Which task is a generative AI task? Writing a product description from verified facts, or creating an image from a written brief, is one: the model produces new content. Assigning a category, forecasting a number, looking up an order status and calculating a rate each do a different job.
Take a single customer email. You can categorize it, summarize it, or draft a reply to it. All three can run on the same model, but they’re three different tasks. What counts is what you ask the system to produce.
How to Tell Whether a Task Is Generative
Start with the output, then name the operation. Is the result a choice from a fixed set, a number, a stored record or the product of a formula? Or is it newly composed content, such as a paragraph, an image, a function or a message? If it’s composed, the task is generative.
Here’s a teaching example, not a real AffStudio customer message: The lamp arrived with a cracked base. Can you send a replacement?
You could ask a system to assign the label damaged_item. That’s choosing from a list. You could ask it to summarize the message in new wording, which means composing a sentence that didn’t exist before. Or you could ask it to draft a reply using the return policy, which means composing again, with the policy as raw material. Only the first request is classification. The other two generate text.
The tool’s name doesn’t settle the question. NIST describes generative AI as models that emulate the structure of their input data in order to produce derived synthetic content, including text, images, audio and video. That’s a description of what the models are, not a rule that they can only do one kind of work. The task is defined by the output you request.
Run-to-run variation isn’t a reliable test either. A classifier can be probabilistic, and a generator can be set up to answer almost identically every time.
Generative AI Task Examples from Real Work
Start with customer support. According to a case study published by Intercom, tado°, a smart-climate company, sees support demand swing with the weather. In busy periods, conversations rise by as much as 400%, to an average of 11,000 customers a week, against roughly 10,000 a month the rest of the year. The team built a workflow in which Fin, Intercom’s AI agent, first works out who it’s talking to: which customer group, and whether the question is about heat-pump automation, the EV Smart Charging app or voice commands. Then it guides the person and pulls from tado°’s help centers. Emily McKay, a CX Content Writer at tado°, says satisfaction scores stayed stable in peak season and improved on the previous year.
The case study doesn’t say which steps are generative, so what follows is my reading. Identifying the customer group looks like classification. Pulling help-center material is retrieval. Composing the answer from that material is the generative step. The study also notes that a CX Content team owns the voice and tone of the help articles, saved replies and workflow answers, which means people wrote the approved source material. Remember the limits: this is a vendor-published story, the satisfaction figures are company-reported, and it doesn’t measure whether individual replies were correct. A convincing reply can still fail a customer if it answers the wrong question or leans on outdated information. Fluent isn’t the same as resolved.
The coding version of the same split comes from developer Simon Willison, who described his method in a March 2025 post. When he knows what he wants, he plays function designer: he writes the signature and a plain-English spec, and the model writes the body. His example is an async Python function that downloads a SQLite database, rejects oversized files and checks the data is valid. By his account, writing it himself would have taken most of fifteen minutes, and the model produced it in about fifteen seconds. Generating code from requirements is a generative task.
The more useful lesson is what he keeps for himself. He says testing can’t be outsourced. In a later example, a model-written GitHub Pages workflow deployed his page, but he noticed two deploy jobs running, read the documentation and fixed a repository setting himself. He also credits 25-plus years of experience for how well his prompts worked. That raises a question I’d put to any reader: when does checking cost more than doing? Willison could judge the output quickly because he knew the territory. Someone who can’t judge it may spend as long verifying as writing. Keep in mind it’s one practitioner’s method from early 2025, not a rule for every tool.
The same pattern appears elsewhere. Shopify’s help center says its assistant takes a product title and keywords and drafts a description, then warns that the merchant stays responsible for accuracy because generated text can include benefits nobody listed. An image from a brief, a five-bullet meeting summary and a batch of made-up support requests for testing a help desk all produce something that didn’t exist before. For a closer look at how teams chain these steps, see AI marketing workflows.
Which Tasks Are Not Generative AI?
Four operations are easy to mistake for generation. Marking a message spam or not spam is classification: pick one label. Forecasting next month’s sales is prediction: estimate a number from past data. Retrieving an order’s status is a lookup: the answer already exists in a record, and nothing needs composing. Calculating a conversion rate is arithmetic: conversions divided by visitors.
