Advanced Computing in the Age of AI | Thursday, March 28, 2024

Why Businesses Don’t Deploy Machine Learning (And How to Overcome It) 

Imaged credit: Shutterstock

Artificial Intelligence -- including its subset machine learning -- is one of the most transformative technologies to be available to businesses in decades. The emphasis it brings to data insights, the speed of delivery and the creativity that underpins it are radical. It’s a data world and we just live in it.

Yet many businesses haven’t adopted machine learning (ML) on any scale. AI requires enormous and expensive leaps of knowledge about data, which itself is often an under-governed, poorly-structured Achilles’ heel.

Questions about where to deploy ML plus concerns over costs, security, compliance and ethics can stop the journey before it even begins. Let’s review some of these concerns one by one.

Deployment Indecision

When it comes to ML, its power is its incredible flexibility. Well-publicized use cases range from sales forecasting and churn predictions to manufacturing efficiency and fraud detection.

But what if these use cases don’t apply to your business or aren’t considered in need of solving? That’s when ML can suddenly become a solution without a problem.

That’s the wrong conclusion, however. Instead, ML is so flexible it can be used to make improvements anywhere there is data. This puts the onus on businesses to creatively self-identify their own use cases, which can be hard to do when technology is complex, new and multi-faceted. After all, how can you use a tool you don’t fully understand?

Assumptions that Data Analytics Delivers the Same Results

Many business leaders think data analytics is enough. A reliance on data analytics is a proven, additive step for any business to take. But using it pre-supposes you know where to focus its insights.

Data analytics only gives answers to the questions you thought to ask. But the beauty of ML is that it answers questions you haven’t considered.

ML balances a much larger number of factors and datasets against each other. It explores and experiments to find hidden trends and correlations. It translates shortcomings or opportunities in your business. Patterns of data causation or correlation you didn’t know about previously will be infinitely more valuable than the answer to any specific question.

Concerns over Data Quality, Quantity and Type

Too often, businesses blame the impossibility of ML on the data itself. However, longstanding insufficiencies in data governance and controls result in datasets with widespread gaps, inconsistencies, duplications and inaccuracies that are ML's undoing.

Businesses are also often quick to give up because they believe there isn’t enough data to work with – despite that businesses are creating more data than ever. Yes, models need enough data to understand the question and return results with confidence and accuracy, but the amount required depends on the purpose and complexity of the project. Many use cases need less data than you think.

Another concern is whether the data involved is structured or unstructured. Structured data – text, numbers, SQL databases, tables, Excel files – is easier for ML because relationships are clearer. Trends and anomalies stand out. Unstructured data –audio, video or document files – has no format from which to identify patterns. Since 80% of business data is unstructured, many ML projects are abandoned as structuring enough data is often considered too labor intensive.

Lack of Skills and Hefty Price Tags

Studies show that most businesses don’t have the right talent to carry out ML. Recruiting the right skill set is expensive. Whether you need data architecture, data science skills or both, demand is greater than supply.

The cost of the technology is daunting, as well. Not only is software expensive, but ML requires high-specification infrastructure typically built on more powerful and expensive GPU resources versus common CPU-based infrastructures.

Data Safety Concerns

Before deploying an ML project, businesses must think about data safety, including security, privacy, and governance.

Security impacts two areas. First, some of ML’s most popular use cases involve the analysis of customer records. Applications and their data become high value targets for hackers. Secondly, algorithms, like any code, can be rife with risks.

ML also threatens privacy adherence, as GDPR and other privacy regulations guard against data subjects being subject to automated decision-making without consent, which is not always part of privacy policies and consent processes. Finally, data governance, ISO accreditations, and data management and movement policies often constrain the natural experimentation that data science depends on – or are undermined by it.

You also must consider the ethics surrounding the black box problem – can decisions be explained? – plus whether any bias exists in the datasets or in where the model places value. Bias can be a silent killer to the efficacy and acceptability of a project.

Solutions Exist

The obstacles that prevent the deployment of ML seem considerable. But they can be solved. Here is where businesses should start:

Find a use case and set goals. Many businesses overlook constructive or profitable areas for deployment. To pick the right ones you need to:

  • Conduct a strategic review of every process.
  • Assess and score each process’ suitability for improvement considering hard (e.g., revenue, costs) and soft (e.g., cultural, competitive) benefits.
  • Consider the right tools for the right roles. Some processes need automating; others need better decision-making.
  • Perform a data audit to assess what is available and where versus what is needed.

Audit skills.

  • Data privacy expertise to ensure legal and appropriate use of data gathering and how it fits with privacy regulations.
  • Data security knowledge to prevent breaches or compromised datasets.
  • Data architecture capabilities to source, prepare, collate, categorize, cleanse and classify data.
  • Infrastructure expertise to ensure the environment is suitably performant, secure, efficient and capable.
  • Change management experience to adapt processes, operations and roles when deployed.

Assess technology infrastructure. Ensure you have the tools, infrastructure and capacity to build and deploy the project.

Evaluate costs. To address affordability and bring down costs, businesses can access expertise, skills, infrastructure, and technology through a managed service.

It’s easier to make excuses than to make progress. If ML projects are approached correctly with the right partners, businesses can tackle the obstacles and reap the benefits. It is a data world after all. We might as well get to know it better.

 

About the Author

Julian Box is the founder and chief executive officer at managed data services provider Calligo. Box founded Calligo in January 2012 and is responsible for delivering Calligo’s vision of building a client-centric business, delivering managed data services that optimize every stage of the data journey, while ensuring that privacy continues to be at the heart of the organization and its services. He has over 30 years of experience helping organizations streamline operations through the innovative application of technology.

 

EnterpriseAI