The term last-mile problem comes from the telecom industry, which observed that it costs inordinately more to build and manage the last-mile of infrastructure to the home than to bring infrastructure to the hub city or residential perimeter.
Businesses are starting to discover a similar last-mile delivery problem in AI: It is much harder to weave AI technologies into business processes that actually run companies than it is to build or buy the AI and machine learning (ML) models that promise to improve those processes.
“The path to deploying ML is still expensive,” said Ian Xiao, engagement lead at Dessa, an AI consultancy. He estimates that most companies deploy only between 10% and 40% of their machine learning projects depending on their size and technology readiness.
In fact, the last-mile problem is a bit of a misnomer when applied to AI deployment in the enterprise. There is a collection of last-mile delivery problems. In addition to building the digital infrastructure required to integrate AI into business processes, companies are grappling with last-mile issues related to understanding what AI is, empowering users, updating models and even contract management.
The technical and cultural challenges associated with deploying AI in the enterprise are more diffuse and, for some companies, more intractable than the hurdles a vendor might face when integrating AI into its products and services. In the long run, AI’s ability to shine a light on business processes — where a process works well and where it can be improved — will eventually close the gap between the relative ease of buying and building AI and the challenge of deploying it successfully.
Here are seven last-mile delivery problems associated with deploying AI in the enterprise and how to solve them.
The digital gap
According to Xiao, the biggest AI deployment impediment for most companies is indeed the last-mile infrastructure for connecting AI into the business. Robotic process automation (RPA), integration platform as a service (iPaaS) and low-code platforms help close the gap, but these are relatively new technologies. Many companies have not yet adopted them or and may not have the requisite skills on hand to make them work.
Once this kind of infrastructure is in place, however, Xiao expects to see data science teams doing less work on AI and machine learning in the name of innovation and more on integrating AI and machine learning into core operations.
Ramesh Hariharan, head of innovation & technology at LatentView Analytics, an AI consultancy, also expects to see a convergence of complementary technologies to plug the digital gap.
“In the long term, the development of machine learning models is expected to become more automated, thereby enabling much broader adoption of machine learning,” he said.
RPA, for example, will be used more and more to pass on data to ML models in real-time, he said, as well as to generate data to build new ML models. iPaaS can make it easier to deploy AI models that work with applications using a microservices architecture. The adoption of no/low-code could extend the capabilities of citizen data scientists who are adopting tools like Tableau and Jupyter that remove barriers to analytics and machine learning.
Hariharan said the biggest change enterprises will see from AI and the complementary tools that help make it work is in the types of decisions traditionally made by employees. More and more of these workplace decisions are likely to be automated, as machine learning generates more accurate predictions and becomes more trusted by business managers.
The decision-making gap
AI can often deliver better, cheaper and faster predictions than humans in a wide range of settings, but it is up to business managers to figure out what these settings are. Hariharan recommends that managers not look at AI predictions in isolation but instead focus on how to automate the decision-making process. This involves going from simply embedding predictions into existing processes to reengineering the business process. Reengineering might entail eliminating some steps entirely, automating others and bringing all this together through software-driven tools.
A good place to start is to look at a specific department, he said. For example, a marketing AI could accurately predict the potential lifetime value of the customer. To take advantage of this system, the marketing and customer management departments need to reorganize their workflows. They need to integrate these predictions into decisions about which customers to prioritize.
The AI that predicts the lifetime value of a customer can be deployed as a service that is used by other applications. The customer-triaging software that calls these services can be built using low/no-code platforms. The iPaaS can be used to call the AI, as well as other steps in the service. The RPA can be used to provide inputs about the customer to AI in real-time.
The AI value gap
CIOs and business leaders also need to clarify their expectations of how different categories of AI technology provide real business value.
“The biggest challenge to integrating AI into existing business processes is confusion over what AI is really good for, especially in the context of business operations, as opposed to specialized technical and scientific applications,” said Jason Bloomberg, founder and president of Intellyx, a digital transformation analyst firm. For all its power to enhance business processes, AI can’t run a business — its current state falls far short of that kind of general-purpose application of AI.
