The adoption of artificial intelligence (AI) is increasing.According to a recent McKinsey survey, 55% of companies use artificial intelligence in at least one function, and 27% attribute to AI at least 5% of the benefits before interest and taxes, much of them in them incost savings form.
Since IA will drastically transform almost all the sectors it touches, it is not surprising that suppliers and companies seek opportunities to display AI in all places that can.But not all projects can benefit from AI and try to apply it inappropriately can not only cost time and money, but it can also sour employees, clients and corporate leaders in future AI projects.
The key factors to determine if a project is suitable for AI are business value, the availability of training data and cultural disposition for change.Next, we show you how to ensure that these criteria conform to your proposed project before your incursion into artificial intelligence becomes a sunk cost.
Start with the simplest possible solution
Data scientists, in particular, gravitate to a first AI approach, says Zack Fragos.000 premises in more than 90 countries around the world.But IA cannot be applied everywhere.
Despite being a very traditional business line, Domino's has adopted change, especially during pandemic.Customers now have 13 digital ways to ask for pizzas, and the company generated more than 70% of sales through digital order channels in 2020.That has opened many opportunities to realize the promise of AI.
The key to Domino's in the AI application, says Fragoso, has been to adopt a simple approach."In the end, the simple solution works faster, has a better performance and we can explain it to our business partners," he says."The explainability is an important part: the more people understand the tools and methods we use, the easier your adoption will be".
The focus 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 get from there, there must be added value in the performance of the model," he saysRocky.
For example, predicting the time it takes to cook a pizza and put it in a box is simple."We take it directly from our operations research.Cooking times can be introduced, "he says.But there are some problems that can only be solved with AI, he adds, such as those that require recognition of images or natural language processing.
For example, last year, Domino's launched a loyalty program that rewarded customers for eating pizza, any pizza, from any pizzeria."We built a pizzas classifier using millions of photos of different types of pizza and put it in an application," explains fragous.
This project offered two types of business value.First, the customer experience improved, he says.Second, he created a collection of pizzas images that the company used to detect quality and its temperature."It was a very complete artificial intelligence project," he says.
A more practical project that Domino's undertook was a predictor aimed at improving the accuracy of his pizzas tracker, since customers want to know when they should go exactly to the store to collect their food, or when they must wait for their delivery to arrive,He says fragous.Adding automatic learning to the traditional coding of the Domino's pizzas tracker.
When building the model, Domino's swam to his principle of "the simplest"."The first iteration was a simple regression model," he says."That brought us closer to us.Next, a decision tree model, with which we could examine more facets.Then we went to a neuronal network because we could capture some of the same variables as in the decision tree, but the neural network produces the response more quickly.We want the customer experience on the website to be really fast ".
There is a place for automatic learning, acknowledges Sanjay Srivastava, Digital Director of Genpact, especially when a company seeks to build processes that improve continuously on the basis of experience.But sometimes everything that is needed is a simple correlation, which can be obtained from basic statistical models.
"The practices of ten years ago around random forests and other statistical tools kits can obtain the response much faster and much cheaper than building a whole team of mlops around it," says Srivastava."You have to know when to resort to existing techniques that are much simpler and much more effective".
A common area in which AI is often presented as a solution, but it is usually excessive, it is in the chatbots, that specialist considers: "In some scenarios, it makes sense.But in 90% of cases the questions to be asked are known, because you can look at the questions that have been asked in the last three years and the answer to each of them is known.It turns out that 90% of chatbots can get ahead with simple pairs of questions and answers ".
Historical data: the AI key to predict future results
Any finite data set can be adjusted to a curve.For example, you can take the winning numbers of the previous years lottery and create a model that prefers them perfectly.But the model will still be better to predict future profits because the underlying mechanism is completely random.
COVID-19 Pandemia has been an excellent example of how this happens in real life.There was no way to predict if the closures were going to cause the closure of the factories, for example.As a result, companies saw a decrease in increasing income in many areas, according to the McKinsey AI state survey.
For example, 73% of respondents saw increase their income in corporate strategy and finance last year, while only 67% did this this year.The difference was even more marked in supply chain management.Last year, 72% saw income increases in this area, but only 54% did this year.
"The fundamental characteristic of AI or Automatic Lear."You are married, chained, handcuffed by history.AI is good in circumstances in which history is likely to be repeated, and you think about the story is repeated ".
For example, he says, some of his clients have tried to use AI to predict future income.But often, income is influenced by factors that cannot be predicted, that cannot be controlled and of which the company has no data.And if any of these factors have an external impact on the results, the entire model can be released.
"Then it makes no sense to choose the AI," he admits."Are you going to invest hundreds of thousands of dollars in a solution that can become immediately irrelevant for a change in a variable?"
The AI can still have a role here, says that specialist, by helping to model several scenarios, or by bringing ideas that may not be evident in another way."Your probability of success increases if your approach is narrower".
The AI will also fall short if the presence of AI changes the behavior of the system.For example, if AI is used to filter hate expressions, people quickly learn which patterns seek AI and write things for filters.
"The best minds in the world have tried to solve these problems and have not been successful," says Carroll.
