The New AI Workforce: How Domain Experts Are Training NexGen AI Models

We've reached an inflection point in the enterprise-level application of generative artificial intelligence (AI)—The Great Reassessment. While the capabilities of generative AI models have amazed us, I'm seeing companies discover that these broad solutions often fall short when applied to specific business needs in particular sectors—in the hype cycle of AI, we’ve definitively entered the Trough of Disillusionment. The reality is simple: what generative AI currently offers costs too much and does too little. The future of AI lies not in these general-application models but in specialized vertical solutions powered by Expert Intelligence (EI).

‍

Verticalized AI models are a natural progression in the story of AI. Let me catch you up: early research evolved and merged into large language models (LLM), which were built to “predict” text in a given sequence based on the context of a prompt. With GPT-4, we saw a human element introduced called Reinforcement Learning from Human Feedback (RLHF): the LLM model gets rewarded based on humans' real-time scoring or grading of its responses. We are moving up the “cognitive stack” of expertise—while anyone can tell an algorithm that a stop sign means stop, not everyone can tell it what coding solution is more elegant and less buggy. We’ve gone from zero human feedback to human feedback to realizing the necessity of expert human feedback.

‍

This happens in part through a process called alignment. We’ll discuss this further below, but for now, think of it as the process by which you’d teach an AI algorithm what to prioritize and what your expectations are. For example, how do you teach AI to drive a vehicle safely?

‍

The truth is that these generalized LLM solutions (GPT-4, etc) can do an okay job for most applications … about 80% of the time. However, the internet-scraped data (YouTube transcripts have been a major resource for some major models) that powers these models, even combined with RLHF, can only take us so far. What about the other 20%? It depends on the results you expect, but hoping that a general LLM solution will fundamentally transform your organization without expertise in the mix? That’s not a recipe for profound success.

‍

‍

‍

For companies who do want to realize the “pot of gold” promise that the AI hype cycle has created—and are willing to spend time and resources to develop a strategy—what’s the next best step?

‍

To understand this, I’m going to cover a few critical points in this piece:

‍

• Why general AI models are starting to hit a plateau 
• The evolution of AI training processes
• How companies like Intuit get it right with an expert intelligence approach
• The role of the alignment process
• Future implications for enterprise AI solutions

General AI’s Plateau & The Rise of Alignment

AI has hit a developmental plateau, and one clear factor is the data—these algorithms consume largely public data, and diminishing returns are the result. The reasons for this include computational limitations and algorithmic boundaries, but, in my view, the key factor is the data. A general model fed on a “high carb” diet of general information cannot be expected to perform with the nuance of high-demand applications.

This coincides with the reported high error rates for general AI applications caused by a lack of real-time information, limited contextual information, and training data quality. The results are concerning: for instance, OpenAI's Whisper created false text in 80% of public meeting transcripts. As more sectors seek to leverage AI for specific applications—like coding, medical analysis, and legal documentation—what’s become clear is that generative models trained on internet data struggle when applied to specific, mission-critical use cases. Expert people are critical and essential to any high-functioning, domain-specific AI application.

This is where AI alignment comes into play: the process of ensuring AI systems follow human values and goals. It's not just about making models more accurate; it's about making them more helpful, truthful, and reliable in ways that match human expectations. Consider the example of self-driving cars. Alignment, in this case, is the process of teaching the algorithm human expectations, including a hierarchy of goals with human safety at the top of the list. The process then uses a combination of scenario-based training and human input to learn how to handle routine and complex situations—including situations where an accident is unavoidable. Part of the process also includes explanations, on the algorithm’s part, for its decisions so that human auditors can continually refine its decision-making process. Without alignment, you’d never come close to an autonomous vehicle anyone would step into. 

Alignment (noun): “The process of encoding human values and goals into large language models to make them as helpful, safe, and reliable as possible. Through alignment, enterprises can tailor AI models to follow their business rules and policies.”

—IBM Research

‍

For AI models to be more helpful and efficient, we need a different mindset and approach to developing them. This isn't just my observation—we're seeing vertical AI solutions grow 400% year-over-year as companies recognize this reality. Through alignment, enterprises can tailor AI models to follow specific business rules and policies, effectively encoding their unique values and expertise into the system.

