Practical AI: or why everything that says it is, isn’t

Complex algorithms do not an intelligent program make, yet, true AI can be a very involved undertaking fraught with high overheads, a substantial investment in IT support, and an ongoing need to monitor and manage the training that an AI system requires

Artificial intelligence computer brain circuits electronics grid

Most IT professionals have reached the point where when a vendor says the words “artificial intelligence,” they duck and cover. It seems that every application being offered is, in some sense, leveraging artificial intelligence. Of course, just because an application purports to be AI doesn’t mean that it is. Complex algorithms don’t necessarily confer intelligence.

In a recent webinar that I conducted, one of the questions was “What’s the difference between advanced analytics and AI?” While this is not an academically approved answer, analytics generally are static snapshots of large data sets that typically use statistical correlations to extract relationships from the data. AI, on the other hand, is more like a motion picture, where near-real-time and archival data is analyzed for meaning in a dynamic and evolving way. The beauty of AI is that once it is trained, it can evolve to incorporate data it has never encountered.

What all this means is that AI is much more flexible and potentially much more powerful than business intelligence software. As noted in a previous blog here, AI is currently being applied to extract meaning from massive data lakes that conventional business intelligence applications would not even consider. However, for all the power of AI, it suffers from some pretty major deficiencies.

Street of dreams, feet of clay

The problem with AI is that it generally isn’t simply a capability, it is a process that demands significant support to make work well. These include:

  • The need for clean data: While it has been represented elsewhere that AI is perfectly happy with data that has not been cleansed, the fact is, that if you are looking for any sense to come out of an AI, you must start with clean data. A previous blog here explained the problems associated with data preparation, so I won’t dwell on it. Suffice it to say that if early computing coined the term “garbage in, garbage out” to describe the perils of automated data processing, then AI amplifies that notion several orders of magnitude.
  • The need for customization: AI-enabled systems usually do not work straight out of the box. They require customization to match the existing computing environment and to deal with the particular data sets of the company. This can take a while and is often done by the solution vendor. For those who want to roll their own—write their own applications using available cloud-based APIs—the customization process can stretch to infinity, depending on the skill of the IT organization.
  • The need for training: Most AI applications are task-specific. One way to think of it is that if you implement AI to cook pancakes, heaven help you if you ask for an omelet! The fact is, that AI requires training and, as things change over time, may require retraining. Of course, this can be done automatically, but if the AI operates across multiple computing domains—imagine an AI that assesses accounting data, market data, and production data—then if an additional domain is added to the mix, the AI may start delivering flawed results. Someone needs to monitor the inputs as well as the outputs, and this creates overheads.

AI can be a valuable tool, but you can’t just plug and play: This is one dog that you don’t want to have running wild. Of course, all of this begs the question “Is there an easier way to do AI?” The answer is yes.

AI that just works

I have defined practical AI as AI that is embedded in a tool set that you already know how to use: one that requires no additional training before use and that delivers value out of the box. This means that, while the AI is still learning, it comes prepared with access to cloud-based data that enables it to hit the ground running, so to speak. As far as the user is concerned it is something that he or she is familiar with—CRM applications, for example—and that just works better all of a sudden.

In fact, the CRM example is telling, because a company I interviewed a while back is set to change the CRM paradigm with a series of applications that embed AI into the sales process to quickly identify prospects from a client database, qualify those prospects, and then provide the essential information that a sales representative needs to close sales more effectively.’s portfolio of solutions is powered by Neuralytics, the AI engine that combines big data, predictive analytics, and AI. Yet, this AI is tame: It comes pretrained for the sales environment and literally delivers value out of the box., of course, is not the only practical AI application in the market, but it is representative of the class. 

What’s practical?

How do you know beforehand if an application is “practical”? Practical AI can be hard to spot. Some applications that claim to be AI, aren’t, as noted above. Yet there are several telling characteristics that you can use to cut through the fog:

  • Practical AI, does not require a great deal of customization: The question to ask is “Can I use this application out of the box or will I need to budget time and personnel to configure it to work with my systems?” While every program requires configuration, practical AI will require only rudimentary configuration when implementing: generally an identification of computing resources and databases that can be accessed.
  • Practical AI does not require pretraining: All machine learning applications require training, but practical AI comes pretrained for the intended application and discovers additional rules as it is used in its intended purpose. Many practical AI applications can also access a dynamic set of rules that are maintained by the vendor in the cloud.
  • Practical AI is invisible to the user: Rather than dealing with arcane new conventions that require user training, practical AI applications look and work like conventional ones, but they work better. The user simply sees better performance.

Questions that you should ask when considering an AI-enabled application are:

  • “How long will it take to set up?”
  • “How long before I can trust the results of the system?”
  • “How much support and training are required to maintain and use it?”

If the answers seem squirrelly, the application isn’t practical.

How I learned to stop worrying and love the AI

While AI can mean lifetime employment for data scientists and indebtedness beyond your wildest dreams, it doesn’t have to be that way. AI done poorly, like any IT application, can consume resources and fail to deliver any real value. However, AI amplifies this notion because AI is typically applied to company critical computing. A failure can rise to the level of company viability threatening.

Practical AI can, on the other hand, operate like any other application, albeit much more efficiently and with a much greater positive impact on the company value chain. While what’s under the hood can be very complex, the user can blithely pursue company objectives using tools that simply work better than the ones they replaced.

The bottom line is that AI is a quantum leap in terms of capability, but only if care is used in its adoption. Enterprises considering AI should focus less on developing their own systems and instead seek practical AI alternative in the marketplace.

Copyright © 2018 IDG Communications, Inc.