How AI Will Trans­form Busi­ness

What does AI mean for busi­nesses big and small? What key op­por­tu­ni­ties and chal­lenges does it present? Two ex­perts on the topic weigh in: Rot­man School Dean Tiff Mack­lem and Sco­tia­bank CTO Michael Zerbs.

Rotman Management Magazine - - FRONT PAGE - By Tiff Mack­lem and Michael Zerbs (MBA ’89)

What does AI mean for busi­nesses big and small? What key op­por­tu­ni­ties and chal­lenges does it present? Two ex­perts on the topic weigh in.

We hear so much about ar­ti­fi­cial in­tel­li­gence (AI) these days, but many lead­ers are at dif­fer­ent lev­els in terms of un­der­stand­ing what it means for busi­ness. What do they need to know?

First of all, I want to be clear about some­thing: AI MICHAEL ZERBS: is al­ready here. You are us­ing AI when­ever you type a mes­sage on your iphone and a word gets auto-com­pleted; when­ever you type a word into a search en­gine and it mag­i­cally com­pletes it­self; and when­ever you use Google Trans­late. These are all AI ap­pli­ca­tions, and they share two key char­ac­ter­is­tics.

In each case, it’s about a ma­chine/agent per­ceiv­ing its en­vi­ron­ment. If it were a real, nat­u­ral in­tel­li­gence such as you per­ceiv­ing the en­vi­ron­ment, that would en­tail us­ing your eyes and ears; but be­cause it’s a ma­chine do­ing the per­ceiv­ing, it uses sen­sors and al­go­rithms. Once the en­vi­ron­ment is per­ceived, the agent takes an ac­tion that is ori­en­tated to­wards a dis­tinct goal. That goal could be ‘com­plet­ing the word that you had in mind’ when you started to type; or it could be ‘au­tonomously driv­ing a car from point A to point B’ in a rea­son­able amount of time, with­out caus­ing an ac­ci­dent. So, the key char­ac­ter­is­tics of AI are that a ma­chine per­ceives the en­vi­ron­ment and then takes ac­tions to op­ti­mize a par­tic­u­lar goal.

Lead­ers should also be aware of ma­chine learn­ing — a branch of AI that ‘cre­ates’ in­tel­li­gence by learn­ing from data. It’s a way to take a lot of data and ex­tract use­ful in­for­ma­tion to help us achieve a goal. Com­put­ers have be­come ex­tremely pow­er­ful: Things that were once the­o­ret­i­cally pos­si­ble were only pos­si­ble if you were will­ing to wait for a very long time; all of a sud­den, you can do these things very quickly.

Tiff, the Univer­sity of Toronto — and the Rot­man School in par­tic­u­lar — have done some in­ter­est­ing re­search around the eco­nomic con­se­quences of AI. Can you talk a bit about this?

What is so in­ter­est­ing about AI is that it’s not an in­TIFF MACK­LEM: ven­tion like, say, in­sulin. In­sulin was a hugely im­por­tant in­no­va­tion that has mas­sively im­proved the qual­ity of life for di­a­bet­ics — but it’s not some­thing that you can take and ap­ply to all sorts of things. AI is more of a ‘gen­eral pur­pose tech­nol­ogy’: It has widerang­ing ap­pli­ca­tions, and we are only just start­ing to see what those are.

The way econ­o­mists think about AI is, we ask, ‘What does it ac­tu­ally do?’ Like any dis­rup­tive tech­nol­ogy, it dra­mat­i­cally drops the cost of some­thing — and in our view, that some­thing is prediction. Take the anal­ogy of com­put­ers. What they dra­mat­i­cally re­duced the cost of was arith­metic — and as a re­sult, things that in­volved a lot of arith­metic were quickly au­to­mated. Next, peo­ple be­gan to re­al­ize that there were lots of prob­lems that we didn’t nec­es­sar­ily think of as arith­metic prob­lems that could be made into arith­metic prob­lems.

