Artificial Intelligence applications in business are becoming increasingly common. Part 1 of this article explored a possible structure for such projects and a few potential benefits. Part 2 discusses some successes, pitfalls, and other implications involved with projects of this kind.
Artificial Intelligence (AI), despite the cautionary tales presented in popular culture, is neither fanciful nor a technology of mere convenience. AI can provide real benefits to businesses seeking to automate operations once handled solely by humans that today can profitably be performed by computers.
"The most successful cases we have had are those where we were able to use AI to augment human beings facing decisions based on external information," notes Paolo Righetti, head of analytics at Teleperformance Knowledge Services (TKS), an independent consulting arm of Teleperformance Group, a customer management services and digital integrated business solutions provider.
One stumbling block Righetti has faced is the fear, uncertainty, and doubt (FUD) many people have with the mere concept of AI.
Overcoming the FUD Factor
Righetti says, "AI-based transformation is more about the enterprise itself than technology. The humanities and social sciences are an important ingredient to be successful when you need to have interaction between AI and human intelligence. You need to understand how to blend these two and how to empower the person with an augmented intelligence, which is the most feeling, powerful, and difficult part of an AI project."
AI projects raise questions that have to be answered and go beyond simply managing change.
"Whenever a project is not just AI but a blending of human and artificial intelligence, you need to have a perspective that is not an AI perspective," Righetti explains. "How do you design an experience? How do you convince human beings to blend their intelligences with a system? How do you not alarm the mind of people who are already doing 10 different things, but you are trying to give them additional information that they might not trust and might not even be in a position to use?"
TKS's methodology is to overcome these factors by insisting on a holistic assessment of the processes and procedures that might provide a reasonable ROI from AI-based transformation, as well as the most thorough inclusion of all client company data that may bear on the activities the project seeks to alter. This process has to take into account not only the needs of the enterprise and the mechanics of adopting processes to full or partial AI execution, but the human and even political factors as well.
Why AI Projects Can Fail
Enterprises trying to undertake such projects on their own are usually ignorant of these factors, and combined with the widespread tendency for corporate departments to silo their information, AI conversion projects provide their own special challenges.
Righetti cites an example of this problem at a European telco that was looking to use AI in a campaign of retaining its customers and tried tackling an AI project on its own. The project started well. The company used a skilled data-science team that was able to identify risk factors that could lead to customer losses, and the team was also able to come up with a substantial list of existing customers that appeared to be at risk because of these factors. Where the project went wrong was that the data was turned over to a campaign sales department that decided the best plan was to contact as many of those customers individually as quickly as possible to minimize costs of the retention effort. Unfortunately, rather than trying to conciliate these customers, the sales department looked at the customer list through its own prism and simply used its traditional "productivity"-based approach rather than design and deploy a proactive "caring" campaign.
"The end result is that they didn't retain anyone, and the entire data science team got dismissed because it was not producing value," Righetti recalls. "Even though the data science team was absolutely state of the art, the big mistake was the lack of holistic coordination. Nobody inside of the company was able to use effectively the output that the data science team produced. When you have identified the customer who is at risk, you need to make a corrective, caring call. You need to identify the issues. But a caring call is not a sales call." Righetti has observed that in past cases, the cost of retaining a customer is about half the cost of attracting a completely new customer. In this company's case, not only was the effort of the data science team counterproductive for the company and the team, abut also there was an additional opportunity cost because the retention effort results were fruitless.
Righetti refers to this kind of problem as a "lack of maturity," although the term applies, in this case, to the telco company's overall lack of understanding about what was needed to use AI to solve their retention problem rather than behavior on the part of any individual. In fact, a study released in June of last year by Pactera Technologies says the failure rate for AI projects is a dismal 85 percent, and studies by Gartner Group in the past have shown that figure has held fairly steady for several years. The Pactera study cited reasons such as senior management not seeing the value in such an investment, although problems cited in other studies included data quality, data classification, and inadequate modeling.
"The vast majority of companies are still in the learning curve. That's one reason for failing projects. We see there are still many companies that have not yet understood how to quickly generate value with AI," Righetti elaborates. "They take figures of costs with a very old-fashioned approach."
The issue of whether or not an AI project adds more value than costs, a traditional business-oriented definition of project failure or success, isn't obvious if a company is evaluating incomplete or misleading data while deciding to start a project in the first place. The lack of maturity can also mean companies don't buy the right AI tools for their situation and they don't really understand how to apply an AI solution to meet their needs, according to Righetti. Retention of information in silos, covered in Part 1 of this article, is another common problem.
"AI is a leveraging in a holistic way of all of the enterprise-wide data available to improve, also enterprise-wide, the full range of front and backend risks and processes to deliver a better customer experience while generating efficiencies. If this approach is not holistic and there are silos, that means we are segmenting data and processes, and so this means you have much less information" to apply to a solution, Righetti says.
Righetti's third big failure cause is a lack of coordination, which was also a factor in the telco's problem.
"There is very limited understanding of how to orchestrate different functions that belong to different departments that are born out of different cultures," Righetti explains. Management consulting, process design, and data science typify different mindsets from different people that need to be coordinated by someone able to work across departments and cultures and make them work together.
