An image of the businessman standing, overlaid with chess gpieces and white copy space - Strategic planning concept
PHOTO: Shutterstock

A mortgage lender enables a new customer to onboard for a loan from his mobile device. An insurance customer submits and manages a claim via her smart device with swipes and clicks instead of manually inputting data. A transportation carrier manages all its invoices, bills of lading and customs documents digitally, transforming that data into a single pane of glass to optimize routes, inventory, schedules and closing out loads. All of these examples were the result of artificial intelligence (AI) projects that went right, said Anthony Macciola, chief innovation officer for ABBYY

Content was digitized, and unstructured data was transformed into structured actionable information and automated into various business processes. Then, a level of learned intelligence was added. It all went off without a hitch, more or less.

But sometimes, maybe often, an AI project doesn’t deliver as originally conceptualized. Then you have angry customers whose apps are not working as they should. You have angry colleagues that were expecting results from your AI project and aren’t getting them. And you have an angry boss.

The best way to avoid all of that is to have a plan in place for the project from the very beginning. While this may seem obviously intuitive, in truth many AI projects start out without a discernible goal — the cornerstone of any plan. And even if there is a goal, too often there are misunderstandings about other aspects of a project. “Do you have a business problem you are trying to solve? Do you have a use case in mind? Can you articulate it clearly? [If not], you’re likely to struggle with project scope, metrics and any definition of success,” said Indico CEO Tom Wilde. For a plan to succeed make sure it has the following elements.

Related Article: What Is Explainable AI (XAI)?

A Solid Use Case

A detailed use case is essential for an AI project. “Many organizations approach AI with the notion that it will tell them what the right answer is inside a large pool of data,” Wilde said. In reality, AI is great at discovering what maps to or matches an already defined desired state — for example, if it is shown what compliant contract language looks like, AI can then automate the process of identifying which contracts are compliant and which are not. “If you can’t define the desired state, don’t expect AI to do it for you,” Wilde said.

The Right Data

“The golden rule of AI is that 90 percent of the work in any AI project is in getting the data,” said Paul Brown, co-founder and CPO of Koverse. “Data is fuel for the algorithms, so you need all the necessary data in a timely manner in one place for processing.” Also remember, he added, the data technology must support requirements for performance, security and scale. 

Another tip, from TIBCO’s chief analytics officer Michael O’Connell, is that big data or more data is not necessarily better. “There are two different types of data: data that represents the business problem and data that doesn’t. The former is essential.”

Related Article: How Artificial Intelligence Will Impact the Future of Work

How it Will Impact Others

When building an AI strategy, a common mistake is to overlook or not think through the implications of the project within the organization and for customers, according to Kapil Kanugo, head of AI and advanced technologies at Accenture Technology. “AI needs to be designed in a responsible, ethical and explainable manner so that there is proper governance with stakeholders and product owners. A virtual agent, for example, may interact with your end customers and branding needs to take a close look as to how the voice, personality and resolution impact their brand. Failure to account for such factors may degrade brand influence and brand affinity.”

Company roles within an organization also need to be taken into account. For example, many AI projects focus on automating an existing business process, Wilde said. “And that’s a good start. But it’s amazing how different people in an organization view how a specific business process works. It’s important to start with a common understanding of the different steps in a process before you can apply AI to it.”

The Training Period

The training period is the most expensive process in terms of computation, human resources and time, noted Somdip Dey, embedded AI scientist at the University of Essex and a research scientist at the Samsung R&D Institute. It can take hours to days to weeks for the training process to complete based on the chosen application and dataset on which the work is being pursued, he said, and if there is any issue during the training process, the development team has to start the training from the beginning. “This leads to lost hours, which translates to money spent on nothing. The most common mistakes in a training process are either choosing the wrong model for the target application or choosing the wrong features/labels during the training methodology,” Dey said.

The Deployment

Another tricky area for planning is the deployment, according to Naren Thiagarajan, co-founder of FloydHub. “After a model is trained, it needs to be deployed to be usable. There are multiple ways to do this — for IoT devices it should run on the edge, for mobile apps it can be hosted as an API. It is critical to figure this out early on and plan on how the models will be updated as the team continues to improve it.”

The Right People

Proper planning includes making sure that the right people are available for the project, according to Rosaria Silipo, principal data scientist at KNIME. “The time required to find data scientists and data engineers with the right expertise should not be underestimated,” she said. “Hiring time is also not independent from the choice of tools. It might be easier to find the right expertise for some of the tools rather than for others.”