For ex­am­ple, we used to take pho­tos on ana­log cam­eras us­ing Ko­dak film, then take the film to Blacks to be de­vel­oped via a chem­i­cal process. One day, some­one said, ‘You know what? Given that arith­metic is so cheap now, maybe we could pro­duce pho­tos dig­i­tally’. All of a sud­den, we were tak­ing more pic­tures than ever be­fore — and Ko­dak and Blacks went bank­rupt.

As in the case of com­put­ers, with AI, the ini­tial ap­pli­ca­tions we are see­ing are very ob­vi­ous things. So, based on your pre­vi­ous pat­terns, Netflix uses AI to pre­dict which movies you might like to watch, and Amazon uses it to pre­dict which books you might want to buy. These ap­pli­ca­tions are handy — but hardly trans­for­ma­tional.

How­ever, we are start­ing to see more mean­ing­ful ap­pli­ca­tions. For ex­am­ple, in health­care, AI ap­pli­ca­tions are dra­mat­i­cally drop­ping the cost of di­ag­no­sis. Say you no­tice a new mole on your arm: Is it just a sun spot, or is it a melanoma? You can now take a pic­ture with your phone and an AI al­go­rithm can tell you. That is go­ing to lead to bet­ter health out­comes. The way to think about AI is, it’s a very pow­er­ful prediction en­gine — and the uses we’ve seen to date are just the tip of the ice­berg.

Tiff, within five years, the Rot­man School’s Creative De­struc­tion Lab (CDL) has far ex­ceeded your ex­pec­ta­tions. Can you give us some back­ground on it, and some idea of what is com­ing down the pipe­line?

The name for the Creative De­struc­tion Lab comes from TM: econ­o­mist Joseph Schum­peter, who was one of the first to think deeply about the process of in­no­va­tion. He coined the term ‘creative de­struc­tion’ to cap­ture the idea that in­no­va­tion cre­ates new in­ven­tions that im­prove our lives; but at the same time, it can de­stroy a lot of value and put peo­ple out of work. So, there are pos­i­tives and neg­a­tives to in­no­va­tion.

The fun­da­men­tal in­sight that led us to launch the CDL was this: We have great sci­ence in Canada, but his­tor­i­cally, we have done a lousy job of com­mer­cial­iz­ing it and reap­ing the eco­nomic ben­e­fits. Too of­ten, the pat­tern goes some­thing like this: Cana­dian sci­en­tists in­vent some­thing amaz­ing, and an Amer­i­can en­tre­pre­neur de­vel­ops it into a prod­uct and reaps all the eco­nomic ben­e­fits. Our in­sight was that there is a mar­ket fail­ure in what we call ‘the mar­ket for judg­ment’. The prob­lem for a new en­tre­pre­neur is, we’ve got great sci­en­tists com­ing up with amaz­ing in­ven­tions — but they don’t know the first thing about whose prob­lem it could solve, or how to start-up and scale a busi­ness. Most of what oc­curs be­tween in­ven­tion and rev­enue is man­age­ment.

Rot­man Pro­fes­sor Ajay Agrawal’s found­ing vi­sion for CDL was to at­tract re­ally promis­ing deep sci­ence-based ven­tures, and con­nect them with some of Canada’s most suc­cess­ful en­trepreneurs to re­solve the fail­ure in the mar­ket for judg­ment. These men­tors are peo­ple like Tony La­cav­era, who founded Glob­alive Hold­ings; Ted Liv­ingston, who started Kik; and John Fran­cis, who started Grounded. They come in and vol­un­teer their time, and we con­nect them to promis­ing sci­ence-based ven­tures. Our MBA stu­dents also do pro bono work for these ven­tures, adding the re­quired ca­pac­ity and an­a­lyt­ics to fig­ure out who the cus­tomer is, how the ini­tia­tive should be fi­nanced, and how it can be scaled. This gives Rot­man stu­dents a unique en­tre­pre­neur­ial ex­pe­ri­ence at the fron­tier of new tech­nol­ogy.

