The five most common mistakes with AI — and how to avoid them

"I've tried it — it doesn't work for us."

That sentence is common. Almost always, it's not because AI is bad for their specific use case. It's because they have made one of five identifiable mistakes. And once you know what those mistakes are, they are easy to fix.

Mistake 1: You treat AI as an encyclopedia

The most common and dangerous mistake. You ask AI about a specific law paragraph, a statistic quote, a company registration number, a historical date — and copy the answer directly.

The problem is that AI hallucinates, and it happens regularly. It generates answers that sound completely credible but can be completely wrong. A model that doesn't know the answer rarely says so. It gives you an answer that sounds right.

It is a fundamental part of how these models work, not a bug that will be fixed in the next version. The purpose is to generate the most likely next word — not to search a database of verified facts.

The solution:

• Always verify specific facts (numbers, dates, names, laws) against the primary source.
• Use AI for what it's good at: phrasing, structure, and analyzing text you provide it.
• Ask the model: "How confident are you about this?" — a good model will flag its uncertainty.
• Use Perplexity, or Claude and ChatGPT with web search turned on, for factual questions. Then you get source references you can check.

Mistake 2: Your prompts are too short and vague

"Write an email to a customer."

That prompt produces an email. It is technically a correct answer. But it is an email that fits everyone and therefore no one — without a specific tone, without the right context, without the detail that makes it relevant.

AI is not telepathic. It fills in the gaps with its own assumptions about what is average and generic. If you want something specific, you must specify it.

Compare these two prompts:

❌ Weak: "Write an email to a customer."

✅ Strong: "Write a follow-up email to a customer who received a quote for 45,000 SEK for web design 10 days ago and hasn't replied. The customer is the CEO of a real estate company in Gothenburg. We previously built the website for their competitor. Tone: friendly but professional, no sales tactics. Max 120 words. End with an open question, not 'let me know'."

The strong prompt gives a specific, usable result. The weak prompt gives you something to start over from.

Rule of thumb: A good prompt takes at least 2 minutes to write. If it took less than 30 seconds — add more context.

Mistake 3: You are choosing the wrong model for the task

Not all AI models are equal — and even the same model doesn't perform equally well on all tasks. Using the smallest, fastest model for a complex legal document often gives a shallow answer, which makes it look as if AI cannot do the task.

The most common version of this mistake is using a free model for tasks that require a paid model — and concluding that "AI can't do it" when in reality, it's just a case of using the wrong tool.

Mistake 4: You accept the first answer

Treat the first response as a draft that you improve together with the AI.

Those who get exceptional results from AI models iterate consistently. They ask for three variations. They ask the model to explain its reasoning. They say "this was good, but remove X and make Y more prominent." They treat the interaction as a conversation with a skilled colleague, not as a search query.

A concrete iteration loop that works:

Step 1: Run your original prompt. Read the response.

Step 2: Identify what is good and what is missing.

Step 3: "This was good as a starting point.
  What I want to keep: [X].
  What I want to change: [Y] — instead, I want [Z].
  What is completely missing: [W].
  Rewrite with these adjustments."

Step 4: Repeat 2–3 times until you are satisfied.

Step 5: Make the final fine-tuning adjustments manually — it takes 5 minutes and is the only way to ensure that it actually sounds like you.

For a longer document this usually takes a fraction of the time it takes to write everything from scratch.

Mistake 5: You expect AI to replace your knowledge

The final and most subtle mistake.

AI is not an expert in your business. It has never met your customer. It doesn't know what your industry actually entails — in the specific, nuanced way that 10 years in the field provides.

AI is best at helping you communicate, structure and scale your expertise, but it cannot replace it.

The practical mistake: you ask AI to provide a strategy for your business without giving it the context it needs — and are then disappointed when the answer is generic. Of course, it's generic. You gave it generic information.

The inverse principle: the more specific you are about your situation, your customers, your challenges, and your context, the more specific and useful the AI's output becomes.

❌ "Give me a strategy to increase my sales."

✅ "I run an accounting firm with 8 employees in Borås.
Primary target group: manufacturing companies with 20–150 employees.
Biggest problem right now: 40% of incoming inquiries are not profitable customers (too small or wrong industry).
Average customer lifetime: 2.3 years. NPS: 61.

Help me identify the three specific points in my current sales flow that filter out the most unprofitable customers — and what I can concretely do for each of them."

One last thing

All five mistakes share a common denominator: they are about treating AI as a ready-made answer system rather than a reasoning support that you collaborate with.

Change that basic mindset, and everything else changes with it.

AI is like a highly skilled, extremely fast, never-tired intellectual tool that knows a little about everything and nothing about your specific context. It is up to you to provide the context. And it is up to you to verify that what it produces is actually correct.