Barry Levett
CEO
AI can and should be a game changer for many paytechs, but the devil is in the detail. In my latest article, I’ve decided to work up some examples of where AI is already beginning to make the difference and also look at a couple of areas where the application of AI models will generate positive results much more slowly.
AI is going to make the difference but not everywhere for everyone
When people talk about major tech ‘leaps forward’, we often think about the creation of the transistor, integrated circuits, personal computers, networking, cloud computing or open source programming. Some of these inventions changed everything in IT, some were gamechangers in terms of the cost of building software.
While AI is not a new thing, the launch of ChatGPT 3.5 in November 2022 put large language models (LLMs) onto the public stage and fuelled widespread expectations of immediate changes to everything. AI, which suddenly became synonymous with LLMs in the minds of the public, was declared the next gamechanger. However, many are already pouring cold water on AI given the inherent limitations of LLMs. The naysayers are not seeing the developments of AI in other spheres being built into tools we use every single day.
Yes – AI will change the world! However, its success depends on the problem you are addressing and the quality of the AI model being used. Let’s look at some areas touching the paytech world where palpable, even transformational, change is possible with the use of AI-enabled tools.
Hype cycle doesn’t apply to AI
The thing about big IT developments is very often they are over-hyped at the beginning, and much of the promise fails to materialise, resulting in the ‘hype cycle’ peaking and pushing the overhyped trend into the inevitable ‘trough of disillusionment’.
However, AI will buck the normal hype cycle rules because much of the hard work and invention was completed during the development of AI’s forerunner – machine learning (ML). So, AI did not reach mainstream use in late 2022, it merely came of age.
An ML system was only ever as good as the applicability to the problem being solved, the size and quality of the data set it was trained on and the amount of computing resources available. As resources increased, ML systems grew in tandem. With the internet solving information challenges and plummeting compute costs, enormous strides forward were made very quickly.
So, today you can get hold of a whole range of different-sized AI models, already pre-trained, for free (e.g. from Ollama.com), or on a low-cost subscription base from dozens of providers including Microsoft and GitHub.
The action is not in the LLMs but in other AI models solving real world problems. 6 key drivers are at work:
- Many real world challenges can be boiled down into classification problems, prediction problems, optimisation, or decision making problems. ML is well placed to handle all of these sorts of problems using well understood models.
- The relatively recent explosion of high quality data sets used to train models.
- ML tools such as TensorFlow are easy to use, often free, and built into development environments.
- Wide availability of compute resources at very low cost to do the model training and to support very large models.
- Fast return on investment as the implementation is easy for many companies using free pre-trained models and/or services provided by AI companies.
- The ratchet effect: AI will only ever get better, never get worse!
So, while AI research started in the 1950s, the pre-requisites for widespread use and application of it have only recently become widely available and for this core reason – this time it’s different! There is not a one-size-fits-all though. Many AI systems are speeding ahead while others are lagging, depending mostly on the use case.
AI helping developers to build better code faster
AI can be applied to IT challenges very rapidly and relatively inexpensively. At Mypinpad, we are already seeing the application of AI in software development exploding. Our developers can use tools like GitHub Copilot and its new competitor Cursor, to help them build, review and quality test code much faster than they could even a few years ago.
AI is being used to help write code more quickly, spot likely software code bugs in what’s already been written – flagging problem code and prompting alternative lines of code and then automatically applying them once the engineer has given the AI-generated change suggestions their seal of approval. Engineers can even train this new technology to apply agreed changes across large blocks of code.
It’s helping developers to embrace the Shift-Left movement which integrates quality assurance (QA), security protections and testing much earlier in the development cycle. Just a few years ago, the code was written by a developer and QA tested by someone else. Change requests were then passed back to the original developer, before a second iteration of the code could be developed. Now, much more is done at draft 1 – often by the same engineer. Ultimately, better software will be written faster. This productivity leap for the DevOps world is already beginning to show itself in terms of the speed of the development cycles which Mypinpad is able to push through.
