4 AI adoption challenges that mask the real value of implementation in the enterprise

February 17, 2021
Partha Pratim Ghosh
Co-founder and Chief Scientist

Beyond the tipping point of AI adoption, where do businesses stand on the matter in 2020? AI enthusiasts have forecasted the development of generalized intelligence by 2050, and the marvels of AI are constantly on our radar as we speak — and not without reason. The AI effect continues to propel the frontiers of capabilities that artificially intelligent systems bring with them — beyond the hype of AI in our news radars, how do businesses fare when it comes to actually solving problems with AI and creating tangible value in the process?

While some business leaders conceive AI as the ultimate solution to their every unknown, others only know enough not to confuse AI with machine learning (ML). Moreover, the AI hype has repeatedly overplayed the value that generalized intelligence solutions deliver through a plug-n-play, API-based implementation. Beyond a few successful, sustainable use cases of AI, most businesses are still struggling with some fundamental AI adoption barriers. Here are four AI adoption challenges that impede the extraction of meaningful business value from AI implementations:

1. Lack of production-grade data

While data science explosion is a consequence of the exponential expansion of data footprints across business functions and beyond, nearly none of this data can be fed to data-intensive machine learning and deep learning algorithms in a business use case. Machine learning algorithms require high-quality data in huge quantities to attain accuracy levels of significant business value.

Deep learning models are especially susceptible to ordinary data quality problems like missing class labels in training data or class imbalance — which can significantly affect their performance in production environments.

2. Sustainability of implementations and pilots

While the implementation of AI solutions is composed of ground-level challenges like lack of skilled talent for practical implementations — sustainability of these implementations is a question that bugs CIOs after successful deployments. Most implementations fail to address data heterogeneity problems beyond theoretical considerations — and as a result, a lack of stress testing for production scenarios can lead to sunk costs.

Moreover, maintaining the use cases beyond deployments calls for oversight from skilled talent — which is expensive, and erodes the competitive advantage that a successful solution was conceived to bring. To make matters worse, unattended AI implementations can lead to AI disasters — Google’s image tagging algorithms misclassifying pictures of African-American subjects as ‘gorillas’ is a leading example that haunts complex models in the industry till date. As a result, staying ahead of the bias implicit in the training data requires built-in self-correction mechanisms or constant tweaks from diverse teams to avert costly disasters of black-box implementations.

3. Where is the data, anyway?

While the availability of data grows with the size of the enterprise, so do the challenges associated with leveraging this data to reach the training stage. While the famous 85–15 split between the time taken for pre-processing and the rest of the owl is no exaggeration, AI adoption in mid to large-sized enterprises suffers from a fragmented, unmeasured and variable data footprint across processes, geographies and product lines.

Before successful warehousing, implementation partners or in-house teams must develop connectors that comprehend varying file formats and integrators (APIs) that can collate data from a variety of systems and sources. As a result, solutions tend to be contextual rather than universal, and encounter unforeseen challenges like keeping connectors and authentication variables intact and updated within the implementation.

4. The cost of implementing and maintaining standards

AI use cases are increasingly becoming the subject of legal, geopolitical and ethical scrutiny, and for good reasons. However, imbibing the challenges of achieving the best standards that are in compliance with privacy, ethical, geopolitical and compliance standards of the timeframe in question can be a moving target that is difficult to hit without a dedicated cross-functional team that overlooks AI-based solutions.

Geopolitical factors like evolving cross-border relations, data residency requirements and introduction of new policies on a subject which finds itself in legal infancy in most judicial and regulatory frameworks can affect existing implementations without affording a reaction time for business leaders. Use cases which overlap with safety-critical socio-economic components — for example, energy, healthcare, pharmaceutical, banking, network management, and defense — call for additional hygiene standards that build adequate safety and data privacy mechanisms into the implementation.

So where is the real, and more importantly, sustainable value of AI for enterprises? The first step is to recognize and understand what AI can or cannot do in a business scenario. Second, CIOs should also evaluate the actual costs that come with preventing AI implementations from becoming black-box solutions that cannot be touched, tweaked or steered by anyone within the organization. Third, businesses should invest in understanding the AI market before jumping the gun — in fact, generalized solutions that expose intelligence through APIs rarely deliver meaningful and scalable advantage beyond a generic function.

What enterprises need is AI solutions that recognize the true currency of business value. Here are four dimensions along which such a solution should be evaluated to keep expectations grounded:

 Context-awareness: Does the solution work with claimed efficiency on real-world datasets? If it claims to use ML techniques, can it achieve the quoted levels of accuracy and KPI improvements with the quality and quantity of available data?

Sustainable advantage: What are the costs of preprocessing and improving and maintaining it post-implementation? Will any costs incurred for the same subtract from the business value and competitive advantage that it claims to deliver?

Implementation readiness: Does the solution require expensive talent to extract business value? Or does it carry the necessary connectors and APIs to reach the boundaries of your enterprise’s data footprint? Can it internalize data quality issues, and does it carry self-correcting mechanisms for handling bias and inaccuracy?

Hygiene standards: Does the solution internalize data residency, privacy and security concerns, or will you need to invest resources for maintaining these standards? Does it show awareness of environment-specific regulatory and policy discourse?

Enterprise AI that delivers value in the long run will be the CIOs answer to these questions. In a rapidly evolving technological ecosystem, the implementation environment is no longer a constant — at the same time, there are ever-newer constants to be accounted for. Time for an AI solution that actually solves these problems rather than creating them. Experience how AI purpose-built for business can solve these big problems with elegance — explore Sainapse today.

About the author
Partha Pratim Ghosh
Co-founder and Chief Scientist