Compliance is the testing ground for AI as investment banks, asset managers and hedge funds seek to roll out the technology across their entire business.
Even before AI, compliance had a tricky balancing act. Like many other middle and back-office functions compliance is seen as a cost centre, where expenditure should be minimised. At the same time, regulation has become more complex, in part because there is more but also due to the rise of new assets, more complex contracts and the wider global reach of investors.
As firms pilot use cases, the compliance department will help businesses decide the balance between control and autonomy. Under pressure from fragmented workflows and higher volumes of communications data, compliance is the perfect fit for what AI is looking to achieve. Let it do the ‘grunt’ work so that teams can focus on the higher value tasks.
Certain AI technologies are more advanced than others. Within communications surveillance, transcription has come a long way over the last five years. It can accurately take down entire conversations with few errors, no matter the accent of the speaker. For trading desks who still rely on voice, an AI tool can turn thousands of hours of conversations into text that can be indexed and searched.
Moving further downstream, another AI agent could then identify worrisome keywords, a second agent could generate a suspicious trading report and a third could submit that report to the chief compliance officer or even the regulator. There is the potential to have an unbroken chain of AI agents monitoring, flagging, writing and submitting suspicious trading reports.
But no firm would put in such a system in practice. There’s too much risk. It might hallucinate, generate a false positive or negative, or miss a suspicious trade entirely.
Even the most advanced models cannot be trusted to be fully autonomous. From our conversations with compliance leaders, most firms have found AI to be good enough, but they remain cautious about where the regulatory lines sit. Errors will happen but to the same level as human error now. The pilot projects are less about testing if AI can do the job but more where to put in controls or checkpoints and where AI can be given a freer hand.
Building guard rails
AI has so far been given a high degree of autonomy in tasks that humans previously didn’t do. Transcribing every voice conversation was never feasible, but with AI it’s possible.
The line begins to blur when analytics are involved. An AI agent can to a high degree of certainty find and identify trades that are abnormal. Here is where many firms start putting in ‘checkpoints’ and where new workflows are being introduced.
This is already becoming a practical problem. We’ve heard examples of firms operating surveillance models they had inherited and could not fully explain, with the vendor chain sometimes unable to explain them either. As regulatory expectations around explainability harden, deploying a model that nobody can properly interrogate becomes a compliance risk in its own right.
A human at this point will look at the report and see if the trade is suspicious. Crucially, the report will need to have a full audit trail. The regulatory test is likely to be whether the firm can explain the outcome, evidence how the system reached it, identify who is accountable and demonstrate that a human was able to challenge the model’s conclusion.
The person looking at the report should have easy access to all data an AI is using to make the assessment, they are essentially double-checking the AI’s work. If the report is accurate, it can be passed further along, if not, the mistake is caught early.
Human intervention then becomes mandatory when these reports go out externally be it other counterparties or the regulator. The risk is no longer just being wrong but the reputational risk of being wrong. Here a human makes the final judgement call.
AI is being added at speed across the industry but including it in an ad-hoc manner creates more risks than benefits. The technology is not a simple plug and play. Workflow will need to be designed from the ground-up to fully realise its benefits.
That involves formalising procedures that were previously unwritten. Take for example, a full audit trail, it is necessary for firms to verify AI’s output. Whereas, in the past this existed more informally as a human could readily explain why they’ve flagged a trade and what data they’ve used.
AI can be used to transform compliance workflows and analyse communication at scale. Human oversight through defined parameters and final decision making will help reduce risks. All of that requires deliberate action through building AI-focused workflows that address and minimise its potential pitfalls.