So, which use cases stand out in terms of the use of AI to improve the quality and speed at which it is able to roll out solutions to the market? I decided to look at one area where I predict rapid improvements as a result of application of AI-driven algorithms, and two areas which are more of a ‘slow burn’ in our market. It will be different based on specific market challenges.
Use Case #1: Fraud Protection (Rapid Improvements due to AI)
The key to spotting fraudulent transactions is analysis of data associated with that transaction. Location, value of a transaction, time of day, account balance available, type of items being purchased—are all relevant factors which AI models can check and risk score in the blink of an eye.
I remember hearing a story from an executive working at a well-known scheme that one way to ensure you get your card blocked in the US is to purchase two tanks of gas (i.e. petrol), one for yourself and another for a friend, and then go to a Nike store and try to purchase some trainers. Whether or not this was a good fraud identifier I have no idea. However, the process of identifying patterns which are hard to spot and implementing this logic into the fraud systems manually is both error prone and inefficient.
AI classification models are designed to track many more variables than any human can. They find patterns which are normally invisible to us. These patterns can be immediately applied without changing existing payment flows. It just plugs in.
Use Case #2: User Behavioural Analysis in Cyber Security (Rapid improvements due to AI)
Similarly, AI is being heavily deployed in cyber security to spot more cyber threats based on multiple parameters which again are often linked to the user behaviour of a specific employee and whether their online activity breaches pre-defined risk parameters. So, a cyber security platform today might spot that a member of staff is using a potentially risky cloud-based file sharing app such as WeTransfer to move documents to an email address outside the organisation.
However, if the platform has historical data on that member of the marketing team’s use of WeTransfer to send branded brochures to an outside designer (and therefore gauges that it is for legitimate purpose and does not compromise any personal data controlled by the company), the app’s usage can be allowed for this person but perhaps not for another member of staff. In other words, it is about marrying user behaviour with other risk-linked parameters to determine if prescribed risk levels have been reached or exceeded.
Use Case #3 – Customer Service (Slower Burn)
It is in this area that AI grabbed much of the mindshare of technology businesses in the first six months after the launch of Chat GPT 4.0. There was barely a tech firm anywhere which did not contemplate, proto-type, or even go live with a chatbot to support their customer service function—us included.
However, as we have discovered, these generative AI-based chatbots are only ever as good as the documentation that supports them. Chatbots are like an automatic “Frequently Asked Questions” on steroids. You ask the chatbot a question about your pension policy or latest transactions from your current account and hopefully you get the right answer.
The key here is not to knock back the AI systems but rather to focus on the quality and quantity of documentation provided to customers. It’s good old GIGO (garbage in, garbage out)—the better the content, the better the performance of the bot. Once this pre-requisite is in place, excellence in customer service via AI will follow.
Use Case #4 – New Products & Services (Slow Burn)
We have seen new products and services emerge like Google Maps that are highly dependent on AI and some clever algorithms. It has live traffic information, reports of accidents etc, but there will also be opportunities for AI models to be applied to day-to-day financial transactions like extending unsecured loans, motor insurance renewals, based on heuristics—enabling rapid decision-making based on multiple parameters. These will roll out over time and many will not know to what extent AI powers the decision-making under-the-hood.
Summary
AI is already being widely applied amongst the tech community to improve the quality—making code generation cheaper and quicker to test and finalise. The efficiencies gained here are legion. The cost of applying many of these AI models is close to zero. Take TensorFlow for ML models, Azure AI and Ollama for pre-trained AI models—they’re all free! It’s about developing these models into particular use cases in specific markets now.
Remember the ratchet mechanism applies to AI i.e. models will only get better, and in some areas they are going to deliver rapid improvements. The developers you retain will be writing more bug-free code and building software solutions faster as long as they have the requisite AI skills, knowledge of algorithms, and ability to work with relevant tools like GitHub Copilot, Cursor and Ollama.