This for­mula has been hugely suc­cess­ful: Five years ago, we set a goal of the ven­tures go­ing through cre­at­ing $50 mil­lion dol­lars of eq­uity value. To­day, we’re clos­ing in on $1.5 bil­lion dol­lars of eq­uity value cre­ated. Of course, this is no­tional — they haven’t ex­ited; that’s based on the money they’ve raised.

We re­cently ex­panded be­yond the Univer­sity of Toronto: About a year ago we part­nered with the Univer­sity of Bri­tish Columbia to cre­ate a CDL at the Sauder School of Busi­ness; last spring we an­nounced part­ner­ships with Dal­housie’s Rowe School of Busi­ness, HEC in Mon­treal and the Haskayne School at the Univer­sity of Cal­gary. And in Oc­to­ber, we an­nounced a part­ner­ship with NYU’S Stern School of Busi­ness. The suc­cess­ful en­trepreneurs — or Fel­lows, as we call them — who men­tor CDL ven­tures come from Canada and be­yond: Peo­ple like Bar­ney

In­no­va­tion im­proves our lives; but at the same time, it can de­stroy a lot of value.

Pell, who holds a PHD in Ma­chine Learn­ing and led the NASA team that flew the first AI into deep space; and Shivon Zilis, who founded Bloomberg’s AI in­vest­ment arm, Bloomberg Beta.

About three years ago, when we de­cided to re­ally fo­cus on AI at Rot­man, it was a case of com­ing back to where it all started. The Univer­sity of Toronto is home to Ge­of­frey Hin­ton — a Com­puter Sci­ence pro­fes­sor who is one of the pi­o­neers and global gu­rus of ar­ti­fi­cial in­tel­li­gence. Around 2012, Prof. Hin­ton and his team of Phds started win­ning lots of global prizes in AI — par­tic­u­larly around pic­ture recog­ni­tion, which is a clas­sic AI prob­lem. All of a sud­den, Sil­i­con Val­ley lead­ers were com­ing up to Toronto to hire many of his PHD stu­dents. These peo­ple are now run­ning the AI labs at Ap­ple, Google, Uber and Face­book. So, this is another sad Cana­dian story, be­cause the Cana­dian gov­ern­ment ac­tu­ally funded a lot of this re­search — through what is known as the ‘AI win­ter’, when progress was very slow. Sud­denly, Prof. Hin­ton’s team started hit­ting home runs, and Sil­i­con Val­ley swooped in.

For the stu­dents and pro­fes­sors who want to stay in Toronto, we want to help them start and grow their busi­nesses right here. We’ve now got 100 Ai-ori­ented com­pa­nies go­ing through the Lab each year. As far as we can tell, it’s the big­gest con­cen­tra­tion of AI ven­tures of any pro­gram in the world. And in­stead of Sil­i­con Val­ley steal­ing our tal­ent, its lead­ers now fly up to Toronto reg­u­larly for CDL meet­ings, be­cause they want to in­vest in these ven­tures.

Can you de­scribe what these ven­tures look like?

Ev­ery year we are see­ing an ever-broad­en­ing sweep of apTM: pli­ca­tions across sec­tors. In the be­gin­ning, many ven­tures used AI to pre­dict some sort of fault or mal­func­tion: Think of pos­si­ble prob­lems with cars, planes, trains, drones, pipelines or any kind of big ma­chine. How many times have you gone to the air­port gate and you hear, ‘We have a me­chan­i­cal prob­lem; there’s go­ing to be a de­lay.’ This costs the air­lines bil­lions of dol­lars ev­ery year, and it’s a huge in­con­ve­nience for trav­el­ers. If they could do a bet­ter job of pre­dict­ing these prob­lems, air travel would be much more pleas­ant and safe, and the air­lines would dra­mat­i­cally drop their costs.

At CDL re­cently, we’ve been see­ing all sorts of ap­pli­ca­tions in health­care, in­clud­ing new types of di­ag­nos­tics that are up to 100 times cheaper than what we use to­day, and ap­pli­ca­tions for more per­son­al­ized medicine. The re­al­ity is, peo­ple re­act dif­fer­ently to dif­fer­ent drugs and treat­ments — and AI can pre­dict how, say, a par­tic­u­lar can­cer pa­tient will re­act to a cer­tain treat­ment.

