5 Steps You Must Follow to Make AI Efficient

Artificial intelligence (AI) is here to stay. With almost every technology company saying, "We use AI" and "We can make you X percent more effective," businesses are not only interested in how to use AI but are under internal pressure to make it work.

So why are businesses so often underwhelmed by the results of AI?

The biggest challenges of AI

AI comes with three main challenges.

1. Adoption. Getting people to move away from the traditional methodologies they know and have long relied on is quite difficult.

2. Untrustworthy. Many current AI methodologies are 'black box,' meaning users can't see how or why the outcomes are delivered, and consequently don't understand or trust the results.

3. Not fit for purpose. Automated systems can deliver false positives when poorly tuned and applied, which can decrease efficiency rather than improve it.

At the same time, many people who bring in these technologies often don't realize that even AI needs a helping hand. Simply throwing compute capacity at the problem is akin to the monkeys and typewriters theorem.

How you can use AI well

Given these challenges, many businesses are left wondering what they can do to help their AI truly meet the promise of 100% effectiveness.

While there's no magical silver bullet, to get the best results, start at the foundation – the data. If you want results that match the patterns in the real world, and that help you solve real-world problems, your starting point has to be the underlying data that represents the real world. This is something called network analytics.

It sounds obvious, but as anyone who's worked with data or AI will know, it's not simple. However, this five-point recipe can start you on the journey to getting real-world results from your AI, and move you toward becoming a data-driven organization.

1) Resolve your entities: Match entities, create a single customer view

Data about individuals, organizations, or real-world concepts should be joined together in a process called Entity Resolution. An example of this is creating a statistical single view of the customer. However, match rates are often as poor as 50% (on a good day!) due to data quality issues, or are confused by trying to match complex entities such as businesses. Fortunately, some techniques can raise resolution rates to 85%+ by using multiple data points to overcome data quality challenges.

2) Create networks: Link the networks together, uncover the relationships

Entities interact in the real world and form networks: businesses have directors, customers, suppliers, and employees. People have networks too – those they live with, places they visit together, or others they interact with.

However, to consume these networks using AI, you need to generate statistics about the networks, implying that networks do not go on forever. This in turn means they have to be"bounded," with the question then becoming where to make the edges of a network. Furthermore, you have to deal with over-connectivity (e.g. many people pay utility bills, but are they all connected? Probably not). But once you have good-quality networks in your data, you'll never look back.

3) Apply network analytics: Understand your network, analyze the relationships

Now you are ready to develop your AI models, whether for detecting financial crime, assessing risk, or identifying contagious churners. You'll find that network analytics not only predicts 100% more of what you are looking for, but false positive rates plummet. The outcome can be up to a threefold increase in effectiveness by only focusing on those alerts of interest.

4) Operationalize analytics: Bring AI into the process to produce compelling outcomes

Often crucially missed is the impact that can be achieved by using network analytics to operationalize AI to produce compelling outcomes. Imagine the traditional AI model detecting insurance claims fraud: "This claim has 67% risk of fraud" (based purely on the claim data alone) – where does the investigator start? How long will it take them to investigate? Will they remember when the system was right or each time it was wrong?

False positives can be the scourge of AI solution adoption. Now imagine the network analytics equivalent: "This claim may be fraudulent because...

1. The claimant's partner was previously insured and made a similar claim;

2. The credit card that bought this policy, bought two other policies both having made similar claims in the last 3 months;

3. The ratio of whiplashes to vehicle damage claims across this network is 3:1;

4. The network has a 150% growth in claims over the previous month;

5. The vehicle was involved with another vehicle with claims that were rejected for fraud."

An investigator can make a much faster decision when presented with an alert that has been elaborated using network analytics. This dramatically improves the throughput of the investigation team, often cutting effort by 70%. A similar approach can be used for AML investigations or tax compliance issues, as well as assessing the next product to cross-sell to a business.

5) Corporate memory: Blend humans with AI, enhance effectiveness

Blending human decision-making with AI is often underlooked and is a missed opportunity, as it can significantly enhance future effectiveness. Every time a human makes a decision, that outcome can be stored against the appropriate entity within the network. As all these nuggets of insight are on a network, their impact is amplified as they influence the scoring of other surrounding entities on the network. For example, the average over-indebtedness or payments in arrears on a network could be a useful variable to use in your models.

If you want to up your game on data-driven decisions, the next-generation approach is to use network analytics. The benefits are significant, and once you have entities and networks in your data lake, your data scientists will wonder how they ever managed to deliver outcomes without them. You should aim to make this a fundamental utility within the data lake that can operate both in batch and real-time.

Technology is now available to accelerate your journey toward network analytics without having to reinvent the wheel. Your data scientists can deliver game-changing results across your organization using AI, by combining it with network analytics.

The State of AI in Financial Services