Bharath Thota, Kearney partner, once worked with a world conglomerate of products and consumer goods of more than 30.000 million dollars.The CFO management team wanted a better visibility of the conglomerate financial metrics to see if their growth was going up or down.The existing process was that they received the reports in PDF 30 days after the information period was closed.
The data science team applied the AI to predict what the numbers would be like."They had a good intention," Hota acknowledges."They wanted to offer the address a futuristic vision".
How to fix error 3194 onios https: // t.CO/HECP7WUJD9 https: // t.CO/MUEHLQLYT
— Crios Music Fri Jul 15 11:52:40 +0000 2016
The mistake they made was in the financial data that introduced into the algorithm.Financial analysts who fed this data had to make many assumptions, so the data set ended up containing many individual biases.
"The direction was excited," says this."They had something that looked forward, not backward.But when the quarter ended and they saw those predictions again, they were completely wrong ".
The entire project lasted months, Thota acknowledges."They had to find out how to build this thing, do architecture, investigate AI platforms, get everything to work together.".
When such a project fails, people lose interest and trust in AI, in their opinion.In the case of this particular company, the solution simply consisted of creating for the CFO management team a financial command box that would provide them with the metrics they needed, when they needed them.
Over time, this says, some of AI was also used, in the form of natural language generation, to automatically provide key information about the data to executives in simple terms.
"It was a visibility problem," says that specialist."And there was a simple solution to provide that visibility".
Data challenge
Most AI projects require data.Good data, relevant data, data that are correctly labeled and without biases that can biased the results.
For example, a company that wants to prevent cats from entering a chicken coop can choose to install a camera and image recognition technology to detect cat entry.But success depends on having an adequate training set.
"You will need to have many photos, and those photos must have labels about whether there are cats in them or not," says Gartner Whit Andrews, to which he adds that the collection of this data takes a long time and is expensive.And once gathered, can the company reuse the same data set for other projects?
But what if it turns out that the company needs to know how many cats enter the chicken coop?In that case, this original set of images with the number of cats that appear in each of them will have to be re -estimated.
"A cat may not be so expensive, but a herd of cats is a problem," says Andrews.
In addition, if only a small percentage of images contains several cats, getting a precise model will be much more difficult.
This situation is frequently given in marketing applications, when companies try to segment the market to the point that data sets are infinitely small.
"Almost all the companies that I know use the segmentation to address customers," admits Anand Rao, a global partner and leader in PricewaterhouseCoopers.
If they collect data waiting for them to be used for one purpose and end up using them for another, the data sets may not meet the new requirements.
For example, if data collection is established so that there is a balance of data points of each region of the United States, but the commercial question ends up being about the needs of a very narrow demographic segment, all inferences will be useless.Let's say, for example, that the company is interested in the purchase habits of Asian-American women of a certain age range, and there is only a couple in the sample.
"Be very clear what decision wants to make with his segmentation," says Rao."Try to ensure that the sampling you are doing is representative, but also captures your questions".
The sampling problem occurs in any system that attempts to predict rare events.For example, if a company looks for examples of fraudulent behaviors, in a data set of one million transactions, there are a handful of known fraudulent transactions, and an equal or greater number of fraudulent transactions that have been overlooked.
"That is not very useful to infer," says Rao, and adds that this happens a lot with the automation of business processes when a company has many people who do particular tasks every day, but do not capture data on how those tasks are being done, or not capture the correct data necessary to train an AI on how to do it.
"In those cases, you have to build a system to capture that information," he adds."Then, a few months later, return and build the model".
And for projects that do not need data, AI is not the right path.For example, some business processes, such as insurance and subscription, are based on rules, says Rao."You can build a rules based system interviewing experts and gathering traditional formulas.But if you can do it with rules and scripts, it does not need ia.It would be an exaggeration ".
Use an AI for such a project may require more time and precision may not be better, or only slightly better, or improved performance may not be needed.
"So you will not have the return on investment because you are spending time on a problem that you could have already solved," he says.
An error of 300 million dollars in AI
In November, the real estate company Zillow announced that it annulled homes worth 304 million dollars that it had bought based on the recommendation of its Zillow Offers service, promoted by the AI.
It is possible that the company has to amortize another 240 or 265 million dollars next quarter, in addition to dismissing a quarter of its workforce.
"In our short period of operation of Zillow Offers, we have lived a series of extraordinary events: a worldwide pandemic, a temporary freezing of the real estate market and then an imbalance between the supply and demand that caused an increase in prices ofHousing at an unprecedented rate, "said his CEO, Rich Barton."We have been unable to precisely forecast the future prices of the house, to adjust our models based on what we have learned and move on.But, based on our experience to date, it would be naive to assume that unpredictable pricing and disturbance forecast events will not occur ".
The AI learns from the past, says Tim Fountaine, McKinsey's main partner."If something has not happened in the past, it is impossible for an algorithm to predict it".
And IA do not have common sense, add."An AI algorithm designed to predict the production of a factory that has never seen a fire before, will not predict that production will collapse if there is a fire".
Predict the prices of real estate is an interesting use of AI, he acknowledges, to conclude with these words: "But you can see that everyone is becoming a little shy with that type of applications".