AI/machine learning

Four Points Can Improve Your Odds of AI Success

ODDS—organization, due diligence, data, and scrub. These four important steps can make sure you are ready to implement artificial intelligence in a way that leads to a successful project.


At a recent discussion session at The Data Standard, we asked the question, “Are you ready for AI?” Josh Odmark and I hosted the session and introduced the topic. Here are the highlights of how to get ready for AI and improve your ODDS of the project giving you a great result.

O. First, the organizational framework must be set up to create a supporting structure. This starts with support and buy-in from top-level management who are clear on the objectives. A lack of management support will usually mean that the project will either end when the going gets tough or not get deployed in the end. Furthermore, a project needs to be set up with all the usual bells and whistles of a project: a timeline; to-do items; stakeholders; budget; and, most importantly, project management with enough spare time to see it through.

D. Second, the project should start with due diligence. In this context, that refers to making sure that the current situation is known, the problem is clearly defined, and the form of the solution is clear. When using the word “clear” twice in the prior sentence, I’m speaking as a mathematician who wants a numerical definition that can be objectified. Especially, the solution needs to be defined as precisely as possible, including objective numerical success criteria that can be assessed at any stage in the project and allow the project to stay on track. An essential part is the desired accuracy of the model. Projects are often either terminated too early (“this will do”) or too late (“let’s see if we can do even better”), wasting resources in either case.

D. Third, the data must be available. Data science, artificial intelligence, machine learning, or whatever you wish to call it relies on having as much data as possible. Some data may be in your possession. Other data could be generated, acquired, or bought. Some data cannot be obtained at all. What data is even needed, or desired, must be determined at the very start of the project. And this must be done by domain experts and not the data scientists! Sometimes you have less data than you really want, and the question is whether this is good enough or not. Unfortunately, this cannot be answered until you try to make the model, but all stakeholders should be aware that the project is taking a risk.

S. Fourth, the data must be scrubbed. Most of the time, the term is “clean,” but I have used “scrub” here for two reasons. Of course, I want the acronym of this checklist to work nicely but also because this process actually involves more than just cleaning. Before we clean the data, we must ask whether the data is relevant to the question we are asking and whether it is representative of the underlying problem. Often, these two difficult items are not done, which leads to models that do model the situation but do not solve the problem that was posed. Note that being relevant and representative are two technical words from statistics that can be made quite precise. Following this, the data must be cleaned, which means outliers and bad data are removed, missing values are filled in, spikes are possibly smoothed over, and so on.

Now you are ready to do AI.

Having done AI, you face the toughest task yet: convincing people to use your model in real life. Change management is the process of getting all the users to change their behavior from whatever they were doing to the new work flow that includes your model and its solution to the problem. This change may involve many features that have nothing to do with AI, or even nothing to do with the problem. Take, for example, industrial predictive maintenance. The problem is that equipment fails catastrophically sometimes, which is why companies want an accurate forecast. The change is that maintenance personnel, having acted in a fire-fighting capacity all along, must now learn to plan their work and spare parts orders. This is a significant change in the way the company operates and may require additional software tools to realize, all of which has nothing to do with making an accurate failure forecast.

Want some help with improving your ODDS? Feel free to reach out.

Patrick Bangert is vice president of artificial intelligence for Samsung. Before holding that position, he founded and was CEO of Algorithmica Technologies, a machine learning company specializing in oil and gas applications. He holds a PhD degree in applied mathematics from University College London. After a few research positions at Los Alamos National Laboratory and NASA’s Jet Propulsion Laboratory, Bangert became assistant professor of applied mathematics at Jacobs University Bremen in Germany. In 2005, he founded Algorithmica to bring the machine learning methods down from the ivory tower into practice. Bangert is cochair of SPE’s Digital Energy Technical Section.


Join the SPE Digital Energy Technical Section

If you are not yet a member of the Society for Petroleum Engineers, please join SPE. You will be a member of a vibrant community of more than 164,000 professionals in the oil and gas industry. SPE provides active value to its members through conferences, publications, webinars, and a huge industry network. Once you are an SPE member, you can join sections (geographically local groups) and technical sections (topical special interest groups). If you have an interest in all things digital, then please join the Digital Energy Technical Section (DETS), one of 16 SPE technical sections at present. You can join by going to your online membership profile and selecting “Professional Online Communities.” There you will have the opportunity to edit your selection. Below the technical communities, you will find the technical sections. Please select “Digital Energy” and be sure to click “Save” at the bottom. Welcome to DETS!