The deployment of artificial intelligence agents to risk management processes without adequate attention paid to data quality will expose capital markets participants to undue vulnerabilities.
The warning from Clearwater Analytics came after a survey by the company found almost all institutional investors questioned said that data quality was the biggest barrier to trusting AI outputs.
Kirat Singh, President of Risk and Alternatives at Boise, Idaho-headquartered Clearwater, said that while agentic use in risk workflows was on the rise, many organisations might not recognise the amount of data-quality assurance preparation needed to ensure they don’t expose themselves to even more risk.
“You don’t know what data the agent has pulled or if it’s consistently kept the same assumptions across different asset classes and different models,” Singh told Data Management Insight. “And that’s why, when you’re looking at risk or deciding how to hedge a position or what leverage to take on, you can’t completely trust your agent.”
Industry Survey
Singh’s comments follow the release of Clearwater’s latest global research report, entitled GenAI and the Data Divide, in which it found that 99% of respondents to a survey said they had trust issues with the quality of their data. The report concluded that this was the key reason why many firms had failed to move their AI projects from pilot to production.Data quality as a determinant of AI success has become a key focus within the data management space as near-autonomous agents have been increasingly deployed across capital markets. Without good data going into the models, the chances of good outputs emerging are slim.
With the spectre of agentic errors compounding through downstream workflows, vendors have been busy girding their offerings to ensure their clients’ data is AI-ready.
AI Critical
The GenAI and the Data Divide report also found that 95% of institutional investors consider AI important or very important to their goals, and 79% expect AI-enabled firms to outperform their non-AI peers.
Nevertheless, just 40% said that AI informs more than a quarter of their operational decisions. And while 79% of executives rate their data as complete, only 56% rate it as accurate.
The findings indicate that a key benefit of AI can become a liability if the data is bad.
“If data quality is an issue, you can’t go fast – and what does AI let you do? It lets you go fast,” Singh said. “Until people feel confident about the data quality going into it, it doesn’t help that you can go faster.”
The chief reason why care must be taken when applying AI to risk management tasks is that the technology is trained to always give an answer, even if it is a wrong one, Singh added. Without strict instructions, it will also make its own assumptions that, without later verification, could throw out the intended risk parameters.“If there’s not enough information in the system for the agent to actually answer your question, what a lot of them will do is one-shot it for you,” Singh said. “There’s a bunch of different aspects of risk that you need to manage.”
Singh cited the example of the failed hedge fund Situational Awareness as one that was precipitated by its AI being ill-prepared to manage risk.
“They didn’t understand leverage,” he said. “That’s the trouble with just asking an agent ‘what’s my risk?’ It doesn’t necessarily take into account all the factors… liquidity, margin and so.”
Clean Data
Clearwater said that its clients had benefited from its own data quality provisions and had consequently seen more of them move agents and generative AI projects into production.
The company saw an almost threefold rise since January in the number of clients building their own custom workflows as the number running workflows rose 69% over the same period. Those clients’ use of their models had also evolved, with automated workflow runs doubling and running at a 96% success rate.
Singh said Clearwater’s clients hadn’t rolled back their AI projects because the company ensures the data they use in their models is clean, complete and verifiable.
The solution to ensuring safe agentic use in risk modelling is to write very tailored, specialised instructions for agents and to limit their access only to data that has been verified.
“Rather than letting the agent say it’ll come up with a model on how to price, for example, an RMBS… you say instead, ‘here’s a model that I’ve picked to price an RMBS, here’s the term structure that the markets imply, here’s the data you use to calibrate that model, and, of course, here are the configuration parameters in those models’,” he said.
“If you do it that way, then your model will be consistent and the agent will be very helpful.”
Subscribe to our newsletter
