Beneath the Surface of AI: Deep Mechanisms from Algorithmic Correction to Recruitment Decisions

Is this person nervous? What is my bank account balance? Where should I return these shoes? In the sales department, who is more worth hiring, Jane or Bob?

Just a few years ago, answering these questions required humans. Today, machines are not only faster, but in most cases, they perform better.

From the phone in your pocket to the newly addedvoice assistantsin corporate boardrooms, artificial intelligence has touched many aspects of personal and professional life. The proliferation of user-facing AI tools has given rise to a group of niche mid-sized companies focused on B2B and B2C applications, which are leveraging this advanced technology to expand their markets.

Ordinary consumers frequently interact with AI-based chatbots and recruiting tools, and many do not realize it. These tools may seem simple and intuitive to users, but the complex technology underlying them is anything but.

Building and correcting products

Bringing a product to market often requires a significant amount of time and data. But what happens when AI needs to forget or be corrected?

According to Josh Sutton, CEO of Agorai, in an interview with CIO Dive, the method for reconfiguring algorithms depends on the type of AI used. Deep learning, a subset of machine learning, is "enhanced pattern matching" that integrates different data points to understand outcomes and answers; symbolic AI, on the other hand, is more based on representing the world in a way that mirrors human understanding.

Correcting symbolic AI systems is relatively straightforward: developers can modify or remove unwanted components. However, errors in deep learning algorithms are similar to bad behavior and require more complex remedies.

Sutton pointed out that just as a golfer who develops a bad swing cannot correct it overnight, AI forgetting bad behavior also requires gradual reinforcement to adapt the "body" or algorithm to behavior patterns different from those it was trained on.

For consumer-facing AI products such as chatbots, there are multiple ways to evaluate and improve AI output.

Kevin Gao, founder and CEO of customer service and communications provider Comm100, told CIO Dive that the company helps clients train deep learning-based bots that interact with customers. Training begins with predicting the questions customers might ask, building a question bank, and training the bot.

The AI provides a confidence score for answers. If the score falls below a threshold, the question is extracted and evaluated, potentially becoming a new question or being merged into an existing one. Gao said human agents can also flag whether an answer was helpful, triggering expert intervention for review.

A well-trained bot starts with quality data, but even in an era generating petabytes of data daily, obtaining high-quality training data remains a challenge. Sutton said data ownership issues and the lack of a comprehensive data marketplace force many companies to piece together solutions to train systems, even if those solutions may contain biases or incorrect information.

This problem is exacerbated by the fact that a few companies, such as Facebook, Google, and Amazon, hold far more data than ordinary enterprises can imagine, using it to maintain market and consumer dominance.

Today, chatbots are advanced enough to anticipate customer needs, respond to questions in different formats, and interact with multimedia. But even the best bots sometimes need human reinforcement.

Gao said that if a customer becomes emotional or triggers specific keywords, the enterprise may want a human agent to take over the conversation—although customers usually do not notice the switch. Chatbots are reliable, but AI is still just a tool to augment human capabilities, not operate independently.

As AI tools become more prevalent, security and privacy are recurring concerns for executives. Jeff Epstein, Vice President of Product Marketing and Communications at Comm100, told CIO Dive that deploying a chatbot on a website is easy, but building security is hard.

He said tools must integrate seamlessly into systems, communicate smoothly, and be easy for enterprises to deploy. Nothing exists in isolation.

For example, a chatbot that connects to a bank and needs to access and retrieve account balance information must establish connections with mission-critical systems that have strong firewalls and APIs. But security cannot come at the expense of functionality; the entire technology network must "work well together within that technology sandbox."

Did a machine hire you?

AI applications are being integrated into more business tools and platforms, but many employees interact with the technology before their first day on the job. AI, machine learning, and deep learning are steadily entering the recruitment process—and their role goes far beyond speeding up resume screening.

Juan Betancourt, CEO of Humantelligence, told CIO Dive that only 30% of predictive success is based on traditional resume information that hiring managers see, such as experience, GPA, education, and references. The top indicator, "EQ" (emotional intelligence), accounts for 60%.

The idea of machines quantifying and assessing the nuances of human personality is unsettling to many. How can a string of binary code assess the subjectivity of the human condition and the complexity of workplace culture better than a real person?

As it turns out, it can do better.

Betancourt said today's digital tools can scientifically measure dozens of traits of an individual's personality, from behavioral characteristics and work styles to motivational factors and value systems. With the right tools, companies can encode their culture and use AI to fill gaps and weaknesses.

Betancourt said AI can help companies examine top-performing employees, extract the traits of their success, and then analyze lower-performing groups to identify areas for development. When bringing in new talent, algorithms help break the pattern of "people hiring those similar to themselves," enhancing diversity and objectivity.

Ankit Somani, co-founder of AllyO, noted in an interview with CIO Dive that AI can also improve recruiting efficiency. For example, hiring engineers involves more human touchpoints than hiring warehouse workers, and companies need to ensure each touchpoint delivers value.

Low-touchpoint positions have greater potential for efficiency gains. If the time at each touchpoint is reduced, job seekers can apply for more positions.

For positions such as waitstaff or sales representatives, hiring typically involves coming to the store, meeting with the manager, and quickly getting a "you're hired" decision. These positions have the highest turnover rates, between 50% and 60%. Betancourt said the widespread introduction of AI-based candidate profiles will have a huge impact on this area by finding the best matches and reducing turnover.

But AI in hiring can be intentionally or unintentionally abused or misused. Somani said the "black box" problem of AI—where inputs go in, decisions come out, and there is no explanation—needs to be addressed. Admittedly, humans often make decisions based on fewer data points than algorithms, but if machines are to narrow down candidate pools, algorithms need to show why certain candidates are selected and others are not.

As the interview process adopts more media, including automated questioning via text, phone, and video, there are also more ways to evaluate candidates. Margaret Olsen, Senior Vice President of Engineering at Cogito, told CIO Dive that, for example, in a conversation, "how" something is said can be as informative as "what" is said. AI applications can identify emotional content in voices, capturing subtle cues that humans might miss.

Testing for empathy and language complexity in voice is still an early tool, but as the market matures, more applications will emerge. However, facial recognition is further off. According to Kurt Heikkinen, President and CEO of Montage, in an interview with CIO Dive, only 6% of candidates said they would be comfortable with the use of this visual tool in interviews.

Heikkinen said the use of advanced analytics such as voice and facial recognition also raises several privacy issues, such as whether job applicants need to be informed in advance that these tools are being used, whether automated decisions could partially determine their screening results, and what data is collected.