Why Most Generative AI Investments Fail (and How to Fix It)

The following is a guest article by Andy De, Chief Marketing Officer (CMO) at Lightbeam Health Solutions

A recent MIT report, The Gen AI Divide: State of AI in Business 2025, suggested that 95% of all generative AI projects fail. This means only one in 20 ever succeeds. Given the $30 billion to $40 billion in Gen AI investments, that’s a sobering failure rate that represents not just wasted capital, but also lost time, missed opportunities, and growing skepticism of AI.

According to this report, 80% of organizations have made investments in Gen AI purchasing tools like Copilot, ChatGPT, or Claude. However, only 40% have deployed them, with 20% reaching a pilot phase. Only 5% have made it to full production. Scaling an AI pilot program has presented significant challenges. Only technology, media, and telecom organizations have shown meaningful structural change, while other industries are still mostly experimenting.

Why do these investments fail? Gen AI and LLMs are held back by five fundamental limitations that leaders too often ignore in the heat of FOMO-driven adoption.

1. Hallucinations: Gen AI is known to spew inaccurate and dubious outputs, given the lack of curated and reliable sources on the Internet; can you afford to have your Gen AI-driven use case hallucinate in a healthcare context where lives are involved — no
2. AI Bias: LLMs often over-rely on certain sources of information, agnostic to their credibility or balance; this creates dangerous real-world risks, especially when pertaining to race, color, gender, and sexual orientation, which are not acceptable
3. Non-Determinism: Gen AI/ LLMs can produce different outputs when given the same input multiple times, leading to unpredictability and inconsistency in its results; that is simply not acceptable in a clinical context and constrains applications and use cases for clinical diagnosis and decision support
4. Security Issues and AI Manipulation/Hacks: Text generation models, like LLMs, function on broken-down bits of language (also known as tokens) and their statistical correlations – in other words, they technically cannot process language like humans do; they cannot fully detect the shift in user intents between neutral prompt A and malicious prompt B; this can be triggered by manipulating the prompt, often referred to as “Prompt Injection,” the prompt injection technique exploits this limitation of AI, similar to phishing with humans, the perennial nightmare of IT teams everywhere
5. Limited Use Cases and Applications: Despite their impressive output capabilities, Gen AI applications are limited in their ability to tackle complex, multidimensional societal issues; they excel in defined, narrow tasks but lack the general understanding needed to address broader challenges, such as strategic decision-making or ethical dilemmas; this highlights a significant gap between the capabilities of current Gen AI technologies and the requirements for solving real-world problems that involve high stakes or deep contextual understanding

Given this reality, Gen AI/LLMs have found relatively modest usage in healthcare. The one bright spot is Ambient Clinical Intelligence, or Ambient Listening. Gen AI is being subsumed by the next wave as we speak: Agentic AI and AI Agents, which offer far more impact and scalability.

Agentic AI and AI Agents, and Differences with Gen AI/LLMs

Traditional modalities of AI, including machine learning, NLP/NLG, deep learning, and Gen AI/LLMs, predict or classify data to deliver static responses to queries from users.

Agentic AI/AI Agents, in contrast, can autonomously plan, adapt, and act to complete tasks, workflows, and processes using APIs, files, browsers, or other tools with little to no human involvement.

Figure 1: Difference Between Gen AI/LLMs, Agentic AI, and AI Agents

Breaking it down further: AI agents are autonomous software programs. Using deep learning, they understand their environments, process information, and act as designed on everything from routine tasks to highly complex operations.

Agentic AI is the autonomous operating system (OS) that enables one or multiple agents to work together and accomplish complex tasks, workflows, or processes. In other words, it’s the underlying intelligence.

Agentic AI/AI Agents in Population Health Management

AI automation is a true “game changer” for healthcare. By offloading manual, repetitive, and error-prone tasks it frees up staff to focus more on patient care. Here’s what this can look like in your healthcare organization:

Figure 2: Agentic AI and AI Agents for Population Health Management

AI Agent #1: Conversational AI Agent for Transitions in Care Enablement

• Automates follow-up to recently discharged patients and helps them with transitions of care
• Impact: Reduce readmissions and staff workload while improving patient satisfaction

AI Agent #2: Conversational AI Care Management Enrollment Agent

• Uses conversational AI and interactive voice response (IVR) to call high-risk and rising-risk patients, capture their information, and enroll them into a care management or remote patient monitoring program
• Impact: Improves outcomes and reduces costs by enrolling high-risk and rising-risk patients into care programs faster without overburdening staff

AI Agent #3: Patient Referral Agent

• Automates referrals and ensures patients can easily connect with their specialists of choice, schedule appointments, and send reminders and follow-ups
• Impact: Reduces referral leakage, closes care gaps, and improves patient outcomes

AI Agent #4: Care Gap Closure Agent

• Enables automated outreach to patients to schedule their annual wellness visits
• Impact: Drives higher quality scores, stronger financial performance, and better preventive care

By deploying a care continuum AI Agent suite, organizations can create a comprehensive automation layer from hospital discharge to preventive care. This unlocks measurable improvements in quality, efficiency, and outcomes, including:

• Increasing call center agent capacity with no additional FTEs in the current fiscal year
• Boosting clinician, nurse, and care manager productivity and capacity without increasing headcount
• Reducing avoidable ED admissions and readmissions, saving millions in costs
• Maximizing reimbursements under value-based care contracts
• Lowering fatigue and burnout among clinicians, nurses, and care managers, strengthening retention
• Improving patient satisfaction and loyalty while lowering the risk of unplanned acute events

This is not just a future concept; it’s a model whose time has come. Turn automation into a force multiplier for your clinical and operational teams. For providers, physician practices, ACOs, and MSOs, adoption is no longer optional; it’s how you’ll survive and thrive in value-based care.