For most companies, the pursuit of artificial intelligence (AI) solutions often focuses on two goals: reducing costs and driving efficiency. These are noble pursuits, and surely COOs who bring projects to their board with cost and efficiency in mind receive praise. Focusing on these goals alone, however, may be ill advised. The path to delivering true, lasting value is through harnessing intelligent automation that can act as a catalyst for end-to-end operational change. The goal, then, is not just building a slightly better version of an existing tool or process; it’s about creating an entirely new way to operate. This approach — what we call pragmatic transformation — still has cost reduction and efficiency at its core, but it has the potential for a much greater impact.
Like many in our industry, BBH has spent a considerable amount of time exploring how AI and machine learning can enhance the back and middle offices. Based on our experience using robotic process automation (RPA) to automate discrete business processes, there was a natural evolution toward combining RPA with AI. After many iterations, we launched two AI and machine learning initiatives and the lessons we learned from our experience were profound. In creating the framework for AI, we went through an operational metamorphosis, emerging with new processes, controls, and buy-in from stakeholders spanning the entire business. We could also be introspective about some of the bigger themes that are shaping our business and our industry: operational efficiency, design thinking, agile project management, the future of work, employee roles and skillsets, and enterprise communication, just to name of few.
As we reflect on our own AI journey from pilot program to pragmatic transformation, we’d like to share three success factors that helped us achieve our goals.
How do we define AL and machine learning?
Artificial intelligence and machine learning are broad descripitons to describe systems that can think, learn, and perform actions based on a defined objective. Al can come in different forms to perform tasks faster, improve decision making, and automate not just the tasks, but the decisions behind the tasks.
Success factor 1: Invest in business problems, not AI
It’s clear that the big question for our industry is not “if” or “when” to invest and apply AI, but “how.” As we looked to implement AI, we learned that selecting the right use cases was crucial; but too often firms focus on an exciting technology solution without actually considering what it’s solving for. This, of course, is easy to do, especially as firms establish innovation labs that aren’t fully entrenched in each business line. The first step for us was identifying a real business problem.
Consider an AI solution we created for our fund accounting business. In 2017, we had nearly one million “miscellaneous” cash wires flow into our accounting platform. Each one needed to be manually coded, researched, and resolved. As you can imagine, this was an onerous task, requiring significant manpower and, more importantly, time. To solve this issue, we developed LINC (Language Identification Network Center). LINC uses natural language processing in a supervised machine learning framework to categorize cash breaks, while making recommendations to operations specialists. The machine does this by recognizing words contained in wire transfer text, and the algorithm then codes the description. The result: 95% of entries are automatically coded compared to 65% before we deployed the solution. The second phase of this is moving beyond coding the exceptions to teaching the machine to resolve the breaks. This is where training is crucial. When teaching the machines, we need to ensure unconscious biases don’t produce unintended results; thorough reporting, benchmarking, and analysis will measure how well we’ve taught the machines.
Our second use case was in the data-rich area of Net Asset Value (NAV) reviews. Security pricing within NAV systems must be reconciled each night to ensure the NAV is 100% accurate. Using the traditional fixed thresholds, we were producing thousands of exceptions for our analysts to review each night, with less than 1% representing a true issue. To address this problem, we created ANTS (Anomaly NAV Tracking System) which uses predictive analysis and machine learning to eliminate the noise of false exceptions and brings forth exceptions that might not have been previously detected. Analysts, in turn, spend less time sorting through recurring exceptions thus enabling them to focus on resolving true NAV threats.
In building these solutions, we learned to differentiate between “automating to replace” and “augmenting through enhanced intelligence.” Automating to replace is typically filled by RPA. RPA, unlike AI or machine learning, is often a solution for tasks that have a series of repeatable, manual steps with no complex judgement required. Augmenting through enhanced intelligence, on the other hand, can automate subjective and judgement-based tasks on top of many of the rote manual tasks RPA is already doing.
To recap, selecting the right use cases is crucial. To start, consider:
- A real business problem you need to solve
- An area that is rich with data – structured and controlled data is a prerequisite
- Whether you want to “automate to replace” or “augment through enhanced intelligence
Success factor 2: One step at a time – operating through iteration
For many COOs, building support for AI adoption can be a daunting task. It certainly was for us. The prospect of expense creep, risk of failure, and redesigns can often detract from the success of a project. Building confidence is key. That’s why we turned to the agile/iterative approach when choosing how to build, deploy, and ingrain AI into our own operating model. For instance, when building LINC, we added capabilities in stages and could refine incrementally, as figure 1 illustrates. This allowed us to celebrate victories along the way, while also keeping our goals aligned with the big picture. And by collating feedback and input from people across the business, we were able to generate support beyond our fund accounting group. In short, we were able to transcend the “silo effect” and give our project much-needed visibility.
We found that for navigating uncharted territories, like AI, you can’t have a standard waterfall approach, since that would assume you know all requirements and scope up front. For AI projects, there is far more uncertainty as to how core systems will react and what data is available to fuel the engine. Starting with a small prototype and iterating based on feedback can help. Agile, however, will not solve these uncertainties, but it will arm your project team with the framework to solve challenges and adapt to new issues as they come.
