Micro Focus is now part of OpenText. Learn more >

You are here

You are here

Why Isn’t AI Working for Your Business?

Soundarya Jayaraman Writer, G2
Photo by Kevin Ku on Unsplash

More than a third of companies use artificial intelligence (AI), while another 42% are exploring their AI options, according to IBM's recent Global AI Adoption Index. AI adoption looks easy, thanks to rapid advancements in AI technology and the availability of off-the-shelf AI tools.


But in reality, as other recent research makes clear, some companies are struggling with AI. Accenture reports that only 12% of AI adopters are currently using AI "to outpace their competitors," while nearly two-thirds (63%) are still in the experimentation phase—"barely scratching the surface of AI’s potential." Multiple reports show that it is common for AI models to never make it into production. 

So where are companies going wrong? There are four common possibilities.

The Problem of Unrealistic Expectations

A reality sometimes lost in between the fear of AI and the hype of AI (such as with Watson, GPT-3, and AlphaGo) is that AI has its limitations. Business leaders often overlook this. They fail to understand what AI can and can’t do for their business and have unrealistic expectations.


Sure, AI is a powerful new technology. It makes sense of unstructured data and gives insights that humans miss. But it isn’t magic. AI relies on mathematical modeling, and its solutions are often probabilistic. Today, enterprises use narrow AI that does specific tasksnot general AI, the kind in sci-fi movies.


Even IBM, which claimed Watson would be the intelligent digital assistant for everyone everywhere, settled for a humbler version of the tech. The company set out to make sense of all medical data to improve cancer treatmentbut in the end, it failed. Ultimately, the division was discontinued. Former IBM executives noted that their unrealistic AI goals set Watson up for failure because the tech was not advanced enough to meet those goals. This brings us to the first lesson.


Lesson No. 1

AI we use in business is narrow AI technology that supports one particular step—and not an entire process. Enterprises must apply AI to the problems it is suited to solve. Don’t try moonshots without rockets.


Want AI to enhance your sales by automating lead qualification and nurturing? Do it.

Want AI to manage your whole sales operation without humans? Absolutely not.

The Data Problem

Data is central to any AI project. Naturally, then, a data problem can mean an AI problem. 


One data problem plaguing AI projects is poor data accessibility. Enterprise Management Associates recently found that the average organization has between four and six separate data platforms (while finding that some organizations may have as many as a dozen data platforms), making it difficult to navigate the data infrastructure. In its Global AI Adoption Index, IBM reports that 20% of companies don’t have the right tools to find and use their data.


Another AI pitfall is bad data quality. Data-profiling startup Superconductive recently found that 77% of data practitioners report having data quality issues. At the same time, according to Juniper Networks, only 35% of companies consider "standardizing, labeling, and cleaning" their data to be a top priority.


When an AI model is trained on biased or otherwise wrong data, that data will skew the model and produce incorrect results. Amazon and Microsoft have both shut down some of their respective AI projects because of inherent bias replicated from inaccurate training datasets. This brings us to our second lesson.

Lesson No. 2

Ask yourself more than twice: Are you data-ready?


Get your data operations (DataOps) in order. Have a proper data infrastructure to collect, store and organize, and manage data. Make sure it’s accessible across the organization.


And don’t stop with data collection and management. Ensure you have the resources to clean, label, and categorize the data so that your machine-learning (ML) algorithms can read it.

The Problem of Skill

If you don’t have the right people to steer the ship, will you reach your destination? Not having the right data scientists, engineers, AI developers, and business analysts can result in businesses struggling with their AI projects. Even mature AI organizations have trouble filling their AI talent gap.


A recent report from McKinsey described a large financial firm that hired 1,000 data scientistsaveraging $250,000 per year per personto embark on advanced AI analytics projects. Or so the firm thought. The new hires failed to deliver as expected; it turned out that the new hires were not technically data scientists. 


McKinsey estimated that a mere 100 (actual) data scientists, "properly assigned in the right roles in the appropriate organization," would have been plenty.

What can we learn from this?


Lesson No. 3

Once you identify a valuable AI use case, list the AI talents the company needs to work on that case. From data scientists to data engineers to business analysts, your employees all have defined roles. Look for them in-house. If you don’t have any, hire them and give them clear deliverables.

The Problem of Approach

A significant but often overlooked reason why AI projects fail is the approach. Businesses often make the mistake of taking a typical software-development approach to an AI project.


In a traditional software-development project, an organization designs, builds, tests, deploys, and operates software applications. There is certainty throughout the whole process on what the outcome is.


Take, for instance, the waterfall model. It’s a linear approach: design first, deliver at the end. But AI projects don’t work that way. They contend with uncertainty all along the way; delays happen, and problems pop up. Even after a successful proof of concept and a pilot run, an AI model can fail.


AI engineer David Talby built an AI model for predicting readmission in hospitals. But even after successful testing, it failed. The model degraded after deployment because the real-world conditions it operated in changed too fast.


When this happens in a business, the executives face a major dilemma: continue, or stop? The time and cost involved with hiring, data collection, data classification, and data "cleaning"—all in a context of persistent uncertainty—lead executives to pull the plug on AI projects before those projects can deliver.


Lesson No. 4

Change your approach to AI development. The traditional software-development method doesn’t work for AI. Businesses have to adopt a more flexible approach to AI project development. 


Design, build, test, and deploy. If the AI model fails, correct, iterate, and test again. Most importantly, keep your best data scientists on the project even after deployment because actual AI operationalization begins when you have a successful model.

And have patience.

Your AI Road Map

Here are three AI tips to avoid all four of the above problems and implement a successful AI project in your company.


  • Start with a minimum viable AI project. Assess its business value in terms of time, resources, money, and return on investment.
  • Choose between buying and building. You'll need to decide whether you want to partner with an AI service provider or completely build your AI in-house. This decision depends on the business value of your AI use case. Get the right AI engineering team and ensure that it aligns with your business teams and your AI goals.
  • Get your data house in order. This doesn’t necessarily mean getting your entire data system in order. Start with the data you need for your AI use case and scale as you progress.

Design, build, test, and deploy. Implement deadlines, goals, and metrics—understanding that you may face delays and additional costs anyway. Assess the performance and impact of AI faster. Change models when needed. As your business and the project mature, evaluate and undertake more ambitious AI.

Just remember: Opting out is no longer an option.

Keep learning

Read more articles about: Enterprise ITDigital Transformation