MiFID II compliance: the essential confluence of regulation, compliance and artificial intelligence

James Foley, Resilient: MiFID II compliance – AI to the rescue!

With the July 2017 publication of its MiFID II Policy Statement, the Financial Conduct Authority (FCA) has signaled the final countdown to the most sweeping changes of governance, protection and transparency for a generation.

While MiFID II will fundamentally change the business models of many firms, especially those in the advisory sector, there are also significant new technology requirements that apply to all firms and which are causing some to have sleepless nights.

One of these new technology decrees is a mandate to record telephone calls for anyone either directly involved in trading or giving advice that may lead up to a trade. The intent behind this is noble, which is to provide transparency against firms or individuals giving bad advice.

However, the new regulations don’t end simply with the need to record these calls, but go further to require that firms actively monitor their calls to identify mis-selling or other bad practices.

With many firms conducting much of their business over the phone, it begs the question as to how, and who, are going to be reviewing these calls – and just how many phone calls need to be listened to in order to satisfy the regulator that the compliance oversight function is operating as it should.

In this respect, the FCA is not entirely forthcoming. The regulator will use words such as “appropriate” when describing the number of calls that need to be as listened to, but this doesn’t help firms in determining what this means in practice.

Should firms plan to listen to all calls, 1% of calls, or somewhere in-between? Estimating this, and being on the wrong side of the line, can make the difference between compliance and non-compliance and all the personal and corporate implications of this. However, while the advice from the regulator may be vague, in practice the regulator is usually very prescriptive regarding whether a firm has adequate or inadequate controls in place. Call recording and review is no exception.

In a review of the Final Judgements that the FCA has made on firms who do record their telephone calls, it appears that over the past five years the FCA expect firms to listen to around 20% of their recorded calls. Whilst this may seem a reasonable degree of sampling from a purely statistical point of view, in the context of a firm that conducts much of its business on the phone, this requirement to eavesdrop to this level could place an onerous burden on the firm. In short, for every five members of staff conducting business on the phone, the regulator appears to expect the firm to have one additional person hired just to eavesdrop on their calls. Some firms have already mooted their need to hire thousands of new compliance staff just to meet the requirements of MiFID II.

Fortunately, while there is new cloud-based technology that makes the recording and storage of calls much simpler than it has been in the past, technology can also play an important part in reducing the compliance monitoring burden too. A new type of artificial intelligence (AI) technology is at hand that should enable firms to meet the new MiFID requirements without the need to hire armies of compliance staff.

Recurring neural networks are an idea that have formed PhD thesis since the 1960’s, but until today there wasn’t sufficient compute power available to the average firm to make this a practical reality. Recurring neural networks are a part of AI that model the way the brain synthesises language, enabling the rapid and accurate transcription of conversations. So, instead of listening to telephone calls, compliance officers can simply specify words or phrases that they are interested in and thereafter they will be immediately alerted to any conversations that contain these trigger words.

This technology can go further to highlight out-of-context words and phrases, or to look at the sentiment of the conversation itself to identify suspicious intent. As such, individual compliance officers can review the conversations of thousands of staff at the same time, with only the need to listen to those which warrant their attention.

Whilst automated transcriptions of calls will never reach 100% accuracy, what this “robot compliance monitoring” can deliver is very rapid turnaround, reasonable accuracy and 100% coverage. After all, isn’t it preferable that firms can have automated compliance monitoring on 100% calls with an 80% accuracy than have humans listen to only 20% of calls with accuracy that, over time and after a long day, may not be as accurate as one would like to think. In this way the robot compliance monitoring may be used simply to highlight to the compliance staff those calls that do merit human oversight.

Recurring neural networks are particularly strong in the area when traditional transcriptions services are the weakest – those that take place in a noisy environment or where accents or dialects can sometime confuse the system. It’s for this reason that all the newer transcription services, from IBM Watson, Google, Apple’s Siri and Amazon’s Alexa use recurring neural networks to provide high quality transcripts of the spoken word.

The problem with these public-domain transcription services is that firms have little control over where the audio transcription takes place, most of which send the audio files overseas to be processed.

With London consolidating its position as the fintech centre of the world, it’s no surprise then than it is London-based firms that are leading the charge in applying AI to these new regulation and compliance challenges. Just as London tech firms pioneered solutions to the FCA’s original mobile call recording requirements in 2014, so too are London firms applying AI technology to the compliance monitoring challenges that such recordings now present.

James Foley, director, Resilient plc