Most people using AI today are getting an 80% answer and shipping it. Sometimes that's fine. A first draft, a brainstorm, a starter outline, a throwaway script. 80% is a gift on those.
The problem is that 80% is also good enough to fool you on the 20% that matters. The output is fluent, confident, organized, and wrong in places you can't see unless you already know the right answer. Same trick a competent politician runs: say it well enough and most people will assume it's true.
#1 · The 80% looks like 100%.
Open any AI tool, type a reasonable prompt, get back something coherent. It will have structure. It will have the right vocabulary. It will sound like someone who knows what they're talking about.
That fluency is the whole problem. You can't tell the difference between a confident right answer and a confident wrong one by reading them. You can only tell by knowing the subject well enough to catch the miss.
If the topic is something you don't know, AI's output is a starting point you have to verify. If the topic is something you do know, AI's output is something you can actually fix. Both are useful. Neither is finished.
#2 · Prompt engineering is the skill nobody wants to learn.
Prompt engineering sounds boring and slightly fake, like “synergy” or “thought leadership.” It isn't. It's the difference between a generic answer and an answer shaped by what you actually need.
A good prompt encodes context, constraints, format, audience, and a target outcome. A bad prompt is “write me a blog post about AI.” Both will give you words. Only one will give you something you can use.
The people getting real leverage out of AI right now are not the ones with the most expensive subscriptions. They're the ones who learned how to talk to it, and who built a personal library of prompts that consistently produce work they'd be willing to put their name on.
A bad prompt: “Write a follow-up email to a client.” A useful prompt: “Write a 90-word follow-up to a client who went quiet after a proposal. They're risk-averse, in healthcare, and the proposal was for a website rebuild. Tone: warm, not pushy. Goal: get them to reply with one specific objection.”
#3 · AI is a confident liar.
Hallucinations are not a bug that's getting fixed. They're a structural feature of how these models work. The model is predicting the next likely token. It is not checking reality. When reality and likelihood disagree, likelihood wins, and the answer comes back in the same self-assured tone as everything else.
This matters most where the cost of being wrong is high. Legal language, medical claims, financial figures, technical specs, anything a customer is going to act on. The model will give you something. It is on you to know whether the something is true.
Treat AI output the way you'd treat a smart intern on day one. Capable. Eager. Plausible. Wrong often enough that you don't sign anything they hand you without reading it.
#4 · Experience is the filter.
The 20% that AI can't reliably do is the 20% that took you years to learn. It's knowing that this phrasing will land with a CFO and that phrasing will get you ignored. It's catching that the “best practice” the model just recommended is the same one that quietly sank a similar project two years ago. It's spotting the missing step in a workflow because you've actually run the workflow.
That kind of judgment doesn't come out of a model. It comes out of having done the work, badly, and then better, and then better than that.
This is the part of the job AI makes more valuable, not less. When everyone has access to a tool that produces fluent, organized, plausible work, the scarce resource becomes the person who can tell what's actually right.
#5 · What actually moves the needle.
The short list of what separates people getting real value from AI from people just generating filler:
- They know the subject well enough to spot the misses.
- They've built prompts that work for their specific use cases, and they reuse them.
- They use AI for the parts of the job where being 80% right is genuinely fine, and they do the last 20% themselves on the parts where it isn't.
- They don't ship anything important without reading it like a skeptic.
That's it. That's the whole edge.
Final thoughts.
AI is not magic and it is not a fraud. It's a very fast, very fluent assistant that will hand you confident, well-organized work that is sometimes wrong in ways you can't catch unless you already know better.
The people winning with it aren't the ones with the best tools. They're the ones who already knew the work, learned how to ask for what they needed, and kept their hands on the wheel for the part that matters.
80% is a gift. The other 20% is still the job.