The strong push for AI integration into modern businesses isn’t without reason – the capabilities of artificial intelligence are considerable. And it’s probably true that the businesses that fail to adopt it will end up being left behind. Used well, it compresses the time and effort required for tasks. Used badly, though, it can result in outcomes that are worse than those found by businesses that never integrate it to begin with. We’ve talked about how AI tools can accelerate what you do, but just as important is knowing how not to misuse them; let’s address that now.
Augmentation, not abdication
The biggest mistake a founder can make is outsourcing judgment to an LLM or AI. Judgment is the reason that AI will never make humans obsolete. You can understand context, ethics, and trade-offs in a way that can never satisfactorily be left to a machine. AI is like a power drill: it can make a DIY task much faster and cleaner; it can also cause a disastrous flood. The difference is how it is handled, and that’s the human side of the equation.
To look at it practically, ask yourself which part of a task is generative, which is factual, and which is judgmental. When you have considered that, apply the following detail:
AI and LLMs can generate options and structure
You can let AI insert facts, but you should always double-check them against a trusted source
Deal with the judgment side yourself. Content, code, and tone are all things only a human can check.
Why AI goes wrong sometimes
There have already been numerous examples in international news of AI applications that have caused expensive or embarrassing errors, which can be extremely injurious to trust. Why does this happen? It’s because AI is only as good as its programming – it has access to all the information in the world, but information without context or guardrails isn’t that useful.
You may have read about how ChatGPT 5 delivered the wrong answer when asked how many “B”s there were in the word “blueberry”. Look at the word: it’s two, no room for disagreement, right? But at least one user has shown examples of the LLM stating there are three: one at the start, one in the middle, and one in the “berry” section of the word. ChatGPT, like any large language model, generally delivers information by predicting the next word in a sentence. It’s bad at counting. And not only that, it’s confidently bad – it will state falsehoods as facts from time to time, so you need to check its work.
Prompt leakage
If you want an LLM to produce content based on a client brief, be aware that it doesn’t understand privacy the way we do. The raw data you feed in – and ask the AI to process in producing your finished document – may not be intended for the eyes of the public. But the AI doesn’t understand that, and even if you tell it that, it may still reproduce the data in its output. This can violate contracts or regulations.
Speculative reasoning
AI applications work by extrapolating from the information it has. This can lead to faulty conclusions, which is understandable when you want a film review based on some actor names, plot points, and personal opinions. It’s another thing entirely if you’re looking for medical advice or niche legal statutes that may differ across jurisdictions. Part of the problem here is overhyping by AI evangelists; people will claim that it can be a lawyer, a doctor, a PhD scholar – but each of these roles requires years of specialized study, and should not be entrusted to something more akin to a talkative search engine.
None of this is to say that AI and LLMs aren’t useful, but their skill lies in reproducing information that is presented to them in a readable or applicable way. An AI is no more a lawyer than someone who has been shown a diagram of the human body is a doctor.
Make AI work because you understand it
AI applications shine when you’ve done the groundwork. Set clear goals, provide clear data, and perform clear checks. If you’re serious about LLM readiness, invest some time in aligning your content with how modern models read, rank, and reason. Understanding search intent and structured content enables you to create content that’s ready for AI comprehension, featuring headings, schema, and conversational clarity. The result will be that AI applications and models and people can understand your work and find it online, in context, and in a way they can use.
High-stakes arenas
The “move fast and break things” ethos behind much of AI adoption has its place in finding profit margins where none existed before. But there are some domains where it can lead to harm, and these areas need to be vetted all the more closely.
Medicine
You can use AI applications to summarize literature, structure already-written notes, or draft information in a way that makes sense to patients who are not medically educated. You should never use it to make a diagnosis, select a drug or treatment plan, or set dosing without review by a trained clinician. The danger of hallucination is bad when the AI is selecting paint choices or diet recommendations; it can be fatal when it misses drug interactions or contraindications, things a doctor would notice.
AI can be beneficial when researching, comparing documents, and converting legalese into plain English. It should never be used to draw up legal briefs, especially without a trained lawyer closely reading it for citations and jurisdictional nuance. AI, for whatever reason, is terrible at referencing information; even if the information is true, it has a habit of citing studies and cases that never existed. Inaccurately cited briefs can be terminal for a case, and misuse of AI can lead to sanctions for lawyers and firms; in short, the risks far outweigh the convenience.
AI applications have many appropriate uses in the workplace, and some of their stated shortcomings are overstated. However, be cautious of the knowledge that these shortcomings exist and never rely solely on it. Artificial intelligence is always at its most powerful when twinned with actual wisdom.
Images by マクフライ 腰抜け, Steve Buissinne, & Rubén González; Pixabay
Editor in Chief of Under30CEO. I have a passion for helping educate the next generation of leaders. MBA from Graduate School of Business. Former tech startup founder. Regular speaker at entrepreneurship conferences and events.
There have already been numerous examples in international news of AI applications that have caused expensive or embarrassing errors, which can be extremely injurious to trust. Why does this happen? It’s because AI is only as good as its programming – it has access to all the information in the world, but information without context or guardrails isn’t that useful.
The “move fast and break things” ethos behind much of AI adoption has its place in finding profit margins where none existed before. But there are some domains where it can lead to harm, and these areas need to be vetted all the more closely.
