Apprenticeship was how people learned a craft for centuries before the current era. It involved a mentor who watched your every move, corrected your mistakes in real time and moved you on to more complex tasks only after you proved you’d mastered the basics.

Apprenticeship holds a cherished place in my family history: My grandfather owned a machine shop that created rubber and metal molds for manufacturing companies. Though highly intelligent, he did not formally study machining or business. Instead, he apprenticed at a machine shop shortly after returning from World War II.

This high-touch, highly effective training modality isn’t a relic of bygone days. Artificial intelligence (AI) and data-driven personalized learning can offer us many of the benefits of traditional apprenticeship.

The 2-Sigma Problem: Scaling the Benefits of One-on-One Tutoring

The apprenticeship model was the holy grail of learning: one-on-one, perfectly personalized and deeply human. However, its main drawback is its lack of scalability due to limited time, budget and availability of experts and learners.

Workplace training was created to meet the need to upskill hundreds, even thousands, of employees quickly and effectively. Doing so required a move from apprenticeship to classroom training — and a sacrifice of relevance and personalization in favor of speed and scale.

This sacrifice is illustrated by the 2-sigma problem, an oft-cited study by Benjamin Bloom, the educator and psychologist behind Bloom’s Taxonomy. In 1984, Bloom set out to measure the effectiveness of classroom instruction against that of one-on-one tutoring. The details of the study are as follows:

• The Experiment: Bloom compared three groups of students:
  • Group 1: This group was taught in a conventional classroom, with about 30 students taught by a single teacher. Student progress was assessed via periodic written tests.
  • Group 2: Students in this group were also taught in a conventional classroom setting via the same instructional style. However, students needed to demonstrate mastery of concepts before moving on. They received the same tests as students in Group 1, but they also received “corrective procedures” and “parallel formative tests” to ascertain mastery.
  • Group 3: This group of students learned the subject matter via one-on-one tutoring with an experienced tutor. Students received periodic feedback via the same formative tests, corrective procedures and parallel formative tests as students in Group 2. (Bloom noted that students’ need for remedial work was quite low.)
• The Result: The students in Group 3 performed two standard deviations (2-sigma) better than the students in the conventional classroom.
• The Impact: Statistically, these results mean that an average student receiving one-on-one tutoring has the potential to perform better than 98% of students in a traditional classroom setting.

The “problem” Bloom identified is that tutoring is too costly and labor-intensive to scale for an entire student population or workforce. He concluded the study with a challenge to the L&D industry: “Find methods of group instruction as effective as one-to-one tutoring.”

eLearning: A Digital Attempt to Solve the 2-Sigma Problem

For the past 40 years, L&D leaders and teams have indeed been seeking a way to solve the 2-sigma problem at scale.

When eLearning emerged, L&D leaders hoped that, by distilling an expert’s wisdom into a course module, we could finally provide a one-on-one transfer of knowledge to every learner. But in our rush to scale content, we lost the most powerful element of apprenticeship: the feedback loop between novice and master. Most eLearning pushes information to the learner without pausing, reacting or adjusting to their specific needs.

Actively observing, reflecting and reasoning are core to how learners gain understanding and skill mastery, and those elements are typically missing in most eLearning experiences — and a quick summary of topics at the end does not constitute true reflection.

Because L&D leaders generally lack the time and budget to personalize an eLearning module for an audience of hundreds or thousands, we often design for the “average” learner. Thus, we herd every learner, regardless of prior experience, expertise or role, through the same linear path.

The primary issue is that digital content alone does not equal a mentor. A mentor knows their learner’s temperament, strengths and weaknesses and recognizes when to remediate, challenge or encourage. Traditional eLearning lacks this insight and empathy, treating every learner the same.

Now, though, AI and its robust data capabilities offer us the opportunity to build a learning ecosystem that adapts and responds to learner needs.