These are just some ex­am­ples. There are many more I could talk about, but what I would un­der­line is that the only lim­i­ta­tion is our imag­i­na­tion.

Michael, Sco­tia­bank is one of the CDL’S part­ners, but you are also in­volved with the Vec­tor In­sti­tute and Nex­tai. Tell us about the strat­egy for these part­ner­ships.

A few years ago, we rec­og­nized that the world was mov­ing so MZ: fast in terms of ad­vance­ments in AI and re­lated fields, that there was no way for us to ‘know it all’ in­ter­nally. So, in Sco­tia­banks­peak, we took an ‘out­side/in per­spec­tive’. Large en­ter­prises of­ten think that only peo­ple who grew up in the or­ga­ni­za­tion know best; but we knew that wasn’t the case.

In terms of strat­egy, the main rea­son for our part­ner­ships is sim­ple: Gain­ing ac­cess to new ideas. Even though many AI ideas don’t di­rectly re­late to fi­nance, you can of­ten look at, ‘What are these en­trepreneurs and sci­en­tists try­ing to achieve?’ and fig­ure out the ‘fi­nance equiv­a­lent’ of that. It could be around find­ing anom­alies, de­tect­ing pat­terns, or just re­duc­ing the cost of prediction at some level.

The sec­ond point is equally im­por­tant to us: There is a mas­sive tal­ent short­age right now, in terms of peo­ple who un­der­stand AI and can ap­ply it. At the sci­en­tific level, how do you prac­ti­cally ap­ply deep learn­ing al­go­rithms? And at the busi­ness level, once you’ve got the tool, how can you use it in a trans­for­ma­tional sense? We thought, we can sit here all day and com­plain about all the change tak­ing place — or we can team up with great in­sti­tu­tions like Rot­man and ini­tia­tives like Nex­tai and do some­thing about it. Canada has a great op­por­tu­nity to be a leader in the AI realm, and we want to be part of that.

The third as­pect is, while a lot of ad­vances have been made, what academia of­ten re­ally needs are prac­ti­cal-use cases. It is very dif­fi­cult to get ac­cess to real-world data. The good news is, banks col­lect a lot of data; the bad news is, it gets trapped in si­los, be­cause his­tor­i­cally, banks have op­er­ated in si­los.

To com­bat this, we have started sev­eral ini­tia­tives. For

ex­am­ple, we anonymized and ag­gre­gated cer­tain data sets, and gave our aca­demic part­ners ac­cess to it, sub­ject to ap­pro­pri­ate se­cu­rity and con­fi­den­tial­ity ar­range­ments. This en­ables stu­dents to ex­plore what kind of in­ter­est­ing in­sights and al­go­rithms can be de­vel­oped, lead­ing to di­a­logues and all sorts of op­por­tu­ni­ties to de­velop so­lu­tions for our cus­tomers.

Can you talk a bit about how you’re lever­ag­ing third-party datasets as part of your AI strat­egy?

If you work for a large or­ga­ni­za­tion, never un­der­es­ti­mate the MZ: chal­lenge of just get­ting at the data that you think you’ve al­ready got. As in­di­cated, it likely sits in dif­fer­ent si­los, which quickly raises or­ga­ni­za­tional is­sues and data-gov­er­nance is­sues. Even if you can get ac­cess to it, is it in the right for­mat? The com­plex­ity of real-world data is a ma­jor is­sue. Get­ting our own data or­ga­nized so it be­comes ‘AI friendly’ is a crit­i­cal ex­er­cise.

Also, Sco­tia­bank is very ac­tive in var­i­ous Latin Amer­i­can mar­kets, and we see a tremen­dous need there to pro­vide bet­ter ser­vice to our small busi­ness part­ners and cus­tomers. It’s not easy for a small busi­ness to get along with any of the big banks, be­cause their pro­cesses weren’t de­signed for small busi­nesses.

