Erik Brynjolfsson, MIT professor and director of the MIT Initiative on the Digital EconomyPhoto: MIT
Chief information officers said 2018 could be the year they deploy machine learning across their companies to automate repetitive tasks and augment human workers. But getting there won’t be easy. Many firms are still figuring out what tasks to automate, and are exploring ways to organize their teams to support more machine learning.
“It’s not because the technology is lagging. It really has to do with the organizational side, the culture, and the co-invention of business processes that takes a lot longer,” said Erik Brynjolfsson, MIT professor and director of the MIT Initiative on the Digital Economy.
Mr. Brynjolfsson, author and co-author of several books on the economic implications of AI, spoke recently with CIO Journal about machine learning’s role in the enterprise. The conversation touched upon a December article in the journal Science, by Mr. Brynjolfsson and co-author Tom Mitchell, which outlines how companies can determine which tasks should take advantage of machine learning.
Edited excerpts follow:
What’s the current state of machine learning in the enterprise?
We’re very far from artificial general intelligence. Even though AI is amazing, it can only do certain narrow tasks well, and many tasks it doesn’t do well at all. If I were a CIO, I would … go through all the tasks we’re working on in our company, and see which ones are suitable for machine learning. Or maybe look at the ones that people have asked us to do, see which ones are suitable and use (the rubric) to prioritize where we do our work.
What’s keeping AI from ramping up more quickly?
It’s not because the technology is lagging. It really has to do with the organizational side, the culture, and the co-invention of business processes that takes a lot longer. CIOs know this. Say you’re trying to implement an ERP system. You don’t just flip the switch a couple of weeks after you buy the software. It takes years. It’s the same thing with a lot of AI applications. You have to reinvent your business processes, and that process redesign is really where the time-consuming work happens.
It’s not just buying some AI and popping it in. It’s rethinking your processes and reorganizing your business so that you decide who has responsibilities.
Does expectations for human workers change expectations for how productive a human should be at work?
If you go through different jobs or occupations, it’s very rare that the whole job could be done by machine learning or artificial intelligence. It’s also rare that none of it could be done. By far the most common outcome is that if a job has 20 to 25 tasks … and you look at each of those tasks separately, some set of them can be done by machine learning. That means that what you need to do is have both humans and machines working together. Often you need to do some reengineering to reorganize things so you carve out the parts the machine can do and send them over to the machine. Maybe you send them to the cloud, or they’re done by another company or another part of the same company. The other parts that only humans can do the humans focus on. But it’s a sort of reassignment of who does what.
What might that look like in practice?
Suppose you are in a health-care company and you have radiologists looking at medical images. Machines are now able to do that extremely well to detect cancer. At the same time, the humans do a lot of other tasks. They talk to the other doctors, they look at other lab reports, they interface and communicate with patients, and recommend treatments. So all those other parts of the job still need to be done and they aren’t done by the machine learning, so you need to have a way that the machine learning part can communicate its recommendation to the human doctor, and the human doctor pulls in all the other information and communicates with the patient. It’s a re-engineering of that role.