The challenge is that generative, general models require significant investment but struggle to demonstrate measurable ROI. The problem isn't with AI technology but rather with how we approach its implementation. The solution is to move beyond basic AI implementations to solutions that incorporate deep domain expertise—authentic, human expertise. This requires a fundamental shift in how we think about AI development and deployment, one that combines technical capabilities with human values and domain-specific knowledge.

You already understand the outcome you want from AI: you want it to do the job you ask it to do with all the nuances of your specific work area. Next, it’s time to build a system, and in this case, the system has two crucial elements: expert intelligence and data collection.

Data Labeling and Domain Expertise: The Foundation of AI Training

Expert intelligence (EI) is crucial for developing effective vertical AI solutions. In my experience, expertise can lie within your company or outside it in your ecosystem of partners, vendors, or community members—like how Salesforce leverages its extensive user community.

‍

Consider Intuit (TurboTax’s parent company) as a prime example: “Our strategy is to be a global AI-driven expert platform.” They don't have a ton of CPAs inside the company, yet they rely on this ecosystem of CPAs that they service because those are also their customers. Rather than trying to build AI systems solely on internal knowledge, they've created a structure with expertise as a fundamental aspect. This approach has proven remarkably successful, with Intuit projecting $1.4 billion in revenue for FY2024 from their expert-based services alone. Google, Anthropic, and others are utilizing similar strategies by hiring experts from companies like Turing to train their AI’s programming capabilities.

‍

Human involvement has worked behind the scenes to train AI from the get-go; we see this in the example of data annotation teams, the workforce that trains general models. However, vertical-specific systems require much more expertise, and companies like Turing are materializing in response. The key insight here is that expertise can be outside your organization. What matters is creating systems that can effectively capture and utilize expert knowledge, regardless of its source. This is particularly important as we move up the cognitive stack, where more than basic labeling and pattern recognition are needed.

‍

Car First, Engine Second: The Critical Step of Data Collection

One of the most common mistakes I see organizations make is jumping straight to model development without first establishing the proper foundation. It's like building an engine with no car. Before you even think about AI models, you need to understand the jobs to be done and create systems that can effectively capture and utilize expert knowledge.

‍

Here's what this means in practice:

‍

1. First, establish desired outcomes and how the system should work to support them
2. Then, design processes for collecting expert knowledge
3. Create feedback loops for continuous learning
4. Finally, implement data collection and model training

‍

Let me give you a concrete example from healthcare. Imagine you're running a hospital and want to train an AI model to provide aftercare tips for cardiac patients. Before considering the AI model, you need to create systems for capturing expert knowledge from doctors and nurses, establish feedback loops, and ensure the entire process aligns with medical best practices.

‍

The key is digitizing your processes and understanding the jobs to be done before implementing AI. This approach ensures that when AI is implemented, it enhances rather than disrupts existing workflows. But you have to build a system that first captures that data as it's produced.

Sculpting a Model: The Role of Fine-Tuning in Aligning Vertical AI

You understand your outcome, and you have a system in place (in combination with expert intelligence) to collect that data—fine-tuning (a step in the alignment process) is the momentum you need to keep the process spinning.

‍

The broader alignment process has evolved from Asimov's simple "do no harm" principles to something much more nuanced. Alignment typically happens in two stages, starting with fine-tuning, where the model is given examples of the process it’s being trained for—summarizing health charts, for instance—and learning to ask contextual questions.

The next step is the critique phase: the model’s responses are graded in real-time (by either a human or another AI or LLM model). In this case, it’s critical that the right parts of the chart are summarized, and accuracy is non-negotiable. Desirable responses are typically fed back into the reinforcement learning model via a human (RLHF) or AI (RELAIF).

Traditional data annotation and alignment are binary and categorical and have relied heavily on non-expert laborers, with projects requiring up to one petabyte of labeled data per LLM. Vertical AI demands a more sophisticated approach. It’s up to the data annotation team to collect, clean, and feed the data to a given model. They also determine whether the model’s output is correct. For example, if the AI is shown a picture of a Stop sign but thinks it’s a Yield sign, the annotator has discovered a critical error. 

