FI’s face the challenge of monitoring billions of transactions, identifying potential suspicious activity and then handing these cases over to armies of investigators. With a SAR being filed every 30 seconds Financial Crime News met with Caspian CEO and Founder Chris Brannigan to hear how Caspian can help make a difference.
1. FCN: What can you tell us about your company and your products & services?
CB: At Caspian our area of expertise is in the automated investigation of high volume, complex risk alerts for financial service firms with a fully explained human readable decision. So, we are not in the business of generating alerts for suspicious behaviours in KYC, transaction monitoring or due diligence, many other machine learning companies do that already. Once the alert has been generated and a well-qualified human analyst must perform an investigation, that is where Caspian technology is applied. We have developed a unique Financial Investigation Platform (FIP) that combines machine learning with human intelligence to fully automate or significantly augment high-volume risk investigations and judgements. As alerts are generated, they are fed into Caspian FIP. The system performs an expert investigation on the alert to pull in relevant data, generate key evidence, make an overall risk decision and output a human readable explanation that provides a logic and rationale for the decision.
Where Caspian FIP is unable to make a full determination or requires further information, then it will present the investigation it has generated to a human Analyst and it will provide guidance to the Analyst to assist their work in completing an expert level investigation. Where FIP is able to make a fully determined decision and supporting rationale, then these completed investigations are either presented to human Analysts for a ‘quick review’ and / or are sampled by human QA in the normal risk process.
Our proven investigative technologies provide banks with substantial increases in speed, accuracy, consistency and transparency across domains such as Anti-Money Laundering Transaction Monitoring, Customer Due Diligence and Entity Risk.
Ongoing decision assurance is also a key area of focus for banks and we achieve this by utilising our tools and processes. These continuously test and improve the machine decisions beyond human regulatory standards.
2. FCN: How did you come up with the idea and what drives you?
CB: I am a founder of Caspian and met my co-founder while we were both undertaking PhD’s in cognitive neuroscience. This was in the 1990’s when we called AI, ‘Parallel Distributed Processing (PDP)’ and even the concept of ‘back-propagation’ was pretty new and exciting to geeks like me. I was particularly taken by the research of Anders Ericsson into elite performance across sports, medicine and business and into how the acquisition of expertise could be codified and measured within a scaffolding framework. The driving force behind Caspian has been the idea that software, analytics and cognitive models can be usefully combined to precisely diagnose, measure and enhance human performance.
When we founded Caspian several years later, we focused more specifically on improving the quality of human decision making in high risk, high consequence environments. Imagine flight simulators for the workforce and you’ll get the general picture. Our main work was in Defence for organisations like NATO, MoD and DoD but we also extended this approach into domains with somewhat less fatal consequences including pharmaceuticals, management consulting, banking and construction. We would immerse human trainees into the environments and task scenarios that they would experience in the military theatre, and then capture how they managed to achieve objectives as the scenario unfolded. We would build a detailed cognitive model of how expert practitioners thought and acted within these scenarios and embed this into the simulator. The system would then precisely measure and diagnose trainee performance against the models for expert practitioners.
In developing these training solutions, we would produce very granular models of expert performance in a range of scenarios, using increasingly sophisticated algorithms. The old PDP algorithms evolved into the more muscular Deep Neural Networks and data from front end workforce systems became accessible and abundant (slightly skipping over IT and data issues). We found that the environment simulator was now not required. We could embed these solutions into the actual work environment and operational systems.
By happy coincidence, at this point we were working for a large Bank, deploying simulators to improve human decision-making performance in areas of Financial Crime such as Sanctions Screening Investigations. We found that this new environment of Financial Crime shared many of the attributes that we recognised from the military domain – complex tasks, high risk, high consequence, deep audit, traceability and large numbers of humans in the process, with expertise relatively hidden in the heads of geographically dispersed individuals.
From the success of simulators in Sanctions Investigations we moved into the area of AML Transaction Monitoring and applying the approach to directly automate and augment the L1/L2 investigations. We’ve now built this into a container-based platform, that can be deployed at scale to banks’ private cloud, and are applying the same technology to other AML problems including KYC and EDD.
3. FCN: What is it that makes your company different from others?
CB: Our technology and approach is different. Strangely, for a tech company in this space, we don’t start from the machine learning and the raw data. We have a heritage in neuroscience. We started by building a model of how expert investigators think and act. The system that we have developed replicates that.
We have spent 24 months working with banking partners and Financial Crime experts observing, capturing and testing how expert investigators gather evidence, judge and evaluate evidence, make risk decisions and then explain those decisions. The resulting cognitive map is a detailed blueprint from which we have developed algorithms that can ingest bank financial data and automatically perform a complex end to end investigation. The machine performs the wide range of cognitive tasks that an expert investigator would perform. Within this framework the machine is able to make a risk determination and then to explain how and why it made that decision, while pulling through the accompanying evidence to audit that rationale against the standard of the expert consensus. No black box.
We compliment this with unsupervised machine learning approaches that work directly on the data to find anomalies that experts don’t know.
This is live in a banking environment, investigating complex alerts, generating evidence, making risk decision recommendations supported by explanations that are validated against the standard of the best human experts.
Looking at the overall market, there are not many companies that focus on automation of end to end investigation. It is not an easy problem to solve. For Caspian, generating models to automate and augment expert decision making has been in our DNA for the past fifteen years. Our expertise and R&D are wholly configured to that goal.
4. FCN: What barriers do you see to the adoption of AI in Financial Crime and Compliance?
CB: There are many challenges to successfully implementing new technology within large, complex organisations. These are germane to any enterprise software and it is up to vendors to minimise these issues for Banks. If we consider barriers that are specific to the AI use case, then our focus is immediately drawn to the area of risk and, to model-risk management.