Of course, there are fin­techs out there that have a very dif­fer­ent ap­proach. Cur­rently, we’re part­ner­ing with one called Kab­bage, which works with small busi­nesses. We re­al­ized that, if they pro­vided us with ac­cess to what these mer­chants were sell­ing and when they were get­ting paid, we could de­ter­mine the credit-wor­thi­ness of in­di­vid­ual mer­chants — and dra­mat­i­cally ac­cel­er­ate the loan-ap­proval process. We can now say to small busi­ness lead­ers, ‘Here is the tra­di­tional process for get­ting a loan; and here is an ex­pe­dited process that we of­fer with our part­ner, Kab­bage. If you agree to share some of your data with us, you can go down the lat­ter path, which is much faster’. This is just one ex­am­ple of how we’re us­ing data in new ways.

SMES [small-to-medium-sized en­ter­prises] are an im­por­tant com­po­nent of our econ­omy. Given that they don’t gen­er­ate or have ac­cess to huge amounts of data, how can they em­brace AI?

On the one hand, for com­pa­nies that have large amounts of TM: data, there are huge economies of scale and net­work ben­e­fits to be had. Just look at Google or Alibaba: These are un­be­liev­ably data-in­ten­sive com­pa­nies that are thriv­ing thanks to AI. On the other hand, dif­fer­ent types of dig­i­tal tech­nol­ogy are ben­e­fit­ing SMES. For ex­am­ple, Cloud com­put­ing. At one time, if you had a small busi­ness, you had to pur­chase your own servers, but to­day, you don’t have to do that — and as a re­sult, these com­pa­nies can scale them­selves much faster than in the past. Also, in a dig­i­tal world, you of­ten don’t need to build a fac­tory, and you can ac­cess global mar­kets di­rectly by sell­ing on­line.

It’s still early days for AI, but as it be­comes more main­stream and gets pack­aged and sold to busi­nesses, there will be ways for SMES to lever­age it. For ex­am­ple, one of the big­gest prediction prob­lems for a small busi­ness is, pre­dict­ing your cash flow, and there is al­ready a com­pany out there build­ing an AI en­gine to do that flow for small busi­nesses. That is a well-de­fined prediction prob­lem: A small busi­ness doesn’t have mas­sive amounts of data — but it does have all of its fi­nan­cial in­for­ma­tion for the life of the busi­ness. So, there will be ap­pli­ca­tions for SMES. Ob­vi­ously there’s an en­try cost, and you need to look at whether you can part­ner with new ven­tures to ac­cel­er­ate your progress; you don’t have to build it all your­self.

We would be re­miss if we didn’t touch on hu­man cap­i­tal. Tiff, what are your thoughts on how AI is go­ing to af­fect jobs and com­pe­ten­cies?

If you be­lieve, as we do, that AI is dra­mat­i­cally drop­ping the TM: cost of prediction, this means that jobs in­volv­ing a lot of prediction are go­ing to see de­clin­ing de­mand and lower wages. On the other hand, jobs that are com­ple­men­tary to prediction will do well. In the realm of health­care, if you work as a ra­di­ol­o­gist, spend­ing most of your time look­ing at x-rays, very soon, AI is go­ing to be able to read x-rays faster and more re­li­ably. But, if your job is to care for those peo­ple, or fig­ure out what treat­ment they need next — those skills are only go­ing to rise in de­mand.

I don’t want to min­i­mize the dis­rup­tive ef­fects that AI will have on so­ci­ety. If we see re­ally rapid progress, it will have se­ri­ous im­pli­ca­tions. We are al­ready see­ing this play out in the world: Which two coun­tries have the high­est lev­els of in­equal­ity in wealth dis­tri­bu­tion? The U.S. and the UK; and the con­se­quences of that are the elec­tion of Pres­i­dent Trump and

Most of what oc­curs be­tween in­ven­tion and rev­enue is man­age­ment.

