Late 2016 the word was out that change in customer behavior, big data, and artificial intelligence would reshape the financial sector as we knew it. Various manifestations around big data and machine learning caught also OP’s attention and we got involved quickly.
Our hypothesis was that traditional vision/strategy/ roadmap/business case/design/implementation/launch path wouldn’t realize the potential for us quick enough. Therefore, our AI transformation journey has been a bit different, and change we embarked on has created interest. In this article, I’m sharing my thoughts on the key choices we made on our journey and, while at it, nip some persisting myths about AI in the bud.
Vision & Strategy
In order to set and AI vision and strategy, concrete first-hand experience from AI is essential. The concept of AI is far too abstract to create genuinely useful Vision without experience. OP we did kicked off AI program which spent the first few months of the journey on use case identification and pilot development work, instead of starting with vision. During the two year program, the funding was allocated with rolling roadmap with 6-9 month visibility forwards.
Commitment & POC’s
Consultant driven myth is that six-week Proof of Concept’s will save the world. Functionally broad enough POC is hard to implement well when working within the constraints of time, availability of data and needed technical set-up to bank environment. Doing the POC at vendor environment is really not an option, if any GDPR data is needed outside bank’s own environment. This leads that genuinely validating the results are inevitably too limited and the POC results are often open to interpretation or just plain poor. Therefore, after seven or so POCs like that, interest and enthusiasm tend to die a slow death. At this point, commitment is put to a test: should we bite the bullet and carry on or take a break with AI?
To identify use cases you need someone experienced with AI. Recruiting a small AI team (product owners and data scientists) for an organisation lacking current know-how is easily six months job. Identifying use cases can’t be put on hold for six months, though. At OP the core team was ramped up in few months and consisted of both consultants and in-house experts. That said, identification of business opportunities cannot be left solely to consultants. At OP, the inhouse team was grown from 5 to 30 during two years.
Identifying business opportunities
There must use cases with tangible value to operational efficiency in AI portfolio. Otherwise, it’s too easy to pull the rug from under the whole AI thing. The use cases should be about seeking out business challenges–risk, efficiency, customer experience, and growth. It’s better to aim implement fifty minor cases than five major ones in a year. Putting all allocation on few individual use cases will likely fail, as there will be plenty of uses cases where risk management and other aspects will prevent the implementation of the use case in full or in part. With fast track approach, it should be accepted that all results aren’t going to be fully integrated into existing products or services. That’s OK. Fragmented results are fine while experimenting.
“OP chose to develop in-house platform for e.g. working with notebooks and using Spark. This was a major investment and I currently think that commercial solutions would have better way to go”
Financing – business or centralised?
Business line interest to invest in AI should not be overestimated. The benefits of the logic can be easily analysed in use cases which promote customer understanding, growth, and risk management. On the other hand, it’s tricky to quantify the benefits in terms of money. How much more will we sell if we understand our customers better and how much better will our services be if we have a better understanding of their needs? Can our chatbot respond to 5% or 50% of our customers’ queries? OP chose to finance the work from the project (not from business line budgets) for at first two years. The impact of AI use cases started to formulate within the first year and OP is currently well past EUR 24 million Euros benefits on operational efficiency only.
Tools and technologies
Not general AI solution meets all IA use case needs. Technically simple end is e.g. customer propensity scoring model can be developed and implement on laptop and results stored to database. On the other hand, there are API’s serving thousands of requests per second. One-time analyses are all about modeling know-how, whereas technically more challenging cases rely on programming competence. OP chose to develop in-house platform for e.g. working with notebooks and using Spark. This was a major investment and I currently think that commercial solutions would have better way to go. OP started to develop AI both to Azure and AWS, but we soon understood that supporting high-volume production systems on several platforms requires significant investment and scaled down to one cloud approach.
Without data, there is no AI. You can’t underestimate the complexity and quality challenges inherent in data. To a large degree, the complexity depends on the number and contents of data warehouses. If your data warehouse comprises three SQL servers, migrating their contents to, e.g., Snowflake and starting analytics, doesn’t take that many weeks. On the other hand, if you have three hundred data warehouses or some of the data resides directly in operative systems, transferring the data to the cloud and data quality management and master data management are a continuous headache.
So did OP manage to achieve its two-year targets with the help of AI? We did and we were surprised by the benefits gained in different areas. We did great in gaining customer understanding and seeking growth, and the risk management benefits were a positive surprise, but the biggest surprise of our journey of change was the significant improvement in our operational efficiency.