The ‘black-box’ problem is well known to machine learning practitioners and is a major challenge for banks that wish to exploit the benefits of AI in areas of complex decision making. Risk Stewards and regulators must know exactly how the machine made the decision. At Caspian, over the past three years we have cognitively mapped how expert investigators gather evidence, make risk judgements and then explain decisions. We have developed AI technology that can make decisions and then explain how it reached that decision, within that expert investigation framework. Every decision can then be fully explained and mapped back to the decision, logic and evidence that a consensus of expert investigators would have taken.
The black-box problem is the first challenge to overcome in FCC. However, AI practitioners are only just beginning to appreciate the exacting standards of evidence and proof required to deploy in the high-risk, high-consequence environment of FCC investigation. Every decision generated by a machine must be traceable and reproducible through the entire cycle of training a model through to its deployment in production. This is a non-trivial problem for AI solutions that are inherently probabilistic in nature. Furthermore, to solve anything more than the most basic FCC task requires an AI solution comprising multiple models chained together within a workflow. This requires mathematical innovation, but it additionally complicates the ability to deliver required traceability and reproducibility of results.
Then there is validation of the performance of the AI solution. Most banks and AI vendors are not experienced in validating the performance of an integrated AI architecture that features multiple models acting in concert. Unfortunately, looking to regulators for guidance does not offer any immediate respite. Then, just when Banks make progress on complex model risk management, there is the cold realisation that these systems require sophisticated levels of ongoing model maintenance as they must deliver to an ever-changing FCC environment.
At Caspian we have spent several years designing this capability into the foundations of our Investigative technologies. The development of this infrastructure requires a lengthy investment in R&D and testing with Banks. It is worth it. My view is that model-risk management can be turned from a major barrier into a significant enabler of AI adoption. Right now, it is not possible for a risk executive within a bank to diagnose with any confidence how their human analysts vary when making cognitive judgements on important risk questions within an investigation. A well designed AI solution enables risk executives to precisely determine how a machine investigator makes risk judgements at an unparalleled level of granularity. For example, how the system makes a Source of Funds determination between attributes such as salary, employment expenses, dividends and bonus and then also how this determination contributed to an overall risk decision at an alert level.
This affords a risk executive the power to deliver risk policy consistently into their investigation activities and risk judgements across the organisation. This is a very powerful risk management tool. We are only just beginning to explore this new capability with Banks, and it is an exciting opportunity for AI solutions in the near future.
5. When are the machines taking over the world?
CB: Not anytime soon!! There is a lot of hype around AI. As a practitioner, I am excited by the opportunities to innovate using current methods. But, the notion of autonomous, self-learning AI is the stuff of science fiction at this point. As someone schooled in the era of PDP and connectionism, the current methods of AI are a reuse of some pretty old algorithms with super powered computation and data volumes that were unimaginable only twenty years ago. In the early noughties, I would guess that there were not one million polaroid photographs of pet cats in the whole world, never mind on a single computer, ready to train a Deep Neural Network.
I often joke that when I am talking to investors I refer to ‘AI’ but when I am talking to banking customers, then I refer to ‘machine learning’. It is easy to get caught up in the hype. The advice that I would give to Financial Crime customers is to focus relentlessly on the business requirements and to use that language when talking to vendors of ‘AI’. I look forward to the day when customers do not mention AI and we talk about measuring risk thresholds, consistency, decision quality, investigative explanations and evidence trails.
From our own work, we have already seen the potential of AI to radically augment human performance on complex tasks of investigation. This baseline is moving quickly into AML level 3 type cognitive tasks, and it will not stop there. ironically the fastest path to the future, and the biggest opportunity for Banks is for human investigators and AI systems to work in concert. The AI will generate evidence, recommendations and an explanation for risk decisions on ever more complex cases. A well trained human expert or teams of experts will act on these inputs and will also continually train the system to improve. The work of the human investigator will evolve quickly, requiring a higher level of expertise and experience than the current L1/L2 entry.
While the Terminator may not be on the horizon, I don’t want to underplay the science and innovation that is required to enable current machine learning algorithms to achieve transformative business results. At Caspian, we’ve built a deep bench of PhD’s across machine learning, neuroscience and computer science working in tandem with Financial Crime experts and experienced developers to turn R&D into powerful investigative solutions. Our mission is to create the worlds most powerful Financial Risk Investigator and to make this available as an easy to consume service to any organisation of any size, 24/7. Inventing a system to do this takes a lot of innovation and effort under the hood.
6. FCN: What is it like working at Caspian?
CB:We’re based in an open plan office in Newcastle upon Tyne and there are just over 40 of us now so we’re still at a size where everyone knows each other. Having a good cultural fit for new and existing team members will always be important in helping us to keep innovating and solving industry challenges.
We work with Agile methodologies and cross functional teams – data scientists, Financial Crime experts, behavioural scientists and developers – working together in small teams to solve their project stories. This approach links our programme of R&D directly to the needs and requirements of the solution. Everyone gets to participate in this.
We’re building solutions that are genuine world firsts so it’s a great reward and motivation for everyone to be part of that. We have strong links with local universities for knowledge transfer, placements and internships.
There is a big opportunity in the financial services sector to make a difference through technology and we’re very focused on being one of the companies that contribute to that difference.
7. FCN: How do potential customers and investors get in touch with you?
CB: Potential clients can gain additional information at www.caspian.co.uk and book a platform demonstration at www.caspian.co.uk/demo. Investor enquiries can be made directly to Chris Brannigan, CEO and Founder at email@example.com
Chris Brannigan is the founder and CEO of Caspian. Chris is a trained neuroscientist and performance simulation expert with vast experience in Banking and Technology Strategy.