Brexit. The con­se­quences have noth­ing to do with the prob­lem, re­ally; but they are fea­tures of the anx­i­ety lots of peo­ple are feel­ing. And, we shouldn’t kid our­selves in Canada: We do have a more re­dis­tributed tax sys­tem, but we’re see­ing the same trends and the same anx­i­ety. As a so­ci­ety, we need to man­age this a lot bet­ter.

Michael, how is AI im­pact­ing hu­man cap­i­tal at Sco­tia­bank?

Ear­lier, Tiff touched on the first-gen­er­a­tion uses for AI MZ: in fi­nan­cial ser­vices, and col­lec­tions and fraud are two great ex­am­ples. It’s ac­tu­ally a sig­nif­i­cant chal­lenge to sys­tem­at­i­cally op­ti­mize col­lec­tions, be­cause you need to de­ter­mine sev­eral things: Out of all the cus­tomers who don’t pay up ini­tially, who will even­tu­ally pay? And for those that will not pay you oth­er­wise, should you con­tact them when they’re one day late or 10 days late? How should you ask for a prom­ise to pay, and how in­sis­tent should you be? This ac­tu­ally re­quires a fairly com­plex re­source-or­ga­ni­za­tion model, and we’ve al­ready seen great re­sults from us­ing ma­chine learn­ing to op­ti­mize our col­lec­tion process.

Fraud is more about ‘anom­aly de­tec­tion’, and we end up chas­ing a lot of false pos­i­tives. That not only wastes un­told re­sources, but can neg­a­tively im­pact our cus­tomers. Say you buy an ex­pen­sive ring for your part­ner, and one of our alerts goes off. Sud­denly, your credit card is de­clined, be­cause this un­usual pur­chase raised a red flag. These sit­u­a­tions will be han­dled far bet­ter through AI.

Sta­tis­ti­cal tools have been around for hun­dreds of years. The way most tools tra­di­tion­ally work is, you say, ‘I have a view that A de­pends on B. So, let me run some re­gres­sion anal­y­sis and val­i­date that view’ and you com­plete it. Through this, good or bad out­comes are de­ter­mined and you mod­ify the model. Ma­chine learn­ing in­verts that process, so the user says, ‘I am not pos­tu­lat­ing any­thing up front: The data will tell me’.

That is a fun­da­men­tally dif­fer­ent way of think­ing, and it turns de­ploy­ment into much more of a ‘test-and-learn’ ap­proach, be­cause the data will tell you dif­fer­ent things, and you then need to test it on a larger scale. To fos­ter this mind­set shift, we have trained ‘dig­i­tal coaches’ in place to help our teams ad­just to the test-and-learn ap­proach.

How do you find a bal­ance be­tween in­no­va­tion and risk man­age­ment?

I strongly be­lieve that at the end of the day, au­toma­tion will MZ: al­ways re­duce risk, be­cause the ma­jor­ity of risk we see in prac­tice re­lates to hu­man er­ror — and ul­ti­mately, au­toma­tion re­duces hu­man er­ror. The idea of test­ing and learn­ing is ex­tremely im­por­tant. By try­ing out dif­fer­ent things early, you will catch mis­takes ear­lier. Tra­di­tion­ally, or­ga­ni­za­tions — par­tic­u­larly large ones — have a men­tal­ity where they try to de­fine ev­ery­thing up front, build it, and roll it out. We just ex­pect that our cus­tomers will like it, and that ev­ery­thing will work prop­erly; but usu­ally, there are chal­lenges. It’s much bet­ter to get cus­tomers in within weeks of start­ing some­thing to ob­tain early feed­back, and run the first al­go­rithm within weeks of de­vel­op­ing the model. The big take­away for me is that the po­ten­tial of AI out­weighs the risk.

Tiff Mack­lem is Dean of the Rot­man School of Man­age­ment. He also chairs the board of the Global Risk In­sti­tute and On­tario’s Panel on Eco­nomic Growth and Pros­per­ity. A mem­ber of the Asian Busi­ness Lead­ers Ad­vi­sory Board, he was for­merly the Se­nior...

Newspapers in English

Newspapers from Canada

© PressReader. All rights reserved.