Adoption of Artificial Intelligence (AI) is on the rise. According to a recent McKinsey survey, 55% of companies use artificial intelligence in at least one function, and 27% attribute at least 5% of profits before interest and taxes to AI, much of it in form of cost savings.
With AI set to dramatically transform nearly every industry it touches, it's no wonder vendors and enterprises are looking for opportunities to deploy AI everywhere they can. But not all projects can benefit from AI, and trying to apply it inappropriately can not only cost time and money, but can also sour employees, customers, and corporate leaders on future AI projects.
Key factors in determining whether a project is right for AI are business value, availability of training data, and cultural readiness for change. Here's how to make sure those criteria fit your proposed AI project before your foray into AI becomes a sunk cost.
Start with the simplest possible solution
Data scientists, in particular, gravitate toward an AI-first approach, says Zack Fragoso, director of data science and AI at Domino's pizza chain , which has more than 18,000 stores in more than 90 countries around the world. But you can't apply AI everywhere.
Despite being a very traditional line of business, Domino's has embraced change, especially during the pandemic. Customers now have 13 digital ways to order pizza, and the company generated more than 70% of sales through digital ordering channels in 2020. That has opened up many opportunities to realize the promise of AI.
The key for Domino's in applying AI, says Fragoso, has been to take a simple approach. "In the end, the simple solution works faster, performs better, and we can explain it to our business partners," he says. "Explainability is an important part: the more people understand the tools and methods we use, the easier it will be to adopt them."
The approach itself is simple: If there is a business problem that needs to be solved, Domino's looks for the simplest and most traditional solution, and then, "if we go up from there, there has to be added value in the performance of the model "Fragoso says.
For example, predicting how long it takes to cook a pizza and put it in a box is easy. "We pulled it directly from our operations research. Cook times can be entered," she says. But there are some problems that can only be solved with AI, he adds, such as those that require image recognition or natural language processing.
For example, last year, Domino's launched a loyalty program that rewarded customers for eating pizza, any pizza, from any pizza place. "We built a pizza classifier using millions of photos of different types of pizza and put it into an app," Fragoso continues.
That project delivered two kinds of business value. First, it improved the customer experience, he says. Second, it created a collection of pizza images that the company used to detect quality and temperature. "It was a very complete Artificial Intelligence project," he says.
A more practical AI project Domino's undertook was a predictor aimed at improving the accuracy of its pizza tracker, as customers want to know when exactly they should go to the store to pick up their food, or when they should expect it to arrive its delivery, says Fragoso. Adding machine learning to Domino's pizza tracker's traditional coding resulted in a 100% increase in accuracy, admits that specialist.
In building the model, Domino's adhered to its principle of "keep it simple." "The first iteration was a simple regression model," he acknowledges. "That got us closer. Next, a decision tree model, where we could look at more facets. Then we moved on to a neural network because we could capture some of the same variables as in the decision tree, but the neural network produces the response faster. We want the customer experience on the website to be really fast."
There is a place for machine learning, says Sanjay Srivastava, Genpact's chief digital officer, especially when a company is looking to build processes that continually improve based on experience. But sometimes all that is needed is a simple correlation, which can be obtained from basic statistical models.
"Practices from ten years ago around random forests and other statistical toolkits can get you the answer much faster and much cheaper than building a whole team of MLOps around it," admits that digital director . "You have to know when to fall back on existing techniques that are much simpler and much more effective."
A common area where AI is often presented as a solution, but often overkill, is in chatbots, he says: "In some scenarios, it makes sense. But in 90% of scenarios, you know the questions to ask because you can look at the questions that have been asked in the last three years and know the answer to each question. It turns out that 90% of chatbots can get by with simple question-and-answer pairs."
Historical Data: The AI Key to Predicting Future Outcomes
Any finite set of data can be curve-fitted. For example, you can take the winning lottery numbers from previous years and build a model that perfectly predicts them. But the model will still be no better at predicting future payouts because the underlying mechanism is completely random.
The COVID-19 pandemic has been an excellent example of how this happens in real life. There was no way to predict whether the lockdowns were going to cause factories to close, for example. As a result, companies saw revenue growth slow in many areas, according to McKinsey's State of AI Survey.
For example, 73% of respondents saw revenue increases in corporate finance and strategy last year, while only 67% did so this year. The difference was even starker in supply chain management. Last year, 72% saw income increases in this area, but only 54% did so this year.
