AI, although established as a discipline in computer science for several decades, became a buzzword in 2022 with the emergence of generative AI. Notwithstanding the maturity of AI itself as a scientific discipline, large language models are profoundly immature.
Entrepreneurs, especially those without technical backgrounds, are eager to utilize LLMs and generative AIs as enablers of their business endeavors. While it is reasonable to leverage technological advancements to improve the performance of business processes, in the case of AI, it should be done with caution.
Many business leaders today are driven by hype and external pressure. From startup founders seeking funding to corporate strategists pitching innovation agendas, the instinct is to integrate cutting-edge AI tools as quickly as possible. The race toward integration overlooks critical flaws that lie beneath the surface of generative AI systems.
Related: 3 Costly Mistakes Companies Make When Using Gen AI
1. Large language models and generative AIs have deep algorithmic malfunctions
In simple terms, they have no real understanding of what they are doing, and while you may try to keep them on track, they frequently lose the thread.
These systems don’t think. They predict. Every sentence produced by an LLM is generated through probabilistic token-by-token estimation based on statistical patterns in the data on which they were trained. They do not know truth from falsehood, logic from fallacy or context from noise. Their answers may seem authoritative yet be completely wrong — especially when operating outside familiar training data.
2. Lack of accountability
Incremental development of software is a well-documented approach in which developers can trace back to requirements and have full control over the current status.
This allows them to identify the root causes of logical bugs and take corrective actions while maintaining consistency throughout the system. LLMs develop themselves incrementally, but there is no clue as to what caused the increment, what their last status was or what their current status is.
Modern software engineering is built on transparency and traceability. Every function, module and dependency is observable and accountable. When something fails, logs, tests and documentation guide the developer to resolution. This isn’t true for generative AI.
The LLM model weights are fine-tuned through opaque processes that resemble black-box optimization. No one — not even the developers behind them — can pinpoint what specific training input caused a new behavior to emerge. This makes debugging impossible. It also means these models may degrade unpredictably or shift in performance after retraining cycles, with no audit trail available.
For a business depending on precision, predictability and compliance, this lack of accountability should raise red flags. You can’t version-control an LLM’s internal logic. You can only watch it morph.
Related: A Closer Look at The Pros and Cons of AI in Business
3. Zero-day attacks
Zero-day attacks are traceable in traditional software and systems, and developers can fix the vulnerability because they know what they built and understand the malfunctioning procedure that was exploited.
In LLMs, every day is a zero day, and no one may even be aware of it, because there is no clue about the system’s status.
Security in traditional computing assumes that threats can be detected, diagnosed and patched. The attack vector may be novel, but the response framework exists. Not with generative AI.
Because there is no deterministic codebase behind most of their logic, there is also no way to pinpoint an exploit’s root cause. You only know there’s a problem when it becomes visible in production. And by then, reputational or regulatory damage may already be done.
Considering these significant issues, entrepreneurs should take the following cautionary steps, which I will list here:
1. Use generative AIs in a sandbox mode:
The first and most important step is that entrepreneurs should use generative AIs in a sandbox mode and never integrate them into their business processes.
Integration means never interfacing LLMs with your internal systems by utilizing their APIs.
The term “integration” implies trust. You trust that the component you integrate will perform consistently, maintain your business logic and not corrupt the system. That level of trust is inappropriate for generative AI tools. Using APIs to wire LLMs directly into databases, operations or communication channels is not only risky — it’s reckless. It creates openings for data leaks, functional errors and automated decisions based on misinterpreted contexts.
Instead, treat LLMs as external, isolated engines. Use them in sandbox environments where their outputs can be evaluated before any human or system acts on them.
2. Use human oversight:
As a sandbox utility, assign a human supervisor to prompt the machine, check the output and deliver it back to the internal operations. You must prevent machine-to-machine interaction between LLMs and your internal systems.
Automation sounds efficient — until it isn’t. When LLMs generate outputs that go directly into other machines or processes, you create blind pipelines. There’s no one to say, “This doesn’t look right.” Without human oversight, even a single hallucination can ripple into financial loss, legal issues or misinformation.
The human-in-the-loop model is not a bottleneck — it’s a safeguard.
Related: Artificial Intelligence-Powered Large Language Models: Limitless Possibilities, But Proceed With Caution
3. Never give your business information to generative AIs, and don’t assume they can solve your business problems:
Treat them as dumb and potentially dangerous machines. Use human experts as requirements engineers to define the business architecture and the solution. Then, use a prompt engineer to ask the AI machines specific questions about the implementation — function by function — without revealing the overall purpose.
These tools are not strategic advisors. They don’t understand the business domain, your objectives or the nuances of the problem space. What they generate is linguistic pattern-matching, not solutions grounded in intent.
Business logic must be defined by humans, based on purpose, context and judgment. Use AI only as a tool to support execution, not to design the strategy or own the decisions. Treat AI like a scripting calculator — useful in parts, but never in charge.
In conclusion, generative AI is not yet ready for deep integration into business infrastructure. Its models are immature, their behavior opaque, and their risks poorly understood. Entrepreneurs must reject the hype and adopt a defensive posture. The cost of misuse is not just inefficiency — it is irreversibility.
AI, although established as a discipline in computer science for several decades, became a buzzword in 2022 with the emergence of generative AI. Notwithstanding the maturity of AI itself as a scientific discipline, large language models are profoundly immature.
Entrepreneurs, especially those without technical backgrounds, are eager to utilize LLMs and generative AIs as enablers of their business endeavors. While it is reasonable to leverage technological advancements to improve the performance of business processes, in the case of AI, it should be done with caution.
Many business leaders today are driven by hype and external pressure. From startup founders seeking funding to corporate strategists pitching innovation agendas, the instinct is to integrate cutting-edge AI tools as quickly as possible. The race toward integration overlooks critical flaws that lie beneath the surface of generative AI systems.
The rest of this article is locked.
Join Entrepreneur+ today for access.
