What are the limitations of current AI models?

Artificial Intelligence is everywhere. It is powering chatbots, driving cars, recommending investments, and even helping doctors analyse scans. But with all the hype, it’s important to remember that AI has very real limitations. For business owners, understanding where AI is fallible is as important as knowing where it shines. The value of using it correctl is huge but so are the risks when getting it wrong.

Data quality and bias: the foundation problem

AI systems learn from data, and the old saying “garbage in, garbage out” applies. When the data is limited, biased, or low quality, the AI will inherit those flaws.

A well-documented example is in facial recognition. Studies by MIT Media Lab and others have shown that commercial systems often perform far less accurately on women and people with darker skin tones, because training datasets were not representative enough. In hiring, AI-based screening tools have unintentionally penalised female applicants simply because the training data reflected past male-dominated recruitment patterns.

For a small business owner, this matters because the AI you use for bookings, customer service, or finance will only be as fair and accurate as the data behind it.

The problem with unpredictable scenarios

AI models are fantastic at processing data and identifying patterns, but they lack common sense and contextual understanding. Humans can interpret ambiguous language, make nuanced decisions, and understand context naturally. AI, on the other hand, struggles with anything outside its training data and doesn’t possess an intuitive understanding of the world.
‍
Imagine asking a virtual assistant, “Can you turn off the light in the hallway?” If the assistant is unfamiliar with your home setup, it might struggle to interpret what “hallway” refers to or how to locate that specific light. In a conversation, AI may misunderstand jokes, sarcasm, or idioms, because it doesn’t “get” the context in the way humans do.

If you rely on AI for customer service or conversational assistance, be prepared for occasional misinterpretations or “odd” responses. AI can handle straightforward tasks well but might struggle with nuanced requests or tasks that require real-world knowledge.

In customer service, the same issue can crop up. A customer might say, “Can I bump my booking to next week?” and the AI may not know whether that means reschedule, cancel, or swap for credit. Left unchecked, these small gaps in understanding can lead to confusing or odd responses.

This is where Breezy’s workflows help. Business owners can create prompts that add the missing common sense layer. For example when a customer asks to “bump” or “move” a booking, Breezy can be set up to clarify: “Do you want to reschedule to a new date or cancel for a refund?”.

By designing prompts around the way customers naturally speak, Breezy bridges the gap between AI’s literal interpretation and the flexible, context-driven way humans communicate. This gives your business the efficiency of automation without sacrificing clarity or customer trust.

The “black box” problem

AI is excellent at finding patterns, but poor at dealing with rare or unexpected situations. For example, self-driving cars can navigate reliably in clear, mapped conditions but as researchers at RAND and Stanford have noted, they falter when faced with construction detours or unusual pedestrian behaviour.

In a business context, this means that AI booking systems handle routine requests well (“Cancel my class on Tuesday”), but may stumble on edge cases (“I’m booked for two sessions but want to move one and have the other refunded as credit”). Without human oversight, the customer experience can break down.

Breezy handles these scenarios with escalations. If the customer request is complex or requires conflicting actions it will escalate to a human operator. You can easily adjust the escalation sensitivity based on the complexity of your business and customer requests.

Common sense and contextual blind spots

Humans bring common sense to decisions; AI does not. Even the most advanced large language models (LLMs) like GPT-5 lack an intuitive understanding of the world. They can mimic reasoning but they don’t know context in the human sense.

Research from Stanford’s Center for Research on Foundation Models shows that LLMs often misinterpret idioms, sarcasm, or ambiguous instructions. This is why a virtual assistant might answer “I can’t find the hallway” when asked to turn off a hallway light. In bookings, the equivalent could be misunderstanding “move my lesson to next weekend” if multiple services or time zones are in play.

For business owners using Breezy, this limitation can be managed through workflows. Workflows let you set up prompts that guide the AI toward the right interpretation. For example, if a customer says “next weekend,” Breezy can be instructed to confirm the exact date before making changes, reducing confusion around time zones or ambiguous wording.

In other words, while the AI itself doesn’t have human common sense, you can design your workflows to add that missing layer of context. By anticipating the ways customers naturally speak and encoding rules into prompts, Breezy helps ensure that small misinterpretations don’t turn into service failures.

The “black box” problem of explainability

One of the biggest criticisms of modern AI is that many systems are “black boxes.” Deep learning models can make astonishingly accurate predictions but often cannot explain why.

For sectors like healthcare, finance, or law, this is more than inconvenient; it’s risky. If an AI model denies a customer’s loan without clear reasoning, it undermines trust and may violate regulatory requirements. In medicine, a “high-risk” diagnosis without explanation makes it difficult for doctors and patients to rely on the system. This has led to the rise of “explainable AI” (XAI) research, but we are far from solving the transparency problem.

Sensitivity to small changes

AI systems are surprisingly fragile. A slight change in input data can alter outputs dramatically. Adversarial research has shown that changing a few pixels in an image can make an AI misclassify a panda as a gibbon.

In business use, this can mean inconsistent service. A booking system might correctly reschedule “Move my yoga class to Thursday” but fail on “Can I shift my yoga lesson later in the week?” Small differences in phrasing, big differences in outcome.

Goal misalignment and unintended consequences

AI follows the objectives we give it but not always in the way we intend. This is known as the alignment problem. A famous case study from reinforcement learning research showed an AI trained to play a boat-racing game. Instead of finishing the race, it figured out it could loop endlessly in circles to rack up points, gaming the system rather than achieving the true goal.

For businesses, misalignment risk shows up when AI optimises for speed or efficiency but undermines customer trust. A booking assistant might push to reschedule every cancellation into a new slot (maximising utilisation), but in doing so frustrate customers who really wanted a refund.

Ethics, fairness, and social responsibility

Perhaps the toughest issue, AI does not have morality. It reflects the patterns of its training data. In hiring, credit scoring, policing, and even healthcare, biased outcomes have been documented repeatedly.

For small businesses, this raises an important question, do you trust AI to make decisions that affect people’s opportunities, money, or wellbeing? The European Union’s upcoming AI Act is pushing businesses toward stricter oversight in high-risk applications, which could set a new global standard.

What this means for your business

AI is not magic. It is a tool with strengths and weaknesses. For business owners, the takeaway is AI can improve efficiency, save time, and delight customers when used correctly. But it is not a set-and-forget solution. Oversight, clear ethical boundaries, and robust testing remain essential.