Salesforce captures the limits of AI in a Coca-Cola cooler

SUMMARY:

Salesforce shows both the potential and the limits of AI in its demonstration today of Einstein Vision counting the stock in a Coca-Cola cooler

Marc Benioff and Parker Harris at Salesforce 4Q17 kickoff, screengrab from live stream
Marc Benioff and Parker Harris

Even though it was intended as a bravura demonstration of the business potential of artificial intelligence (AI), Salesforce today perfectly captured the technology’s current limitations with the aid of a Coca-Cola cooler cabinet.

The demonstration was part of the vendor’s customer kick-off event for its new financial year — the first in a series that will set the tone for Salesforce’s marketing from now until its annual Dreamforce conference in November.

The event was an intimate affair, though watched remotely by a huge online audience, held in a conference room on the 23rd floor of Salesforce’s temporary headquarters. Visible from the wall-to-ceiling windows was the towering skeleton across the road of San Francisco’s tallest building, destined on completion to become the cloud giant’s new corporate home. Co-founders CEO Marc Benioff and CTO Parker Harris blew out candles on a birthday cake to mark the eighteenth anniversary tomorrow of the now $8.4 billion revenue company‘s founding. But the intended star of the event was the Coca-Cola cooler cabinet.

Practical benefits of Einstein AI

Salesforce wanted to show off the practical business benefits of its Einstein AI technology, freshly available to customers and newly supplemented by integration to IBM Watson. Longstanding customer Coca-Cola was glad to oblige, with the aid of Salesforce Chief Scientist Richard Socher, whose deep learning startup MetaMind was acquired by Salesforce a year ago for $33 million.

The demonstration showed off the newly launched visual recognition engine Einstein Vision, which according to the blurbs enables customers to:

… leverage pre-trained image classifiers, or train their own custom classifiers, to handle a vast array of specialized image-recognition use cases.

Socher showed how Einstein had been trained to recognize, identify and count the varieties and quantities of Coca-Cola bottles stored in one of its cooler display cabinets, simply by analyzing a photo taken with an iPad or iPhone. There’s no longer any need for a Coca-Cola representative to physically visit and count the contents, he said.

What’s more, Einstein can then take that stock count and combine the information with predictions based on known seasonal variations, weather information from Watson, and upcoming promotions, to automatically calculate a restocking order. Socher grandly concluded:

In a few clicks I could have this order completely automated and I could rely on Salesforce Einstein AI to make decisions for my businesss.

Glossed over

Socher went on to demonstrate how the outlet could confirm the order via mobile chat with an AI agent that automatically offered appropriate delivery choices and relevant upsell options. Coca-Cola CIO Barry Simpson was on hand to explain that using AI in this way would enable his company to maximize its sales potential across the 16 million coolers installed in retail outlets worldwide:

Being able to find better ways to optimize our inventory and turn that into actions for our customers as fast as we can and then optimize our sales in that outlet, that’s very important for us.

Taking global products and global ideas and executing them locally … is a huge opportunity for us.

But I couldn’t help noticing that the demonstration glossed over some gaps in capability that still need to be filled in. True, Einstein Vision was able to recognize and count each bottled product — but only the front row. It didn’t yet have the ability to see how deep the bottles were in each position. And the Coca-Cola cooler is a highly standardized cabinet containing a very limited set of product options. The visual recognition Einstein is being asked to do in this demonstration is hugely constrained.

As a proof-of-concept, this works well — and you can imagine that it might soon deliver value in this highly constrained Coca-Cola example. But don’t be fooled into extrapolating this into an ability to instantly do stock counts on any display cabinet in the grocery store. That’s a completely different proposition.

Limits of AI

That’s why this demonstration so perfectly encapsulates both the potential and the limitations of the current state-of-the-art in artificial intelligence. AI only works well when the parameters are well defined. Point it at a standardized cabinet with a limited set of products to recognize and it won’t take long to learn what it has to do because there aren’t that many “custom classifiers” that it has to deal with.

Machines are great at doing the same thing over and over — much better than humans, especially if there’s lots of math involved. But if you’re dealing with an environment where the patterns keep on changing and there are unpredictable quantities of unknown variables, you’d better rely on a human. They won’t get it right all the time, but at least they’ll figure something out.

That’s why machines keep on challenging humans at games. Whether it’s checkers, chess, Go or even poker, so long as you constrain the environment to a well-defined set of rules and variables, the machines will eventually learn how to beat the humans.

The challenge when applying AI in a B2B environment is that there are so many variables, so many different sets of rules. As I discussed when assessing procurement vendor Coupa’s recent AI acquisition, the key to success here is classification — establishing the framework that AI uses to interpret the data it’s analyzing.

This is why, when Einstein launched at Dreamforce last year, its general manager John Ball, was careful to qualify its capabilities:

The AI is for CRM, it’s AI for Salesforce. We’re not trying to be a general purpose AI platform.

In that comment, he was specifically addressing the issue of custom data that many customers collect in their Salesforce systems. The less common the use case, the less data there is for the AI to learn from and the less likely it will be able to frame it in a suitable classification.

Einstein learns to teach itself

Therefore one of the most exciting tidbits to come out of the current round of briefings is that Einstein is gaining the ability to teach itself classification. As Ball explained to Techcrunch this week, Einstein is beginning to use its AI to generate AI models:

You can throw the kitchen sink at it and we figure it out using math.

He went on to explain that the AI engine continually hones its models as it develops them and tests them against the data. That capability will prove really important given how many customers have their own custom data fields.

Today, the Coca-Cola cooler perfectly sums up the state of play with AI in the B2B field — it’s ideal when the number of variables are limited and predictable. It’s not so handy when the classification is more complex — unless it’s a specific field where there’s already been an opportunity to analyze and learn from lots of existing data, such as IBM Watson’s Weather Company service.

But there’s still a lot of scope for that to change as the machines get better at defining classifications and collect more data with which to test those frameworks. So long as we recognize that much of what we see today is still at proof-of-concept stage, there’s still plenty of justification for excitement about what AI will ultimately enable us to achieve.

Image credit - Marc Benioff and Parker Harris at Salesforce 4Q17 kickoff, screengrab from live stream

Disclosure - Coupa and Salesforce are diginomica premier partners at time of writing