By Kirk Wylie, CTO of Quantifi Solutions

As trading firms increasingly employ AI to help build the applications used on the desk, it is critical that they have the right guardrails to use this technology safely. When used appropriately, the technology can be genuinely powerful. However, large language models, by their nature, can often produce incorrect outputs with complete confidence. For a trading desk, a wrong number can mean a mispriced position and a loss that could have been avoided if the proper controls were in place beforehand. 

The firms that will be able to extract the most value from AI, rather than getting burned by it, will need to understand which parts of the desk AI should or should not touch, recognise where it can be safely applied, and actively constrain what it is allowed to do in those areas. 

Knowing where the line sits 

Every trading desk ultimately answers to someone with a fiduciary responsibility, whether it is to the investors in a fund or the shareholders in a bank. Someone has to be accountable for the accuracy of the numbers those parties rely on as no fiduciary will sign off on numbers generated by an LLM. To avoid risking any hallucinations that can be produced by AI, it is therefore critical to protect or better yet, silo your core pricing engine, P&L and risk calculations from this technology. 

By being deliberate about where AI is allowed to operate and ensuring that code is human-generated and validated up front, you can protect the numbers that must be certified while freeing the technology to add real value everywhere else. 

Where best to leverage AI 

One of the most useful areas in which you can leverage AI is in the pre-trade layer. Pre-trade tools, process automation and post-trade analysis such as risk reporting all sit under the desk’s own control. A trader using an AI coding assistant based on a validated model library can build tools that would have once needed a team of software engineers or that never got built at all because they could never justify the developer time.  

This does not mean the displacement of traditional financial engineering roles. Building and validating the models underneath remains specialist, hard-won work. What it changes is who can build on top of that foundation, closing the long-standing gap between what a trader wants from a quant library and what they can realistically put together themselves in a spreadsheet. For smaller and tech-light funds in particular, this can be a real advantage. 

Discipline means constraining the machine 

The catch is that narrowing the gap between idea and tool does nothing to guarantee the tool is correct. A great deal of our own work is on forcing the model not to reason from first principles but to call on a validated library instead. If someone wants to price a Mexican treasury bill, the answer should come from a model built and tested for exactly that, not from the AI’s best guess at how such an instrument works. This matters as you never want your pre-trade numbers to diverge wildly from your post-trade numbers.  

To combat this, desks should insist that any AI coding assistant is wired to draw on their officially vetted pricing and risk calculators and input data, rather than left to improvise. The more complex the system you ask an AI to generate, the more oversight it requires. If the underlying library has not been designed to be used safely by an AI in the first place, no amount of prompting will make its output dependable. 

Where the real advantage lies 

The case for AI often hinges on the promise that you can do more with less people. However, replacing a team of engineers with a team checking the machine, or with test automation that is almost as expensive, you have not removed the cost but just displaced it. The productivity gain is only real when the AI is building on a foundation solid enough that its output can be trusted without re-deriving it every time. 

This is ultimately what will separate the firms that benefit from AI from those that get caught out. The ones that pull ahead will not be those that adopt earliest or fastest, but those that are most deliberate about it: keeping AI away from the core pricing, P&L and risk numbers a fiduciary has to certify, pointing it at the pre-trade work where it genuinely helps, and insisting that whatever it builds draws on a validated foundation rather than its own guesswork.