How to Train Your AI Dragon

The good news for organizations seeking to leverage data to create value is that they currently have access to more information – about their products, their employees, their customers, and their marketplace – than they’ve ever had before.

The bad news? That data is often a mess.

According to a recent survey, employees spend nearly an hour per day hunting down information. While 60% of business leaders see the potential of artificial intelligence (AI) to automate mundane tasks within their organization, 60% also say their organizations lack the skills to capitalize on AI.

Fortunately, AI tools are becoming more and more user-friendly, giving organizations more power to turn data into actionable insights – even if they don’t employ a raft of data scientists. An example of this is low-code point-and-click user interfaces that walk users through the step-by-step process of training and deploying custom machine-learning models (ML). As a result, even organizations with very nascent, in-house AI expertise can attain many of the same outcomes as more mature organizations.

Success in AI is about more than technical talent or even domain knowledge. The secret lies in having the right data and content to train models and developing a plan to go after a specific outcome.

Organizations should follow these three steps as they seek to create value through AI:

  • Define ML Models – As with any business process, an AI initiative should begin with the end in mind. Business and IT leaders should first think through exactly what they hope to achieve via AI processes, and then define their ML models in a way that will leverage existing content to meet those goals. To take a relatively simple example: A global apparel company might want to be able to quickly identify and locate photographs that feature a specific model, article of clothing, or accessory. This sort of AI outcome can save a tremendous amount of time for organizations that have thousands of photos with no metadata tags; in fact, this use case can also lead to significant cost savings. Without an AI tool, an organization may need to put an army of data-entry resources to work on the problem or even go through the expense of reshooting images that they already have, but simply can’t find. The very first step for business and IT leaders seeking to leverage AI is to say: “This is what we want to predict; this is what we want to get back from our ML model.”
  • Model Training – The next step organizations must take is to actually train their ML models. This involves providing the tools with a training set of information that includes both assets and data that are aligned to desired outcomes. Even the most powerful AI tool can only complete the tasks that it has been trained to do; unlike a human worker, an ML model can’t “guess” or fill in “gaps” if it lacks the necessary information or context. So it’s easy to see why model training is such a crucial step in attaining business value from AI tools.
  • Operate Your Models – Over time, ML models will continue to evolve, and organizations must keep working to refine them by checking for accuracy and making corrections where needed. You can’t simply release an ML model into the wild and assume that it will always return the right answers. It’s possible for a model to begin to show bias, or to become corrupted, and human stakeholders need to constantly check the model’s results against desired outcomes. There are four important things you should be able to do:
    • Quickly promote new AI models into production.
    • Continuously monitor different models over time to determine if performance is improving or degrading.
    • Compare and analyze the performance of different models – or even different versions of the same model – to determine which is delivering optimal performance.
    • In the case of a corrupt model, quickly assess and deploy a previous, non-corrupt version of the model.

Governance is a critically important piece of any AI initiative – and that too often gets overlooked. For instance, model versioning, the ability to track which predictions are machine-generated and which are human-generated, and the ability to roll back to previous model versions are all essential. Finally, organizations must have the ability to store training data sets in case an auditor or regulator wanted to understand how the model was initially trained.

In short, defining an ML model, training the model, and then operating and governing that model are the three steps to training, or “taming,” the AI dragon.

Christopher McLaughlin is Chief Product and Marketing Officer at Nuxeo where he leads the long-term product and go-to-market strategy for the company’s modern Content Services Platform (CSP). He has over 20 years of experience in the enterprise content management (ECM) space across multiple industries, including financial services, insurance, and product marketing and manufacturing. Christopher specializes in advanced technologies such as AI and machine learning, with a focus on automating workflows and other digital business processes.