The AI Apprentice Model: A Potential Solution to the 2-Sigma Problem

AI is everywhere, but most people are still using its tangible side, generating content like videos, images, text and slides. When L&D leaders and teams start leveraging the intangible side of AI for mentoring and coaching, we find a resolution to the 2-sigma problem.

Consider two veteran sales professionals taking a strategic negotiation course. The performance need: Negotiate higher-value contracts and secure more favorable terms.

• Alex (the generic path): Alex has 10 years of sales experience but nonetheless is served a mandatory, introductory, 40-minute “click-next” eLearning module. His organization’s L&D team used AI to generate the scenarios and audio. When the course begins by asking, “What is negotiation?”, Alex tunes out. He plays the rest of the module in another tab while responding to emails and passes the final quiz without gaining any new skills.

This learning experience uses the tangible side of AI as a content generator. While this use is a good start, when paired with human oversight, the L&D team’s use of AI isn’t yet mature enough to respond to complex and nuanced pedagogical needs with an adaptive learning experience.

• Sam (the AI apprentice path): Sam is a sales veteran with 15 years of experience. When she opens the module, her AI mentor in the internal large language model (LLM) recognizes her role, experience and past performance data.

Skipping the introductory fluff, her mentor drops her into a simulation of a challenging high-stakes negotiation. When she struggles with a client objection, her AI mentor steps in with a “just-in-time” tip drawn from her organization’s sales playbook.

This learning experience leverages the intangible side of AI to mentor and coach learners. It draws upon an LLM, a knowledge tracing ecosystem (see below for a detailed discussion), and uses data to “decide” which content to offer learners, track their performance, and provide growth-oriented feedback. This technology is mature and ready to scale.

The 3-Point Architecture of the AI Apprentice Model

Though both Sam’s and Alex’s learning experiences leverage AI and are billed as “innovative” by their respective L&D departments, their quality and relevance are worlds apart.

Moving from Alex’s generic path to Sam’s apprentice path doesn’t require more slides or more modules. It requires the following three data pillars:

1. Learner insights: This data includes each learner’s role, years of experience, and even their current performance goals. These insights allow the AI coach, evaluator, simulator, curator, or assimilator (or a combination of these roles) to frame the training in a way that is immediately relatable to the learner’s day-to-day work.
2. Knowledge tracing: Knowledge tracing involves the real-time measurement of a learner’s progress toward mastery for the purpose of making in-course corrections. For example, Sam interacts with the content, and the AI maps what she knows against what she still needs to grasp. This responsiveness enables the “skip-ahead” moments that respect a learner’s time: a critical factor for busy professionals who cannot afford to waste valuable time reviewing the basics.
3. The organizational knowledge base: This centralized repository of your organization’s playbooks, procedures, cultural values and technical frameworks serves as your internal source of truth. Using what’s known as a RAG (retrieval-augmented generation), the LLM draws upon these specific and carefully curated assets, alongside any essential guardrails for the learning experience that ensure that the feedback provided to a learner is aligned with your organization’s values. That protects your people, clients and business from the biases, inaccuracies and outright hallucinations that occur with generic LLMs that draw upon the public internet.

The Role of Human Advocacy in AI Learning

Leveraging AI to its fullest potential doesn’t mean deploying a hands-off, automated bot. As we leverage AI to curate our organization’s internal wisdom, sensitive data and proprietary information, we must continue to serve as advocates for the humans who use it.

That means remembering that AI is often not a neutral technology for many of our people. With privacy concerns and AI fatigue at an all-time high, we must ensure that our learning solutions are transparent about the data that’s tracked and offer the learner the choice to opt out.

That said, we must remember that certain elements, but not all, of human apprenticeship can be scaled. AI can scale coaching and mentoring around certain skills, but it cannot yet understand the bigger picture of a learner’s career or empathize with the emotional toll of personal or professional struggle.

As the role of L&D leader continues to evolve from creator of static content to experience architect, we must lean into our unique position as champions of both our people and our business to ensure that our AI implementation technology remains ethical, transparent and, above all, human.