Robot overlords have their limits. And that’s good
Talking about artificial intelligence is in season for Europe’s corporate executives. Just don’t mention its shortcomings.
The C-suite is eager to tout its abilities in riding the 21st-century wave of automation by using sophisticated machine learning or shop-floor robots. Mentions of the phrase “artificial intelligence” on earnings calls are surging, as Bloomberg Intelligence’s Michael McDonough has noted.
In a world where CEOs get more credit for cutting costs and buying back shares than opening factories or hiring staff, technology-driven efficiency is a carrot to dangle in front of shareholders. Stock-market valuations are stretched and spending opportunities are rare — but processing power is abundant and data storage cheap.
That’s why executives are conjuring up the promise of lower costs, more revenue or something in between. Deutsche Telekom and Royal Bank of Scotland are turning to chatbots — a digital replacement for call centres that could shave billions off costs in the next five years. France’s BNP Paribas and publisher Wolters Kluwer are trying to boost revenue, and are using machines to screen financial markets or customer databases and trigger automatic alerts.
Siemens computers are having a go at running gas turbines more efficiently than humans. And don’t forget the blue-collar world: Logistics firms Deutsche Post and DHL are talking up the idea of using robots alongside workers on the warehouse floor.
But there’s remarkably little talk of the limits of automation. What is the acceptable failure rate of these projects? Outside of games like Go or poker, just how suited are machines to the corporate world? Are some algorithms too expensive, as Netflix once found out? There’s a risk that disappointing results lead to an exaggerated corporate pullback, as the Harvard Business Review warned in April.
Machines can fail. Chatbots do so very publicly: Microsoft shut down a bot called Tay after pranksters pushed it to make racist, sexist and pornographic remarks. Earlier this year, Facebook went back to the drawing board after its bots hit a failure rate of 70 per cent, according to The Information.
Failure is fine, but the acceptable failure rate of an intelligent vehicle or a computer-controlled turbine is probably different to a bum steer
Failure is fine, but the acceptable failure rate of an intelligent vehicle or a computer-controlled turbine is probably different to a bum steer on an electricity bill
on an electricity bill. That can be the difference between an easy path to cost savings and a complex, long-term investment that doesn’t work as intended.
Then there’s the question of whether machines are always suitable. Machine learning works best in an environment with rules and huge numbers of data points. That might work with cars driving through heavy traffic governed by laws, or with achieving the best price for selling a big block of shares.
It might not work well in deciding where to invest a hedge fund’s money, for example, or recommending products to customers without much previous data to go on. The minute things get fuzzy — either due to a lack of rules, an unclear evaluation of success or a lack of data — artificial intelligence performs poorly, according to Pictet strategist Edgar van Tuyll.
These limitations mean it’s not yet clear that the cost of automation will be offset by savings in human capital. Hiring a data scientist can cost more than $200,000, according to Bloomberg News. Flightbookings company Amadeus has 40 of them. Siemens says it has more than 200 AI specialists running various projects.
And even Silicon Valley has its grunt workers: Facebook is hiring 3,000 content moderators, on top of 4,500 existing ones. AI cheerleader Amazon has 341,000 employees — three times the number it had in 2012.
There are good reasons to talk about AI and boast of its successes. But opening up about failure will help, too. — Bloomberg