March 23, 2022

AI debt: solve it before it sinks your business

Mike Khanna

Startup CEOs will be used to the idea of technical debt—the cost of rewriting code that was created in a hurry to solve the immediate problems early on. But now there’s a new problem on their radar: AI debt. Almost every company knows that AI is the future. According to a recent McKinsey report, 56% of companies now use AI for at least one business function. Most commonly, they use AI to optimize service operations or to deliver new AI-enhanced product features. But while companies are rapidly turning to AI, they are facing a growing burden of AI debt. Here, I look at why AI debt is such a challenge and explain how Sonasoft can help you overcome it.

What is AI debt?

Before you can solve it, you need to understand where AI debt comes from. As you will see, there are several sources of AI debt.

Failure to deliver savings

Typically, companies turn to AI because it promises significant cost savings through improvements in efficiency, reduction in waste, etc. At its simplest, AI debt is incurred by failing to make those cost savings, leaving your business vulnerable in times of economic turmoil. Often, companies will sell you AI solutions by claiming they can make huge cost savings: “halve your costs overnight”, “boost your team’s productivity by 100%”, etc. In fact, successful AI projects often only deliver improvements of 10-20%. But with many companies looking at profit margins of 1-2%, that can easily lead to a 10-fold boost in profitability. 

Loss of opportunities

AI is a truly disruptive technology. Failure to adopt it will leave your company lagging far behind your competitors. So another definition for AI debt is the lost revenue opportunity from failing to adopt AI when your competitors have done so. Put simply, AI will allow them to undercut you at every stage. Amazon became the world’s richest retailer because of the power of AI. That allows them to deliver true dynamic pricing and intelligent recommendations, driving up their profits and edging competitors out of the market. Likewise, Google became so dominant because it leverages AI in every part of its business. If you don’t adopt AI, you will soon find yourself lagging far behind your rivals.

Failure of AI projects

Unfortunately, solving AI debt isn’t as simple as deciding to adopt AI. AI debt also hits companies that are fully committed to AI. That’s because 87% of AI projects never make it to production. Just think about that in context. According to Verified Market Research®, the AI market was worth $51Bn in 2020. That means companies potentially wasted $45Bn in projects that failed. So, the true definition of AI debt needs to account for that cost too. In fact, this may be the biggest element of AI debt. Solve this, and you can quickly eliminate most of your AI debt. 

AI debt equation

Why AI projects fail

So, why do so many AI projects fail to deliver? I asked my Chief Data Officer, Dr Caroline Zaborowski. She tells me there are three fundamental problems that need to be solved.

Failures in data discovery

Data scientists like me have an adage: “garbage in equals garbage out.” In other words, your AI model can only ever be as good as the data you use to train it and drive it. Often, companies come unstuck because they want the data to deliver more than it can. Put simply, if you only have insufficient data, you can never deliver a successful AI project. Yet we still come across companies trying to sell you the dream that they can somehow synthesize data to overcome this. Or telling you that unsupervised learning will allow them to deliver a perfect AI model without the need for preliminary data discovery. Sadly, this sort of capability remains in the realm of science fiction. In practice, you need to invest real effort in finding and converting all your data into a usable form and checking that it supports your hypothesis.

Lack of data science 

Another common problem we come across is poor (or completely missing) data science. Data is king when it comes to AI. But raw data has to be carefully prepared before you can use it. Failure to do this can lead to models that output rubbish results. So, you need to invest time into cleaning the data, doing feature extraction, and validating your assumptions. All these processes are vital but if you get them wrong they introduce unintended bias. For instance, if you have already assumed a given outcome it is very easy to cherry pick the data or features that support that. 

Failing to plan for production

Possibly the most damaging error made by many companies is failing to understand that AI only works if you think about it as an end-to-end problem. You can create an absolutely perfect AI model but it is literally worthless if you can’t embed it into your business systems. Specifically, that means thinking about how you will run the model in production. How you will feed it with the data it needs. And importantly, how you will actually use the model’s outputs in practice. Time and again I see examples where companies have put together a top-notch AI team who produce great models but are unable to deploy them. 

How Sonasoft cuts your AI debt

At Sonasoft, our focus has been on creating a system that makes it easy for companies to adopt AI successfully. Our engineering team has spent the past year completely redesigning our AI platform from the ground up. The result is SAIBRE, a platform that makes building, running, and maintaining AI models simple and intuitive. On top of this we have a three stage process to ensure your project won’t be one of the 87% that fail.

Data-driven discovery. Sonasoft has a rigorous approach to data-driven discovery in any new AI project. The client shares all their data and their current high-level business intelligence. We ask them what their theories are and their expectations for what the AI model can do for them. Then my data science team analyzes the data in an in-depth and unbiased fashion to validate whether it supports their ideas. If the data is lacking, we will have an honest conversation with the client about the potential pitfalls and explain how they can improve their data. If the data supports the suppositions, we will move on to the next stage.

Data engineering. My team are experts at data exploration and feature engineering. We will ensure that the data is analysed and cleaned up in an iterative and consultative manner. We check everything with the client and make sure that we understand the implications of any operations we do on the data. The result is better data quality, less room for bias, and a more robust model.

Built for production. The best AI model in the world is useless if it can’t be deployed. SAIBRE was specifically designed to make it easy to deploy AI applications in production. Indeed, its name captures this perfectly—the Sonasoft AI Build & Runtime Engine. Deploying a model can be done in just a couple of clicks. And then SAIBRE takes care of running and monitoring the resulting application. All this is backed up by a new UI that allows both non-experts and experts to make the best possible use of the system.

Want to slash your AI debt? Then let us build a robust AI application that can be deployed into your production environment in just a couple of months. To find out more, talk to our sales team.

White Paper

SAIBRE AI Ecosystem

End-to-end AI applications that solve any business problem