‍

Expert-driven alignment differs in that it’s subjective and requires the annotator (the expert) to have deep domain knowledge from the get-go. For example, an expert might be tasked with giving feedback to the AI algorithm about a block of software code—out of several potential coding solutions, which is the best one (that the algorithm should lean toward for future responses)? To distill this further, if you want an AI algorithm that works like an expert, you must train it with experts.

‍

Source: The Information - Creating a vertical AI solution involves a nuanced “sculpting” process.

‍‍

Beyond Code: How Medical Experts Are Training Tomorrow's AI

‍

Let me walk you through how this works in practice using our healthcare example. In this scenario, let’s say I’m talking to a hospital CEO about implementing AI for post-operative care. First, I emphasize that the general models are wrong 80% of the time; we need a more sophisticated approach.

Based on my experience and drawing from best practices across the industry, here's how to implement vertical AI solutions effectively:

‍

• Initial Assessment and System Design: More concretely, imagine you're running a cardiac wing at a hospital and want to use AI for post-operative care. Before you even touch an AI model, you need to understand what success looks like in your specific context. In healthcare, you can't afford to get this wrong. So first, we establish how the system should work: what data do your doctors and nurses need to capture? How will patients interact with the system? What triggers escalation to a human expert?

‍

• Building the Data Foundation: Once you understand the workflow, you need to create digital systems that capture the right information at the right time. You need to integrate seamlessly with how your medical staff already works, collecting their expertise naturally through their normal workflow. This isn't just about gathering data—it's about gathering the right data in a way that doesn't disrupt your experts' primary job of patient care.

‍

• Expert Integration and Feedback Loops This is where the Intuit model really shines as an example. Just like Intuit doesn't have all their CPAs in-house but rather creates an ecosystem of expertise, a hospital can create systems that capture knowledge from their network of healthcare providers. Every time a nurse or doctor interacts with a patient post-surgery, that interaction becomes valuable training data. The genius of this approach is that you're not just solving the immediate patient need—you're continuously improving your system through real expert interactions.

‍

• Continuous Learning Systems Here's what makes this whole thing work: automation with expert oversight. You need batch processing that takes all the expert knowledge collected and feeds it back into the model periodically. But—and this is crucial—you always keep a human in the loop. Similar to how Intuit connects customers with CPAs when needed, your system should be one click away from connecting a patient with a telehealth provider. The goal isn't to replace your healthcare providers but to reduce their caseload while maintaining quality care.

‍

What makes this approach successful is maintaining a balance between automation and expert oversight. I don't ever believe, even in healthcare, that the AI model alone will get you there. Instead, the goal is to reduce the caseload while maintaining quality through expert supervision.

‍

The Future Implications of Vertical AI

We've entered AI's Trough of Disillusionment, and it's become clear that deploying general LLMs across different verticals isn't enough to unlock AI's game-changing promises. The future lies in combining artificial intelligence with deep domain expertise, which requires a fundamental shift in how we approach AI implementation. 

‍

This is an exciting time, as we're moving into a new phase of AI development, particularly with the emergence of the "agentic layer"—systems that can autonomously interact with different models and approaches to find optimal solutions. (I’ll get into this in a subsequent article.)

‍

As vertical AI continues to evolve, I expect to see:

• More sophisticated systems for capturing and utilizing expert knowledge
• Increased focus on domain-specific AI solutions
• Greater emphasis on continuous learning and improvement
• Evolution of hybrid human-AI systems that leverage the best of both

‍

The path forward requires understanding that expertise can come from various sources—internal teams or external networks like Intuit's ecosystem of CPAs. But more importantly, it requires building systems that can effectively capture and utilize this expertise before we even consider the AI models themselves. As we've seen with Intuit's success, this approach can drive significant business value.

‍

The organizations that will thrive in this new era will not simply implement AI but master the art of combining machine intelligence with human expertise. Through proper alignment and fine-tuning, we can create systems that don't just process information but truly understand and operate within their specific domains. That's where the real transformation begins.

‍