"The fundamental thing about AI or machine learning is that you're using the story to inform," says Donncha Carroll, partner in the revenue growth practice at Axiom Consulting Partners. "You're married, chained, handcuffed by history. AI is good in circumstances where history is likely to repeat itself, and you're okay with history repeating itself."
For example, he believes some of his clients have tried to use AI to predict future revenue. But often, revenue is influenced by factors that cannot be predicted, cannot be controlled, and for which the company has no data. And if any of those factors have an external impact on the results, it can derail the entire model.
"how to fool your friends with a diary about coming from a different planet" is a real wiki page that exists for real. look it up.
— Nomashawn Wed Sep 07 11:08:21 +0000 2016
"So there's no point in choosing AI," he says. "Are you going to invest hundreds of thousands of dollars in a solution that can be rendered immediately irrelevant by a change in a variable?"
AI can still play a role here, he says, by helping to model various scenarios, or by bringing to light ideas that might not be otherwise apparent. "Your chance of success increases if your focus is narrower."
The AI will also fall short if the very presence of the AI changes the behavior of the system. For example, if AI is used to filter hate speech, people quickly learn what patterns the AI is looking for and redact things to get past the filters.
"The best minds in the world have tried to solve these problems and have been unsuccessful," says Carroll.
Bharath Thota, a partner at Kearney, once worked with a global conglomerate of products and consumer goods worth more than $30 billion. The CFO leadership team wanted better visibility into the conglomerate's financial metrics so they could see if its growth was going up or down. The existing process was that they received the reports in PDF 30 days after the reporting period closed.
The data science team applied AI to predict what the numbers would look like. "They meant well," says Thota. "They wanted to give management a futuristic vision."
The mistake they made was in the financial data they fed into the algorithm. The financial analysts feeding that data had to make a lot of assumptions, so the data set ended up containing a lot of individual bias.
"Management was excited," continues Thota. "They had something that looked forward, not backward. But when the quarter ended, and they looked at those predictions again, they were dead wrong."
The whole project took months, says Thota. "They had to figure out how to build this thing, do the architecture, research AI platforms, get everything to work together."
When such a project fails, people lose interest and trust in AI, he now admits. In the case of this particular company, the solution was simply to create a financial scorecard for the CFO leadership team that would give them the metrics they needed, when they needed them.
Over time, Thota says, some AI was also used, in the form of natural language generation, to automatically provide key information about the data to executives in layman's terms.
"It was a visibility issue," in his opinion. "And there was a simple solution to provide that visibility."
The Data Challenge
Most AI projects require data. Good data, relevant data, data that is correctly labeled and without bias that could skew the results.
For example, a company that wants to prevent cats from entering a chicken coop might choose to install a camera and image recognition technology to detect cat entry. But success depends on having a proper training set.
"You'll need to have a lot of photos, and those photos have to have tags about whether there are cats in them or not," says Gartner analyst Whit Andrews, adding that collecting this data is time-consuming and expensive. And once gathered, will the company be able to reuse the same data set for other projects?
But what if it turns out that the company needs to know how many cats are in the chicken coop? In that case, you will have to relabel that original set of images with the number of cats that appear in each of them.
"One cat may not be that expensive, but a pack of cats is trouble," says Andrews.
Also, if only a small percentage of images contain multiple cats, getting an accurate model will be much more difficult.
This situation often occurs in marketing applications, when companies try to segment the market to the point where the data sets are infinitely small.
"Almost every company I know uses segmentation to target customers," says Anand Rao, partner and global AI leader at PricewaterhouseCoopers.
If they collect data expecting it to be used for one purpose and end up using it for another, the data sets might not meet the new requirements.
For example, if the data collection is set up so that there is a balance of data points from each region of the United States, but the business question ends up being about the needs of a very narrow demographic, all inferences will be useless. Let's say, for example, that the company is interested in the buying habits of Asian-American women in a certain age range, and there are only a couple in the sample.
"Be very clear about what decision you want to make with your segmentation," says Rao. "Try to make sure that the sampling you're doing is representative, but also captures your questions."
The sampling problem occurs in any system that attempts to predict infrequent events. For example, if a business is looking for examples of fraudulent behavior, in a dataset of a million transactions, there are a handful of known fraudulent transactions, and an equal or greater number of missed fraudulent transactions.