The two key changes that will unlock the true potential of AI

AI’s potential lies in machine learning – and a code of ethics

‍The stars are aligning for AI as machine learning technology matures, data gets richer and more insightful and processing power expands. But AI needs to be reliable to be truly powerful, and machine learning models need standardization before we can see real progress.

I don’t want to sound complacent, but when I hear people proclaiming that AI and the robots are taking over, I don’t immediately start making plans to take down Skynet.

For one thing, the technology, while developing quickly, is nowhere near reaching full maturity. More fundamentally, though, the ecosystem that surrounds AI – particularly in the business world – has fundamental issues that need to be overcome before we start to see AI reach its true, game-changing potential.

Two key developments, though, give me a huge amount of optimism about the extent of AI’s impact in the coming years. One of them might even see off the rise of the machines.

Functionality, fairness and faith

If AI and machine learning (ML) are going to make the kind of transformational difference that people like me are predicting, we’re going to have to trust their output. And, so far, that hasn’t always been easy. If you want to see a PR disaster unfolding in real time, you only have to look at the backlash when an image-cropping algorithm on Twitter was found to be biased toward white faces.

Or look at challenges AI has faced in reaching its potential to ease the burden clinicians face in overstretched health services like the NHS. Here, of course, it’s more than just reputations and revenues on the line, so you need to be completely confident that your recommendations are going to be at least as accurate as a human clinician. In some early AI trials that hasn’t been the case.

It’s hard to blame the organizations involved, here. For all the extremism you can find on its platform, Twitter didn’t suddenly become racist when it launched a new feature or piece of AI as part of its delivery model. It’s not even about the underlying technology. It’s much more about the need to ensure that the ML models being used are robustly configured in such a way that they’re effective, unbiased and safe. As it stands today, there’s no standardized way of doing that.

I’ve been burned myself here. A few years ago, I led a team that developed an AI app to help deploy pharmacists around a Northern Ireland hospital based on both the location and the acuity of the patients. As a proof of concept it was fantastic, but without a formal assurance model in place there proved to be no clear path from concept to scaled-up deployment, and certainly not a cost-effective one. It was very frustrating.

As mundane as it may sound, the lack of assurance standards – a code of ethics for the next generation of coding – is holding back the progress of AI. This lack of clarity around what “good” AI looks like makes CIOs and data scientists reticent when it comes to strategic innovation and true data-driven digital transformation. Without a clear sense of what AI should and shouldn’t be doing, we’re not yet finding out what it can do.

The good news is things are changing. Last December, the UK Government published a report from the Centre for Data Ethics and Innovation that laid out a path to an effective assurance ecosystem. A White Paper and ISO standards will follow, and before long we can expect industry-focused regulators to be working with businesses and their data scientists to give the formal stamp of approval to innovative new ML models that are fit for purpose, fair and safe.

This will remove a significant barrier to AI innovation. But on its own, it may not be enough.

Operationalizing machine learning

Many of the organizations that stand to benefit most from AI are unfortunately those least ready to do so. If you’re starting a business now, then you’ve no excuse not to be data-driven. But if you’re a business with a proud heritage that may stretch back hundreds of years, the chances are your data and your IT systems are going to be siloed and characterized by legacy systems and workarounds. In the financial services industry, for example, there’s a huge amount of technical debt.

This presents an immediate barrier to the kind of advanced analytics that can unlock deep insights into areas like risk and customer retention that are so crucial to banks and insurers. Even those organizations whose data frameworks are relatively mature, though, can find it a struggle to implement meaningful AI into their operations. And here the challenge can be as much cultural as technical.

Machine learning models are the domain of data scientists, specialists (like me) who (unlike me, admittedly) won’t have a background in enterprise IT. They’ll create these models based on business requirements using specialist tools and programming languages, and test them to satisfy themselves that they’re yielding meaningful, useful results. But then what?

You can’t expect a typical IT department to know how to support these specialist tools, let alone work out how to integrate the resulting predictive models into a regular workflow like an online customer journey. And you can’t expect a data scientist to become an expert systems integrator. But perhaps, instead, you could give her a platform that could fast-track the path from predictive model to production code.

At UST we’ve developed one of the world’s first enterprise-level managed services for MLOps. In developing our MLOps platform - xpresso.ai - we drew from the tried and tested software development lifecycle model enshrined in DevOps, but reimagined it for machine learning. Xpresso.ai is a platform that gives the data scientists all tools they need but – crucially – also provides a ready route for deployment. And, with tools for data and model versioning, it creates an environment where ML models can evolve as situations and data sources change over time.

Exactly how these ML tools get used will vary from situation to situation. We’ve already deployed tools that give executives configurable data visualizations of predictive models that they can work into presentations and negotiations, for example. Or it may be a case of an e-commerce team or digital process accessing the model via an API.

The point is that we’ve created a platform that gives both data scientists and those focused on digital transformation exactly what they need.

Looking ahead

And this really is transformational. Our early implementations of xpresso.ai the MLOps platform have been a real eye- opener. In one UK financial services business, the user-friendly deployment of predictive models has made heroes of the data scientists, who suddenly find themselves in huge demand. Not a familiar situation, perhaps, but a welcome and important one for the future of effective AI utilization. Most exciting of all, though, is the potential for these two themes to come together. The structured nature of MLOps means that as assurance models begin to roll out we can embed them as key decision points in the machine learning development lifecycle, creating an assurance audit trail that I believe will give new authority, confidence and freedom to the next generation of AI innovators.

There are exciting times ahead of us.

Read more about how UST is innovating with AI and machine learning  and discover SmartOps, an AI-powered cognitive automation platform.