You’ve identified and qualified the right use cases. You have the infrastructure, tools, and people ready to execute on the build and deploy phases of your artificial intelligence (AI) project. You’re in great shape. But what now?

We know navigating the AI space isn’t easy, but we have a recipe for success: Data + Change Management + Trust. These are the three critical ingredients you need to make AI work for your enterprise.

In this article — part one of a two-part series — we’ll assess if your data is ready for AI.

The opportunity for narrow AI

You’ve heard the warnings: AI will have, “... a more profound impact than fire or electricity” (Alphabet CEO Sundar Pichai) and “... could spell the end of the human race” (Stephen Hawking).

Maybe you tend to apply a dose of skepticism to such statements about “Artificial General Intelligence” or “Strong AI,” where a range of complex human attributes such as sentience, self-awareness, and consciousness come into play, not to mention ethics.

For our purposes, let’s reset the scope to narrow AI — the discipline of making machines good at one narrow task. 

We’re optimistic about (narrow) AI and the potential for us to leverage technology to enhance how we work ... augmented intelligence if you like.

We believe AI, and in particular machine learning, can fulfill and amplify the benefits analytics have long promised to businesses. Exceptional user experiences will increase productivity, improve customer satisfaction, and reduce market unpredictability. 

Industry analysts are also optimistic. McKinsey cites AI’s potential to impact over 400 business use cases — and this is just the tip of the iceberg.

Bringing AI to life can seem daunting, but it doesn’t have to be. To focus on awareness, mental cycles, and effort on the task, start with this simple framework:

Data + Change Management + Trust = Successful Enterprise AI

Up first is the topic of data.

Three must-ask questions to ready your data for enterprise AI

Except for new and experimental algorithms (e.g., low/zero-shot learning) and online learning, AI today is dominated by machine learning techniques that rely on historical data.

Data is, therefore, critical to success in deploying AI. And there’s no shortage of data, as we’ve spent the last few decades digitizing and instrumenting pretty much everything.

To ensure your data is ready for AI, focus on these three questions.

1. Is your data attainable?

Let’s look at attainability first. The following are questions to pose as you assess the best use case(s) for your data:

1. Is your data available? Do you know where the data is?

2. Is your data accessible? You likely know which systems contain your data, but can you access it both for model training and scoring?

3. Is your data aligned? Do your data assets line up with the use cases you want to solve?

When you can answer “yes” to all these questions, it’s time to analyze whether your (fully attainable) data has predictive legs.

2. Does your data have predictive power?

Next, you need to determine the outcome variable. What exactly are you trying to predict? Quite often, this is something the business is highly interested in but is not necessarily explicitly tracked in existing systems of record. Think about the propensity for a customer to repeat buy within a given time, or for a discounted quote to be approved at the first attempt.

Put another way, do you have labeled data? We find it’s necessary to leverage business rules to derive the labels from existing data.

Once you’ve set an outcome variable, then you can confirm or refute your hypothesis that the data has a predictive signal with respect to the outcome.

Previously, this could be a very time-consuming, manual, and repetitive task involving lots of queries, summary statistics, and analysis of data shape and correlation. Thankfully, we can now throw technology at the problem and have AI-augmented analytics sift through the data to pinpoint variables, quantify data quality, find more subtle patterns such as interactions between variables, and expose common machine learning pitfalls like hindsight bias and collinearity in raw data.

This analytical step is crucial and something of a litmus test. If you pass, you can be confident you’re off to a great start.

3. Is your data suitable for operations?

The third question relates to how suitable your data is when deploying a predictive model in the wild. In my experience, the most important consideration here is to ensure you can access your source data at scoring time. It’s not much use to set up a super-accurate model that relies on data from sparse, static, remote or unreliable sources which are either difficult or expensive to integrate and transform at runtime!

Less is more. Start with fewer data assets and expand as needed.

It’s also important to consider here whether your data is equitable. For some use cases, the potential for negative future implications is huge because AI-powered systems can scale and amplify egregious patterns in historical data.

Take, for example, financial services firms migrating from rules-based systems to algorithm-based, automated decision-making in areas such as loan approval. These models carry an increased risk of inadvertent discrimination as pre-existing bias in training datasets is reflected in and perpetuated by AI models.

(For more on the complex and multi-faceted topic of bias and how to prevent it, check out this blog series.)

Finally, for many use cases, a prediction needs to drive an action to improve outcomes. If the machine predicts a customer is unlikely to renew her contract, what should I do about it? Look for actionable variables in your data to guide the best course of action and predict what results actions may produce.

Once you’ve used these questions to confirm that your data is ready for AI, you’re ready to tackle the other two ingredients for successful AI: change management and building trust. Get started by exploring use cases for your AI journey. Download “Einstein’s Guide to AI Use Cases” to help your team brainstorm.