In a rapidly changing financial environment, the race is on to implement AI. It’s no longer a matter of if an organization is going to use AI, it’s when and how they plan to implement it in business. The big question is whether they should develop in-house, buy off the shelf or acquire a custom solution. Here’s what the options are, what they offer and what to look for when choosing an AI solution.
Surveys of major financial organizations show they recognize the need to use AI to leverage their complex data and mitigate business risks. MIT Sloan Management Review surveyed more than 3,000 managers and interviewed executives. Over half said their companies are implementing AI (57 percent), have an implementation strategy (59 percent) and understand how AI can generate value for their businesses (70 percent).
They also learned a majority of companies tried developing their own AI, but only 1 in 10 gained significant financial benefits.
Success depends on many factors and choosing the right development platform is a key factor in how well your project goes.
DIY AI – beware the hidden costs
It’s common for large corporations to have data science departments while smaller organizations may be starting to explore AI. Frequently, these teams are tasked with building their own AI models with varied success.
Nearly 80 percent of AI projects don’t scale beyond a proof of concept, reports CompTIA. “Businesses often face challenges in standardizing the model building, training, deployment and monitoring the process,” the article states.
There can be benefits to in-house development. Organizations have complete control of their projects and full ownership of the platform. Open-source products provide lower-cost solutions that may be attractive to companies with the expertise to support them. But often companies “don’t know what they don’t know” until they’ve invested the time.
Even when projects work, commercial vendors will come out with something that is better, cheaper and faster, quickly outpacing what internal teams worked so hard to build.
Using open-source solutions
Open‑source software is commonly used to build platforms. It’s inexpensive but has limitations. Open-source can leave organizations at risk if it hasn’t been developed for their specific use case or business challenge. Software that hasn’t been thoroughly tested may not perform well, have data biases or leave gaps in security.
Proprietary systems may have hosting issues, need multiple servers, and require ongoing investment for hardware and software updates. Depending on how familiar the company is with AI, it can take months and even years to develop, test, build a governance framework and finally deploy. Yet for smaller organizations with simple AI needs, open-source may fill the need.
A good AI specialist is hard to find
Hiring for in-house teams can be problematic, given the fierce competition. ITCareerFinder, in ranking the top 10 IT skills that are in demand for 2023, notes that employers are moving to skills-based hiring and compensation over degrees, perhaps a nod to the industry’s rapid growth.
At the time of writing, LinkedIn had postings for more than 42,000 jobs in AI, with an industry average base salary of $129,739 for machine learning engineers. Depending on industry and location, salaries ranged into the $200,000s at larger organizations.
Integrating DIY AI with current technology
While many developers know what they want, some AI systems fail because they don’t know how to integrate their new tools with current systems. AI research scientist Prajit Datta writes that “top‑notch AI models will be of no benefit if the current workforce and the final users cannot efficiently work with the intended system.”
Datta adds that while creating the AI model in‑house gives organizations full control of the project, it also “requires significant effort in terms of organizational and administrative overheads. Considerable economic costs must also be taken care of in such an endeavor.”
Worldpay from FIS, the world’s largest acquirer, was looking for a new anti-money laundering (AML) solution. Ian Belsham, Worldpay’s Global Head of Transaction Monitoring at the time, said his in-house data team had tried to build their own AI system, but it was ineffective.
International consulting firm Milliman Inc.’s Payment Integrity Services saves its healthcare clients millions of dollars annually, but Milliman wanted to introduce AI technology to boost its success. They too decided to implement a proven AI solution by collaborating with Mastercard rather than continuing with efforts to build an in-house system.
What to consider when choosing DIY solution
Businesses dealing with thousands of transactions per second need AI tools that are fast, operate in real time, resilient and integrate well. Be sure to consider the following questions:
- What data is the model trained on and how much?
- What AI tools are used and how do they work together?
- Will the model integrate with existing technology?
- Does the right team exist on staff? How often will they need training and at what cost?
Off-the-shelf, market-ready models
Market leaders have developed off-the-shelf-solutions that are production-ready for immediate global deployment.
Speed to market while using global intelligence for greater accuracy and improved results
Technology innovators are finding ways to use the vast amounts of data generated by daily transactions. Many data consortia are available, and many AI and ML solutions access this targeted data in real time. It’s used to protect digital interactions in retail, banking and payments.
Business leaders with ample market share are well-versed in extracting high-quality intelligence from this data while securing privacy and personal information. But how to use it effectively?
