Tag Archives: The Future of Employment: How Susceptible Are Jobs to Computererisation

Hiding from the Computers Part 3: “Unexpected Item in Bagging Area”

Time for some micro. If we go back to the Frey and Osborne job destruction chart (see my post here for a more detailed explanation), we can see that the Sales and Related (the red) category is one of the first to get crushed (click for larger image).

Probability of Computerisation jpeg

But internet shopping, along with Semi-Automated Customer Activated Terminals (SACATs) telling us that there is an “unidentified object in the bagging area”, have been around for well over a decade. To get a sense of what has been going on in the US and UK you need a comparable data set, and the International Labour Organisation (ILO) is the place to find it (here).

Frustratingly, the time series only go up to 2008. Nonetheless, the data line up almost perfectly with the Osborne and Frey chart (click for larger image):

US Employment by Occupation jpeg

The three worst-performing employment categories were agriculture, manufacturing, and wholesale/retail, all showing declines against a 5.5% rise in US employment as a whole for the period. Incidentally, wholesale/retail is the largest employer in the US: 20.6 million of the 145.4 million Americans in work in 2008, or 14.2% of the employed workforce. And the UK: ditto (click for larger image). Retail/wholesale shrank by 1%, only beaten by the incredible 14.0% slump in manufacturing employment.

UK Employment ILO Data jpeg

For more granular and up-to-date data, we have to go to national statistical offices. Here is how employment has evolved in the US retail sector over the last decade according to the United States Department of Labor’s Bureau of Labor Statistics (see here, click for larger image):

US Employment in Retail Sector jpeg

Total employment in the Sales and Related sector grew 3.7% from 2002 to 2012 to reach 13.8 million, with a noticeable downward blip over the Great Recession. Employment for cashiers, however, declined by 1.8% to 3.3 million, with no bounce back after the Great Recession. Why the difference? Well, my guess is that the difference is because of these:

Self check Out Machines jpeg

Where I live in the UK, they have become ubiquitous, although there do exist certain national brand supermarkets that don’t use them. Nonetheless, Semi-Attended Customer Activated Terminals (SACATs) are still relatively basic in functionality and can only replace a part of what traditional cashiers do. So let’s break down the job of a cashier into parts, and then think about which bits existing SACATs are replacing and which parts the new generation of SACATs will be able to replicate. The cashier job description from the UK National Careers Service.

Checkout Operator Profile jpeg

To this list I would add the vetting of purchases of alcohol to see that the purchasers comply with age restrictions.

If we look at this list, we can see that old-style SACATs can only really do bullet point 2 and 7, and partially 6. Moreover, the customer is roped into performing bullet point 1 and 5. It’s actually a quite limited repertoire and stems from the bottleneck variables in the Frey and Osborne paper, in particular finger and manual dexterity (plus visual perception, which they don’t include).

O Net Variable Jan 14 jpeg

The SACAT has no manual dexterity. It co-opts the customer to pick up irregular-sized groceries and present them to the machine with the bar code in the correct position. Incidentally, bar codes technology has, in itself, eliminated the finger dexterity required to input prices into a cash register.  However, a 360-degree checkout called Rapid Scan, currently being trialed by the supermarket operator Asda in the UK (here), makes the need for the customer (or cashier) to jiggle each good redundant and is three times faster than traditional hand-scanning approaches.

360 Degree Checkout jpeg

Such visual perception issues don’t show up in In the Frey and Osborne chart of technology bottlenecks above, perhaps because they think the technology is already there. However, while the visual perception of the Google self-drive car may be better than that of a human’s, it certainly isn’t at a price point ready for mass deployment.

As an example, the old-style SACATs can’t recognise fruit and vegetables. From a Wall Street Journal article generally downbeat on the automated checkout:

What’s so cognitively demanding about supermarket checkout? I spoke to several former checkout people, and they all pointed to the same skill: Identifying fruits and vegetables. Some supermarket produce is tagged with small stickers carrying product-lookup codes, but a lot of stuff isn’t. It’s the human checker’s job to tell the difference between green leaf lettuce and green bell peppers, and then to remember the proper code.

“It took me about three or four weeks to get to the point where I wouldn’t have to look up most items that came by,” said Sam Orme, a 30-year-old grad student who worked as a checker when he was a teenager.

Another one-time checker, Ken Haskell, explained that even after months of doing the job, he would often get stumped. “Every once in a while I’d get a papaya or a mango and I’d have to reach for the book,” he said.

But out of Datalogic, we have just seen the market rollout of checkouts with Visual Pattern Recognition (ViPR) that can identify items without a barcode.

Datalogic Visual Pattern Recognition jpeg

Intriguingly, this checkout incorporates machine learning (ML). It has a hit rate of around 98%, with the 2% of unidentified object requiring recourse to a human operator. However, when identified, the image and the object go back into the database, improving the identification of all such machines in future.

But what about those annoying fruit and vegetables. Well Datalogic has a whole host of patents pending that deal with this particular problem. Here is one that incorporates X-ray fluorescence in the Visual Pattern Recognition software:

Unlike a visual recognition system, a system using X-ray fluorescence utilizes the chemical makeup of the item or produce being tested for recognition. The problems in visual systems associated with attempting to interpret variations in color, texture and size are minimised …. because the X-ray fluorescence system analyzes a fixed chemical composition. For example, kiwi has distinctly different chemical makeup from that of a potato, and so the X-ray identification process that utilizes the chemical makeup for identifying a product can be quite precise. Further, effects of ambient light and other visually sensitive environmental factors become less of a concern and less intrusive into the object recognition process. As an additional advantage, the FIG. 12 system can make it possible to recognize and characterize a paper bag with produce in it, in which case there is not a need to take the produce or other items to be recognized out of the bag to be identified.

