Start Lesson
A bakery owner I worked with spent two hours every week writing her newsletter. She started using AI to draft it and cut that time to fifteen minutes. That same month, a lawyer asked AI to find relevant case law — and it returned six citations that looked perfect. Three of those cases did not exist.
Same technology. One outcome was a time-saver. The other was nearly a career-ender. The difference was not the tool. It was the type of task.
In the last lesson, you learned that AI is a prediction engine — it predicts the next word based on patterns, not understanding. This lesson maps exactly where that prediction ability shines and where it breaks down. After this lesson, you will be able to sort any task into "AI-ready" or "needs human verification" — and explain why.
These tasks play directly to prediction's strengths. In every case, AI has seen millions of examples of the pattern and can produce a strong first draft.
Generation. First drafts of blog posts, emails, product descriptions, job listings. The bakery owner's newsletter is a textbook example. AI had seen thousands of small-business newsletters and could predict the right tone, structure, and length.
Summarization. Hand AI a 30-page report and ask for a one-page summary. It identifies the most frequently emphasized information and restructures it. Works well for meeting notes, research papers, and customer feedback — anything where the source material is right there in the input.
Classification. "Is this review positive or negative?" "Which department should this ticket go to?" AI excels at sorting things into categories because it recognizes the language patterns of each one.
Extraction. "Pull every date, dollar amount, and name from this contract." AI scans text and pulls out structured information faster than any human — because the patterns of dates, dollar amounts, and names are distinctive and well-represented in its training data.
Translation. Not just between languages, but between formats — turning a formal report into a casual Slack message, or rewriting technical documentation for a non-technical audience. Transforming one pattern into another is exactly what prediction does well.
Notice the common thread: every task on this list involves transforming or recognizing patterns in text that is already provided. The AI does not need to know anything beyond the input. That is the sweet spot.
These tasks break the prediction model because they require something prediction cannot deliver.
Reliable math. The model guesses what a correct answer looks like, not what it is. Remember from the last lesson: "7,849 times 3,271" is not a language pattern. It is a calculation. Always verify numbers.
Real-time information. Models are trained on data up to a cutoff date. They do not know today's stock price or yesterday's news. Some tools connect AI to the internet, but the base model works from stale data and does not signal when its knowledge is outdated.
Consistent memory across long conversations. AI processes a window of text (the "context window") and makes predictions based on what fits inside. In a long conversation, earlier details can fall outside that window and get lost — not because the AI decided they were unimportant, but because its prediction only considers what is in front of it right now.
Following rules perfectly. Tell AI "never mention competitors" and it will follow that instruction most of the time — but not always. Prediction is probabilistic. There is no internal rule engine enforcing your constraint; there is only a statistical tendency to comply.
Genuine reasoning about novel situations. AI can mimic the pattern of reasoning because it has seen reasoning in its training data. But when a situation is truly novel — no pattern to match — the output may look thoughtful while being wrong. This is hardest to catch because the format looks right even when the substance is not.
Here is the lawyer's story in full. He asked an AI to find relevant case law for a filing. The AI returned six cases with complete citations — court names, dates, docket numbers. They looked exactly like real citations because the AI had seen thousands of real citations and predicted what one should look like in this context.
Three of those cases did not exist. The AI had not looked anything up. It had predicted what plausible citations would be, and "plausible" is not the same as "real."
This is called a hallucination, and it is not a bug. It is a direct consequence of how prediction works. When the model does not have a strong pattern to follow for the specific fact you need, it generates the most likely-looking output. The dangerous part: it presents fabricated information with the same confidence as verified facts. There is no uncertainty signal.
The rule: Never trust AI on specific facts — names, dates, statistics, citations — without verifying them independently.
Here is the practical framework I use with every client: AI is spectacular for the first 80% and dangerous for the last 20%.
The first 80% is the draft, the structure, the heavy lifting. AI gets you from a blank page to a solid starting point faster than any tool in history.
The last 20% is accuracy, nuance, and judgment. The financial advisor still verifies the numbers. The lawyer still checks the citations. The manager still reads the performance review before sending it.
The people who get the most value from AI use it as an incredibly fast first-draft machine, then apply their own expertise to finish the job.
Think of five tasks you do regularly at work. For each one, ask two questions:
Tasks that are pattern-based with low stakes if wrong — those are your best candidates for AI right now. Tasks that are precision-based with high stakes — those need a human, with AI at most providing a starting point you verify completely.
Write your five tasks down. You will use this list again when we talk about adoption levels in the next lesson.
You now know what AI can and cannot do. But knowing the technology's limits is only half the picture. The next lesson introduces a framework for how to adopt AI in your work — three levels, from simple assistance to full autonomy — and why most businesses should start at Level 1.