"Putting together the business owner with the process person, a manager…even without going into issues of security and compliance and all the other levels of complexity inherent in this process, having those people and the data science team working together is extremely difficult, and that's where the magic happens," he says. "When you're able to orchestrate those teams in a holistic way, that's where you really can transform and generate an incredible amount of value for the company."
Another aspect of lack of maturity is exemplified by unrealistic expectations of what AI can really do for a company.
"Unrealistic expectations could be in terms of how easy it is, how fast you can do it, how much money you can generate, and that's really a matter of lack of experience. This is no simple problem. There is no magic trick. You just need to have enough experience to drive quickly through the assessment," Righetti points out. "That's where we come in, and we generate a lot of value. We build the assessment, we build the business case, we know what is possible. We know how much time it takes. We know what tools you have to have to do it. We know how to orchestrate it in a way that's successful. We are able to set the right expectations in terms of timing."
Mark Simmonds, program director of communications data and AI for IBM, cites some additional common problems.
"Data quality issues is probably one of the biggest reasons, and by that I mean either inaccurate data or missing data," Simmonds points out. "When it comes to machine learning and AI, where you're training a model using data to make a prediction, if that data is incomplete or inaccurate, obviously that's going to affect how the model behaves when it's put into production."
Simmonds also points to issues such as not having the right kind of budget, there not being firm management control of the project timeline in a way that invites scope creep on an implementation project, and the data scientists not understanding the business issues involved.
"You have to have data scientists that have domain expertise in that particular field," Simmonds cautions. "If it's banking transactions, for example, that person needs to have an understanding of the business side of the problem, not just the technical side. It's about putting together the right kind of team for the project, a good project manager, a good business liaison, a good sponsor, domain expertise, and a really good data scientist who can understand the data."
Simmonds thinks the mindset of looking at AI as a solution first and looking for a way to apply it second is also a problem.
"People see AI and machine learning as a sort of panacea. There's so much hype around AI that the hype sets an expectation that it's going to solve everything. If it's done properly though, with the right expertise and understanding of the data, it can do a lot," Simmonds says.
Righetti notes that a final common point of failure is that, in an AI project, clients tend to focus on the change in technology rather than the impact AI can have on an enterprise's ways of doing business.
"Everybody's focusing on AI and not the change that is driven by AI. The AI's costs have the potential to deliver value. The value comes from the change in how you use this AI, not from the AI itself. The fact that everybody talks about the AI biases, the focus, and the expectation," Righetti explains. "AI is a cost; it's not a saving. It's how you use it that can generate efficiencies to create a better customer experience, generate efficiencies and an incremental income value."
Does Company Size Matter in an AI Project?
Even a superficial look at many of the AI successes around the world might give anyone the general impression that only the biggest of companies have the wherewithal to invest in AI or that only companies with the largest customer bases can make an AI project pay off. According to Righetti, that's only partly true.
"Depending on the solution, there are thresholds below which the solution cannot be applied because they have some sort of minimum costs. Below a certain volume of value, customer base, scale, and so on, there is no way you would be able to produce a positive ROI. On a very large scale, it's much easier to get to a point of ROI," Righetti says.
On the other hand, that's not a concept set in stone.
In smaller enterprises or companies where an AI upgrade might be focused on one or two specific processes, "the way you get to positive ROI is by doing something that is a bit difficult. You would need to leverage experienced teams and companies to start this. One of the failures and mistakes of the old approaches is large companies think AI needs to be slow and expensive, which is not true," Righetti points out. "There are a lot of very innovative solutions you can find that can also be for free and you can deploy in a project in weeks."
IBM's Simmonds agrees.
"These days there are lots of AI services in the cloud that can be used at low cost, depending on how much data is involved, so a lot of smaller businesses and people getting started in AI can simply dabble in it and get started for free just to see what AI is like. Organizations can then progress to consume services on a 'pay-as-they-go' basis."
AI Is an Unavoidable Part of Our Future
Simmonds points to ethics as a significant element of AI that needs to be addressed.
"Why was this loan approved and that loan not approved? Organizations have to have transparency into the decision-making made by AI systems. Bias in the data that you're looking at and ethics in terms of to what level an AI system should be able to make a decision is an area that's full of contention."
Righetti thinks the future benefits of AI are unlimited. AI is an essential component of the wider digital transformation trend.
"I think the world in which we live today will become the prehistory once AI is deployed enough," Righetti predicts. "AI will transform the world. Communication, transportation, healthcare, everything will be radically, completely different. The entire world can be optimized with AI. The risk is that AI has a level of intelligence that needs to be managed, controlled, and regulated somehow. It's a new source of energy, the combination of data and AI."
And how do we cope with the fear of change AI embodies? Righetti thinks understanding will cure that over time.
"It's like the danger that [some people saw] coming with the distribution of electricity. Nobody is really scared today by electricity because it's a commodity, it has been regulated, and because we have the understanding of how to deal with it and all the technology to make it safe. AI is at the beginning of this lifecycle. It needs to be understood first and then regulated in a way that it's not damaging. It needs to understood and regulated in a way that is beneficial and not detrimental to the entire world."