Bloomberg sees tools like RPA and digital process automation (DPA) providing AI-supported capabilities like next best action functionality, which is essentially an auto-complete for workflows. Businesses are finding value in weaving natural language processing into DPA workflows for use cases involving virtual assistants and chatbots.
Part of the problem in wringing value from AI is that many businesses delegate their AI projects to IT teams who see it through a technology lens rather than a value-delivery perspective. Pat Geary, chief evangelist at Blue Prism, an RPA vendor, said that AI initiatives need to be business-led to succeed.
“While IT and technical leaders are vital for ensuring smooth rollouts, it’s down to business leaders who best understand their companies’ operational challenges and demands to make judgements about where RPA and AI will make the most positive impact on key business outcomes,” he said.
The AI explainability gap
It’s also important to connect the output of the AI to the results business users care about.
“If the AI isn’t adopted all the way down to the end user, it doesn’t create any value,” said Arijit Sengupta, founder & CEO of Aible, an AI platform.
Users typically don’t want to know how predictive models are created or what percentage of model accuracy they exhibit; they do need to know how the AI affects their KPIs and the business metrics they care about, Sengupta said.
“As business assumptions change, as they are now in the current COVID-19 emergency, end users need to know how the AI will help them,” he said.
Unlike LatentView Analytics’ Hariharan, Sengupta believes that most business users will benefit more from directly embedding better models into business applications they use every day, like Salesforce and Tableau, rather than relying on RPA and iPaaS tooling to pass along intelligence. While embedding requires more integration work upfront and can limit users’ options, he said this kind of tight integration can also help give business users a say about what aspects of the model fit their preferences and align with the business realities they see on the ground.
“If you make AI a one-way process, where data scientists create the AI and business users consume it, you’re not solving the problem,” Sengupta argued.
Putting business users in charge of the AI they are using raises the last-mile delivery problem of explainable AI — or AI that users trust because they know how it reached its recommendations.
“If operators don’t trust the machines’ recommendations, they won’t act. If they don’t act, no additional and actual business value is created,” Dessa’s Xiao said. Fixing this issue will likely require deep collaboration between the business operation and technology organizations, and it will not be easy as many machine learning models are not transparent.
The AI feedback gap
Machine learning models have a limited shelf life because the data used to build the models becomes stale over time. Hariharan said the limited shelf life of machine learning models makes it imperative that businesses find ways to automatically update models to reflect changing circumstances based on feedback derived from new data.
Xiao concurred: “Many people think AI/machine learning is a one-way and one-time journey. It’s more like running laps.” A good feedback loop for streaming new business data into the AI model-generation process is an essential component to enable the self-enhancing capability, which is a key differentiating characteristic of AI vs. traditional automation.
The AI process gap
A business process often involves many tasks that are vaguely defined in the heads of the various workers that carry them out. Traditionally, businesses have hired business process experts to analyze these workflows, which can help identify places where AI and automation can add value. Now, AI is being used to automatically make sense of these processes.
Tom Taulli, the author of the new book, The Robotic Process Automation Handbook, said he is seeing RPA software systems use AI to learn from existing processes and automate them. An example is Automation Anywhere’s discovery bot. Taulli also expects to see increased use of process mining to analyze log files to identify bottlenecks. Currently, most of this kind of work is in Europe, but it is starting to move into the U.S.
The AI dark data gap
AI needs a curated source of data to create useful predictions. While traditional data is easy to digest, unstructured data, which according to IBM comprises up to 80% of business processes, has been a challenge.
Prince Kohli, CTO at Automation Anywhere, believes RPA could help automate unstructured data processing using technologies like computer vision, natural language processing and fuzzy logic, negating the need for data scientists to implement robust intelligent automation deployments.
While many RPA platforms now offer AI capabilities, today RPA and AI are often viewed as two separate entities: one is rules-based while the other is adaptive and predictive. But this dichotomy is changing fast.
“In the next year, RPA and process analytics will become entirely infused with AI and machine learning, accelerating process mining and discovery,” Kohli predicted, “dramatically simplifying human effort in these areas and ultimately paving the last mile.”