This robot, Pepper, welcomes people to the store.
Artificial Intelligence (AI) and machine learning (ML) applications are growing and expanding into all industries and functions. From ranking sales leads (Einstein) to chatting with customers (Bold 360) artificial intelligence seems to be popping up everywhere. However, many of the of the boldest claims for artificial intelligence have not yet come to fruition. AI is not yet curing cancer and Amazon’s recruiting tool acted with just as much bias as a normal human recruiter. And therein lies a key takeaway for anyone wanting to find success with AI: artificial intelligence can’t yet do things that humans can’t do.
Sure, machines can be faster and make fewer errors than humans, but if you’re trying to train a machine to do something, you better make sure you could teach a human to do it first, even if the human would be painfully slow. You also need to make sure that that the machine can actually use all of the information that a human would need to solve the problem. If your task involves performing some analysis on numbers in a spreadsheet, you’re in good shape. Machines are good at math and the data is already in a convenient structure. If, instead, your task is to read a report and write a summary, well, you’re in trouble. Natural language processing (NLP) still has a long way to go before it read with the same level of comprehension as a first grader, much less a financial analyst.
You can see the evidence for these facts in the adoption patterns of AI. Robotic Process Automation (RPA) is one of the fastest growing applications of AI. Rather than solving new problems in new ways, RPA solves existing problems in the exact same ways as humans—essentially learning to copy the actions human take to interact with digital systems and automating the repetitive pieces with “virtual workers.” To date, most AI is copying what humans do in order to do it faster, more efficiently, and more accurately—whether scoring leads, creating dashboards, or detecting fraud.
This is good news and bad news for workers. The good news is that AI, through automation and RPA, can take over many of our most unpleasant and boring tasks. No one will miss populating data throughout five different internal systems or manually moving files for storage. The bad news is that there are jobs in the market that can be automated almost out of existence. Although the end result of automation on the workforce is something experts love to argue about, it’s pretty clear that jobs with low skilled and repetitive tasks will begin to evaporate.
For leaders who are looking to use artificial intelligence to improve output or lower costs, there is also good and bad news. The bad news is that artificial intelligence isn’t advanced enough to solve challenging, complex problems. Or even to offload much of research and development or investor presentations. The good news is that there are real cost savings available through artificial intelligence, they just often aren’t in the most exciting, glamorous, and news-worthy parts of the business.
Executives who want to take on AI projects—and that should be everyone because this is a competency that is becoming “table stakes” for staying competitive in every industry—need to keep some basic advice in mind.
- Start simple. Begin by automating simple, repetitive processes that eat up employee time and wear on employee nerves. This kind of automation doesn’t make headlines, but that improving your profit margin might. RPA is a good way to begin building baseline skills in AI.
- Focus on what humans do already. Once you can automate simple tasks, begin expanding into more intellectual exercises. Items like sales lead rating or inventory forecasting are good places to begin. Humans do these tasks today, usually leveraging solid data and a little bit of insight. See if machines can help or do it better.
- Remember the need for data. Humans are excellent processors of unstructured data. Conversations, reports, web pages, fuzzy memories, and more get incorporated into our decision making processes. If we want to offload tasks to machines, we need to give them access to all of the data that we use when making decisions. If the majority of this data is unstructured, then the tasks is not a good fit for AI.
At AIMatters, we got to experience all of these challenges and roadblocks ourselves as we tried to teach a machine learning system to identify companies’ business models based on financial reports. Although a smart analyst could read an annual report and come up with a solid answer, it took many iterations for the team to find the algorithms and the data required to teach a machine to do the same.
In twenty years, we will be having an entirely different conversation about AI—and hopefully reflecting on the many new insights and innovations this technology has contributed to the world. In the meantime, leaders need to take baby steps and focus on tasks that machines can succeed at—those that humans can do well, with accessible and structured data.
