structured, repeatable analysis – as opposed to chatty waffle.

(As an aside: I wish I didn’t have to write lengthy prompts. Apps that have these prompts pre-populated, along with metadata (which the BIS include on their data pages) and pre-aggregated data offer a better experience. I therefore expect to see this type of functionality added natively to the frontier models in the coming releases. They will need it to drive business adoption.)

And then I finally found the limits of the LLMs. It didn’t matter how I pre-aggegated the data or prompted the LLMs, I could not get them to give me a compelling drill-down into participant-level dynamics of the Cross Currency Swaps market.

Summaries? Excellent.
The insights? Still not there.

Insights

I had to go into the data myself and go looking for stories. How has Hedge Fund activity in Cross Currency Swaps changed? Which currency pairs saw particular growth? Is Prime Brokerage really a thing? I pre-aggregated the data and provided a new summary table:

Initially, both Gemini and ChatGPT returned lazy answers because they read only the first five lines of the table! Once told to consume all the data, things improved dramatically:

• Hedge Fund & PTF Explosion: Turnover surged +430% to $22.9B/day, the largest gain among all participant groups.
• Official Sector Growth: Central banks and public institutions up +151% to $8.4B/day, potentially reflecting more active reserve management.
• Non-Reporting Banks Retreat: Volumes fell 13% to $31.9B/day, indicating industry consolidation to the biggest players.
• Non-Financial Customers: Activity more than tripled (+182%) to $15.5B/day – more corporate funding/hedging flows.
• Prime Brokerage Uptick: Prime-brokered trades nearly doubled (+90%) but still remained small at $1.0B/day. This is consistent with increased hedge fund market participation.

It’s pretty good, right? Cross Currency growth fueled by hedge funds. That is an interesting find.

Executive Summary

• AI can summarise, not discover: Large Language Models can condense structured data like the BIS Triennial Survey, delivering sharp, data-driven summaries quickly.
• Prompt engineering is the unlock: When context and metadata are built into workflows, AI is elevated to the role of an analyst.
• BIS data proves the point: AI produced accurate summaries across Cross Currency Swaps and Compression – but needed human intervention to uncover deeper counterparty dynamics.