In the AI space, some innovators are beginning to use this intelligence to develop solutions to solve immediate business challenges. The outcome is off-the-shelf AI that is fast and easy to deploy, improves results, is highly scalable and is user-friendly.
Proven off-the-shelf solutions are delivering ROI right away.
Making AI as easy as Excel
When Sudhir Jha became Mastercard SVP and Head of Brighterion, he predicted “AI will become as easy as using Excel,” and he has made that happen. This year Brighterion launched two of many production-ready solutions that are enriched with Mastercard global network intelligence. Out of the box, these models have built-in global experience and are improving results above and beyond existing solutions.
Brighterion enriches its off-the-shelf models with Mastercard’s global network intelligence, bringing worldwide experience to each deployment. Integration requires minimal data requirements, or, in some cases, none at all. The full stack, state-of-the-art machine learning toolkit enables unrivaled deployment and scalability while delivering response times as little as 10ms on-premises and 100ms in the cloud.
A large global acquirer using Brighterion’s AI for Transaction Fraud Monitoring increased fraud detection by two to three times while increasing approvals by 7.4 percent.
What to consider when choosing a market-ready solution
As innovators begin to develop and launch market-ready AI there are a few things to consider. Not all market-ready models are alike. To ensure the best investment for a business, organizations should ask:
- How long does it take to see results?
- Is the solution ready for immediate global deployment?
- Will the solution integrate with existing technology and business systems?
- Is it available to integrate through the cloud as well as on-premise?
- What is the expected improvement over current results?
- How many historical data elements are required and how long does it take to integrate them?
Customized solutions – outside expertise and personalized models
There will always be times when custom AI builds are needed to solve unique or specific business challenges.
Domain expertise is important for most jobs, and AI development isn’t just any job. A successful third-party developer will bring data science, development and subject matter expertise to the project team. They will determine the correct combination of tools for their customers’ specific needs, and advise on the necessary historical data.
To do this effectively, Brighterion follows a six-step process called AI Express. The team uses a proprietary AI and ML toolkit that includes automated features engineering, model generation and ensemble tools to build optimal bespoke models. Custom models are production-ready in six to eight weeks.
Some organizations, such as large financial institutions, use AI Express to build a model and test it against existing benchmarks and to determine what incremental lifts are possible.
AI for financial institutions
When Worldpay approached Brighterion, it was with the knowledge that Brighterion had expertise in both the financial industry and AI implementation. Worldpay achieved what Belsham called “astronomical results.” They were soon processing an additional 30 percent of transactions in a shorter period of time while receiving 20 times fewer false positives and 25 times fewer daily alerts. Significantly, fraud detection increased by three times.
Immediate ROI in healthcare fraud detection
Milliman approached Mastercard with two challenges. They wanted help identifying growing healthcare provider fraud and integrating the new technology with the Milliman Payment Integrity tool.
Using historical data from a regional health plan, Mastercard’s AI team built an “ensemble” model, a collection of sub-models that work together to compose the ultimate AI model. The new model uncovered $235+ million in potential savings for fraudulent claims, identified 2,700 high-risk providers and increased claim-level detection by three times.
What to consider when planning a custom AI project
Working with an experienced team at the forefront of AI technology in a specific industry will mean using tried and trusted methods. Organizations should expect the following:
- Clear goals for the project
- Proof of value
- Ability to compare to existing technology
- Deployment roadmap
- Story or case study from your industry
- Ability to experiment with layers of explainability to ensure the model complies with industry regulations
- Access to all AI technologies, not just a select few
- Ability to test the model and remove bias
- Model that deploys at scale when implementing the AI and as the business grows
- Speed to market
Finding the right AI implementation strategy
How organizations approach an AI implementation strategy depends on the size of their organization, their timeline and the problem it needs to solve or manage.
There are “less expensive” means to develop AI models, such as using open-source or other DIY solutions, but these often end up being unreliable, easily outdated and the source of hidden expenses. Organizations are left with incomplete solutions, cost overruns and a lack of on-staff expertise.
Market-ready models are quickly becoming just as effective as custom models, depending on the use case. Brighterion’s Transaction Fraud Management and Merchant Monitoring are two examples of off-the-shelf AI that address commonplace challenges faced by financial institutions. Bankers and merchants can see instant results in AI models that are enriched with global network data.
For custom solutions, Brighterion’s evidence-based, six-step process builds personalized models for unique challenges. Using a broad range of AI tools, the team delivers a personalized working model in less than two months.