AI’s calm before the storm

Five years ago, artificial intelligence (AI) experts would have guessed that the first AI “unicorn” startup (a venture valued at more than $1 billion) in Canada would be from Toronto, Montreal, or Edmonton. That’s where the pioneers of AI research were. That’s where Big Tech (Google, Facebook and Microsoft) did their AI research.

But the experts would have been wrong. And yes, we count ourselves among those experts. We wrote a bestselling book on the economics of AI. We have seen hundreds of AI startups through our program for science-based startups called the Creative Destruction Lab. We thought the first AI unicorn would be where the cutting-edge research took place. Better technology would be all it took.

Instead, Canada’s first unicorn AI startup hailed from St John’s, Newfoundland, far from where Canada’s research leaders were developing cutting-edge AI.

Verafin gave financial institutions the tools to weed out money laundering and fraud. In 2020, NASDAQ (the giant U.S. tech stock exchange) bought it for $2.75 billion (U.S.).

What did we all get so wrong? We were focused on the location of AI expertise rather than where AI could be most easily implemented. Verafin was already in the prediction business, and today’s AI is nothing more than a prediction machine. Verafin already parsed financial transactions to find fraud. You need great predictive ability to find small needles in large haystacks. When AI’s latest wave of progress arose 10 years ago, Verafin took those inventions and improved its product with AI at the centre. Its banking customers wanted prediction; AI supplied it.

Those of us focused on the technology got enamoured with tools that carry on a conversation, produce art masterpieces in seconds, or beat humans at their own games. Compared with that, Verafin was pedestrian and, dare we say it, boring.

Chastened by our own predictive failure, we started to look closely at where AI was really being used — not by the tech stars, but by businesses. In the few cases where businesses were successfully adopting AI, they were mostly replacing other, more primitive forms of predictive analytics.

This pointed to a puzzle. Why was the uptake by business so low, given AI’s significant and ongoing technical achievements? Only about 11 per cent of businesses use AI somewhere in their operations. But most large businesses had spent meaningful investment dollars attempting to use AI. Why were they coming up short?

Verafin gave us a clue. For them, adopting AI was easy. They were supplying predictions to their customers, who knew what to do with them. But in many other businesses, AI could supply predictions — say, about demand or supply-chain issues — but many businesses were not equipped to use it.

When you are faced with uncertainty that you can neither predict nor control, you take other measures. If you are a business subject to supply-chain uncertainty, you develop a warehousing and inventory system to deal with the unexpected. If you struggle to provide services when there are surges in demand you can’t predict, you keep extra staff on just in case. You optimize your company to deal with the things you can’t predict. So, when a new tool comes along that allows you to make much better predictions, you must dismantle the system you are using and design a new one. That’s neither easy nor cheap.

When Air Canada developed an AI system to forecast freight demand and dramatically reduce the likelihood of empty cargo holds, it found that to put it in place, it had to train workers to pack the planes differently. In other words, it couldn’t simply adopt the enhanced prediction capability; it had to adapt the system in which it was embedded.

Businesses can cope with a oneoff retraining like in the Air Canada case. But some cases require a fullscale overhaul of organizational rules and operating procedures. Others require the ramping up of some divisions and the elimination of others. This will take time — even just to figure out whether the benefit justifies the cost.

This is par for the course for radical technologies. AI is coming. It will transform many industries. To take the current, modest uptake as a forecast that dramatic disruptive change isn’t coming would be a big mistake.