So not only can we now identify the difference between a kiwi and a potato, we can do it without taking the items out of their paper bag! Human operator: beat that.

Another area in which humans outcompete SACATs in the visual perception sphere is in asking young-looking purchasers of alcohol to produce ID. Yet here again facial recognition software is entering the mainstream. From Bloomberg:

In the U.K., Tesco Plc (TSCO) is installing face-scanning technology at its gasoline stations to determine customers’ ages and gender so tailored advertisements can be delivered to them on screens at checkouts.

OK, let’s think about money. A pod of old-style SACATs (4 terminals) costs around $60,000, but, as with all such technology, I expect the cost to rapidly decline. I have no idea what the depreciation schedule is for a SACAT, but in line with similar equipment I would guess it is 4 years. Taking a straight-line depreciation, you would be expensing a pod at $15,000 per annum. On top of this, we have to take into account the weighted average cost of capital (WACC) to purchase the original equipment. For Walmart, it is 6% (here). So that’s another $3,600 per annum. Let’s round the whole package up to $20,000 (there will be some maintenance costs in there as well).

So how many cashiers can you buy for $20,000? From the US Bureau of Labor Statistic again:

Cashier Wages jpeg

But $20,000 per annum (roughly £12,000) is the direct wage. Even for a notorious employee-benefit miser like Walmart, the all-in cost is going to be higher (you at least have to provide staff toilets). I will lowball it at $5,000 on top of the base wage. So the cost of running a SACAT pod will be 80% of one cashier. However, a pod runs 24/7, a cashier doesn’t. The cashier also takes days off and holidays. With an average American working 260 days per year, the cashier is working just under 75% of a year. (OK, the cashier may work more days to feed the kids, but if he or she does, even Walmart has to pay more.)

We can now create an efficiency equation to find out at what point the SACAT pod is in equilibrium with human cashiers (that is the point at which the status quo, ratio of SACATs to human cashiers, continues) remembering that a pod has four checkout points (so we are equating hours of SACAT lane availability with human availability in term of dollar expense). Using the numbers above, the human cashier is 21 times as efficient as the SACAT, or the SACAT is only 5% of the efficiency of a human. Here I mean ‘efficiency’ in the broadest possible term—not just speed of processing a bunch of groceries. The number incorporates the ability to identify a banana, help put irregular-sized objects into an annoying plastic bag whose top doesn’t open, deal with special offer vouchers of increasing complexity, spot dodgy shop lifters, ask whether the bearded-teenage trying to buy a beer is really over 18, get a plastic security tag off a bottle of vodka and ask a cashier to work extra shifts on a busy Friday evening but not come in on a Monday afternoon.

Perhaps this is why SACATs have stalled the growth in cashier numbers, but not really reversed the number. Nonetheless, we are only at 5%! And Datalogic and its competitors are hell-bent at taking the SACAT into the other 95% of functionality currently occupied by humans.

In my next post, I will leave SACATs briefly and jump back to macro to see how economists have been struggling to explain developments in the labour market in the face of technological change over the last decade or so. After that, it will be back to SACATs to ask who benefits from them and how their installation, and other technologies like them, can dramatically change the shape of an economy. And, finally, I will ask: what is to be done?

Hiding from the Computers Part 1: The 47%

A rather dismal chart to kick off the New Year (click for larger image):

Probability of Computerisation jpeg

It’s taken from a paper by Frey and Osborne of Oxford University entitled “The Future of Employment: How Susceptible Are Jobs to Computerisation?”.

Let’s break this chart down. The total area under the curve is equivalent to aggregate US employment. The bottom x axis is the probability that any given job will be computerised: it ranges from zero, no chance of computerisation, to 1, a 100% chance of computerisation. Further, the authors have lumped these probabilities into three broad categories: 1) low probability, zero to 30% change of computerisation, 2) medium, 30% to 70% chance and 3) high, 70% to 100% chance.

So if you are doing a job in the high probability category, there is a high risk that your job will disappear over the course of time. How many people are in this category. Kindly they give the rather shocking number: 47%. Conversely, around 33% of the working population sit in the low probability category and can sleep easily at night for a little while longer.

Now the eagle-eyed will have noticed that there are no dates given in the chart over which the rise of the computers will take place. This is because the authors have looked at the problem as engineers, disassembling jobs into their component parts to decide which bits can be replaced by computers and which can’t. They don’t try to predict when the technology will reach the necessary maturity. In their words:

According to our estimate, 47 percent of total US employment is in the high risk category, meaning that associated occupations are potentially automatable over some unspecified number of years, perhaps a decade or two. It should be note that the probability axis can be seen as a rough timeline, where high probability occupations are likely to be submitted by computer capital relatively soon.

So, accordingly to their guestimate, computers will munch through the job market—moving from right to left in the chart—arriving in the medium probability category some time around 2030.You can also see where you don’t want to be; for example, the yellow band ‘production’ stinks along with orange ‘office and administration support’.

To arrive at this chart, the authors looked at 702 occupations and analysed them with respect to how susceptible each one was to computerisation. Computerisation here is taken in the broad sense to include machine learning, artificial intelligence and mobile robotics. To do so, they identified cognitive and non-cognitive bottle necks to computerisation and used these as variable to predict the order and extent of future computerisation. The bottleneck variables are given below:

O Net Variable Jan 14 jpeg

Further, a bottleneck such as manual dexterity could be classed as low, such as fitting a light bulb, to high, as for example performing open heart surgery.

Overall, the paper’s analysis suggests that the job market, as we know it, will be blown up over the next two decades, along with all our economic assumptions. Yet I hear not one politician talking about this risk. I will return to this subject in my next post.