Marketing offers a clean contrast. Writing an ad produces a new text draft, which is generative. Predicting the probability that someone clicks it estimates an unknown future value. Calculating click-through rate divides known clicks by known impressions and multiplies by 100. Those are three different questions with three different outputs.
A generative model can still sit inside a larger system without changing what each operation is. It might phrase the order-status message in a friendly tone, but the status itself comes from a lookup.
Where a formula, a rule or a lookup can produce the answer, asking a model to generate it just adds another way to be wrong. Generation earns its place when the output has to be composed.
The Tricky Cases: Summaries, Chatbots and Classification
Is summarization generative AI? It is when the model writes the summary in its own words. An extractive approach that simply selects sentences from the original isn’t generating anything. Rewording doesn’t give a model license to add facts, though. Suppose your note says, “We discussed a Friday launch. No date was agreed,” and the summary reads “Launch confirmed for Friday.” The wording is smooth and the meaning has changed. That’s an illustrative example, not a failure observed in any named product.
Is a chatbot generative AI? Not automatically. A chatbot can show prepared answers or generate text, and a chat window alone doesn’t tell you which. What matters is how the reply is produced.
Can generative AI classify text? Yes. Google Cloud’s model-evaluation sample prompts a Gemini model to sort text into one of five categories and return only the category name, then scores the labels against known answers. The model is generative, but the requested output is a single label, so the task’s objective is classification. If you only want positive, negative or neutral back, you’ve asked for classification.
Meeting notes make the summary question personal. Thorin Klosowski, a senior security and privacy analyst at the Electronic Frontier Foundation, told the Associated Press that text is cheaper to store and easier to search than audio or video. That’s a point about records. My own view is that a searchable record and an understood agreement are different things. A summary can tell you what was written down, but only you can decide whether you know what you committed to. That’s an interpretation, not evidence about memory.
Try It Yourself: Turn Product Facts into Useful Copy
You can repeat this exercise in any chat tool without setting up an automation. The facts below are fictional, invented for the exercise: a desk lamp, three brightness settings, USB-C powered, price $29.
Start with the weak prompt: Write an ad for this lamp. It gives the model almost nothing, so it will fill the gaps. Shopify’s guidance points the same way: give a title and at least two features, and expect the draft to need a careful read.
Now tighten it: Write two English headlines for a desk lamp, each no more than eight words. Use only these facts: three brightness settings, USB-C powered, $29. Don’t add claims about eye health, battery life, warranties or performance.
Here are two illustrative outputs. I haven’t run a model for this article.
An acceptable one: “A $29 desk lamp with three brightness settings.” That’s eight words, and every claim traces back to a supplied fact.
One with an unsupported promise: “Reduce eye strain with a USB-C desk lamp.” Eye strain appears nowhere in the facts, and the prompt banned it. The fix is to cut the claim: “USB-C desk lamp with three brightness settings.”
Before using any generated copy, run three checks. Facts: can each claim be traced to something you supplied? Constraints: did it respect the word limit and the banned topics? Clarity: does it read cleanly? Passing doesn’t prove the headline will convert. It only shows the copy is faithful to your facts.
Quick Quiz: Can You Identify the Generative Task?
These are original practice questions, not an official exam.
- Which is the generative AI task: writing a product description, forecasting traffic, or categorizing a support request?
- Which operation does not create content: writing headlines, creating an illustration, or assigning a spam/not-spam label?
- A model summarizes a meeting in five bullets, in its own wording. What kind of task is that?
- A workflow has four steps: (1) categorize the request, (2) retrieve the return policy, (3) draft a personalized reply, (4) save the record. Which step is generative?
- A generative model returns only positive, negative or neutral. Why is that classification?
Answers.
- Writing a product description. It composes new text. Forecasting estimates a number, and categorizing picks a label.
- Assigning the spam label. It selects from a fixed set. Headlines and illustrations are new content.
- Generative, provided the model writes the bullets itself. If it only picked sentences, it would be extractive. Either way, check that nothing was added.
- Step 3, the drafted reply. Categorizing chooses a label, retrieval fetches existing text, and saving writes a record.
- The requested output is one label from a fixed set. The model is generative, but the objective is classification.
The practical takeaway: name the output you need, then work out which part of the process truly requires generation.





Leave a Reply