• Your brand-new AI-powered robotic process automation can read a scanned document and capture 70% of the data on the first pass – an enormous improvement over traditional OCR mechanisms.
   
• The set of images that your data engineers found online has achieved a 65% accuracy rate in the computer vision application just installed in your QA process.
   
• The latest AI speech-to-text tool does a pretty good job – except for surnames, and when there’s background noise, and when two people speak at once, and, of course, when the speaker has an accent.

Navigating the Hype: Making Informed AI Decisions

Sound familiar? It should. While AI-enhanced applications have made enormous strides in the last couple of years, the amount of hype has also increased at roughly the same rate. Too many buying decisions are currently being made based on the ‘magic’ of AI, on pressure from corporate boards, and on good ol’ FOMO. What’s missing is an objective statement of a business problem to be solved; a definition of what success will look like; and a rigorous review of the performance claims of any proposed solution. If you have the discipline to look past the hype and approach an AI implementation as you would with any other technology implementation, you’ll significantly increase the value derived from your investment.

The Human Element in AI Implementation

There’s one fundamental truth to keep squarely in focus: all AI platforms require some level of human oversight, human assistance, and/or human-led training. This isn’t in any way intended to be a knock on the technology; most providers will admit that their platforms were designed to be trained with data that was carefully curated and prepared by humans with specific domain expertise. The best platforms provide user-configurable rules for when human assistance is required. And, human oversight is essential for dealing with data drift that’s inevitable when modeling any kind of dynamic process.

So, a key driver to getting incrementally better results from your AI investment is the presence of people at various points of your process. In some cases, this means skilled individuals to label images or annotate text as the foundational training elements for an AI or ML model. It can mean individuals who compare a model’s current output against a set of expected values to determine if recalibration is necessary. And in the classic ‘human-in-the-loop’ scenario, it means having individuals ready to render assistance to an otherwise highly automated process at important decision points (e.g., when the machine can’t render a decision at a high enough confidence level.) 

Challenges in Acknowledging Human Involvement

At this point, you may be wondering why platform providers typically don’t advise their clients about the need for these human skills during the sales cycle. Or why they don’t offer the services as a value-added component of a complete solution. The answer has nothing to do with technology, and everything to do with psychology. The AI industry is a seller’s market right now:  buyers are caught up in the hype and providers are reluctant to introduce too much reality (that might take the shine off the magic that customers think they are buying.) Customers generally have little experience with AI implementations, and since they’re buying a ‘machine learning’ platform, they believe that the machine will indeed teach itself - and therefore no humans are needed. It's the perfect recipe for a serving of buyer’s remorse.

Addressing Common Pitfalls in AI Implementation

That new automated process designed to read data from scanned images tops out at 70%.  The cameras positioned at the end of a production line to catch manufacturing defects are missing half of the defective items. The next-logical-product recommendation engine hasn’t increased sales at all. These are typical complaints about newly installed AI-driven projects, and they can be addressed (or they could have been avoided) with the right humans in the right places.

The answer is not to throw an army of marginally skilled labor at the issue; nor is it to re-assign valuable data engineering or data science resources to ‘fix’ the problem. But appropriately skilled individuals with data wrangling experience and state-of-the-art tools can transform mediocre accuracy and stale insights into the outcomes you had envisioned. They shouldn’t be viewed as a plus cost but as a vital component of an investment to produce specific revenue gains or expense reductions.  

It’s daunting to question the superiority of anything that includes “AI”– you might even be called a Luddite – but if you persevere, you’ll be rewarded with an advanced technology installation that actually delivers value to your organization – 100% of the time.

To find out how Liberty Source provides end-to-end solutions with our DataInFormation service offerings, contact Joseph Bartolotta, Chief Revenue Officer, at [email protected].