Here are real examples of generative AI, grouped the way you actually use them: by business function, not by which model is under the hood. Marketing gets first drafts. Support gets faster first replies. Sales and HR get the admin nobody wants to write. Developers get boilerplate. None of it replaces the judgment call at the end. All of it removes the blank page at the start.
This post covers what generative AI actually is, twenty examples organised by where they show up in a business, what it still gets wrong, and how to start without signing a contract you will regret in six months.
TL;DR
Generative AI is good at producing a first version of something (text, an image, a block of code) fast and cheap. It is not good at deciding whether that version is the right one. Use it to remove the blank page. Keep a human on the decision that follows. That single rule covers about 90% of what goes wrong with it.
What actually counts as generative AI
Generative AI is any model that creates new content from a prompt, rather than sorting, predicting, or extracting information from content that already exists. That distinction matters, because it is the line between "generative" and the other kind of AI most businesses run into: the automation and classification tools that move data around without generating anything new.
- Text generation (ChatGPT, Claude, Gemini): emails, drafts, summaries, structured content
- Image generation (Midjourney, DALL-E, Stable Diffusion): mockups, ad creative, concept art
- Code generation (GitHub Copilot, Code Llama): boilerplate, test coverage, first-pass functions
- Audio and voice generation (resemble.ai, ElevenLabs): voiceover drafts, synthetic narration
If you want the process side of this (how a business actually decides what to automate versus what to generate), the automation vs AI post covers that split in more depth. This one is about the generative half only.
Photo: Pexels
Marketing and content: first drafts, not final copy
This is where generative AI shows up first in almost every business, because marketing produces the most raw text and images per week of any department.
- First-draft blog posts, product descriptions, and ad copy variations for A/B testing
- Social captions and short-form video scripts, drafted in bulk and edited down
- Concept images and mockups for campaigns before commissioning a photographer or designer
- Email subject line variations, tested at a volume no single person would write by hand
The pattern I keep seeing: teams that use generative AI well treat every output as a first draft, full stop, no exceptions. Teams that get burned are the ones that publish the first draft because it read fine on a quick scan. It usually does read fine. That is exactly the problem: confident and wrong looks identical to confident and right until someone checks the facts.
Photo: Pexels
Customer support: faster replies, same amount of judgment
Support teams use generative AI to draft the first response to a common query, summarise a long email thread before a human picks it up, or turn scattered notes into a clean help-centre article. What it should not do, in my honest opinion, is send that first response without a person reading it. Not because the tool is bad, but because a wrong answer sent to a paying customer costs more than the time saved writing it.
Small businesses I have worked with get the most value from generative AI in support by using it to write the internal knowledge base faster, not the outbound message. Draft the article, have someone who actually knows the product check it, publish it once. That one article then answers the same question for every future customer without generating anything new at all.
Photo: Pexels
Sales and internal ops: the unglamorous wins
Nobody puts this on a conference slide, but it is where most of the actual hours get saved:
- Drafting personalised outreach emails from a short brief, instead of starting from a blank template
- Turning a messy call transcript into a clean meeting summary with action items
- First-draft proposals and one-pagers, built from a scope the account manager already knows
- Internal SOP drafts, written from a rough voice memo of "how we actually do this"
None of this is glamorous. Most of it is also exactly the kind of writing that used to eat an hour of a salesperson's day between actual calls. Give that hour back and nine times out of ten they spend it talking to more people, which is the actual job.
HR and people ops
Job postings drafted from a rough list of requirements. Interview question sets tailored to a specific role. Onboarding documents drafted once and reused for every new hire. This is low-risk generative AI use: the output is reviewed by a human before it reaches a candidate or an employee in every sensible workflow, so the hallucination risk gets caught before it costs anything.
The one place I would slow down: anything touching a hiring decision itself, rather than the paperwork around it. Screening candidates with a generative model, rather than drafting the questions you ask them yourself, is where the legal and ethical exposure starts. Draft with it. Decide with a person.
Product and design
Concept images for a new product line before committing to a photoshoot. Ten quick layout variations for a landing page before a designer builds the real thing. Placeholder copy and imagery for a prototype so stakeholders can react to something concrete instead of a blank Figma frame. Generative AI is genuinely strong here, because "rough version to react to" is precisely what it produces well, and precisely what design teams used to spend a day making by hand.
Photo: Pexels
Software development
The example everyone already knows: GitHub Copilot and similar tools generate boilerplate, suggest test coverage, and translate a clear specification into working code far faster than typing it by hand. If you want the deeper version of this (which development roles it actually touches and which ones it does not), the jobs AI can't replace post covers software developers specifically, including the nuance most listicles skip.
What generative AI still can't do
Every example above shares the same limitation, so it is worth stating plainly instead of burying it in a disclaimer at the bottom.
It cannot tell you when it is guessing
A generative model produces its most confident-sounding sentence whether the underlying fact is correct or invented. There is no built-in "I am not sure" signal. You have to build that check yourself, every time.
It cannot own the outcome
If a generated email goes to the wrong client, a generated image infringes a copyright, or a generated line of code ships a bug to production, the accountability sits with the person who published it. That person needs to actually read the output first.
It cannot make the strategic call
Ten headline variations is a generative AI task. Deciding which one matches what the brand should stand for this quarter is a judgment call. The model can produce the options. It cannot pick between them on your behalf and mean it.
None of this is unique to generative AI in particular. It is the same rule I give on every automation project: don't implement AI without a real person owning the output. Automation, or generation, that nobody checks is a liability wearing a productivity costume. Someone has to review it, monitor it, and be willing to switch it off if it breaks. That part does not get automated.
How to start without wasting €5,000
I reckon the biggest mistake small businesses make with generative AI is not choosing the wrong tool. It is skipping the free tier and going straight to an enterprise contract before anyone on the team has actually used the thing.
This reminds me of a construction company I worked with a few years back, though the project itself was financial software, not generative AI. I showed up to demonstrate a new system, walked the team through every feature, and got nothing back. No buy-in, no questions, no next steps. I went home with nothing to show for it.
The problem wasn't the software. Nobody had told the finance team why the company was changing systems or what it was meant to solve. They weren't part of the decision, so when I showed up with "here is your new tool," they had no reason to care. I told the manager straight: get everyone aligned on the why before I come back. Two weeks later, full team buy-in, and the implementation went smoothly.
Generative AI rollouts fail the exact same way. A manager reads about ChatGPT, buys ten enterprise seats, and wonders three months later why usage is near zero. Nobody explained what problem it was supposed to solve, so nobody built the habit of using it. The tool was never the blocker. The buy-in was.
My honest rule of thumb, the same one I apply to every automation project: if an AI project costs more than €5,000 and doesn't have a measurable ROI target, kill it. For generative AI specifically, that usually means you don't need €5,000 at all yet. Start on a free tier, with one real task, with one person who actually wants to use it. Measure the hours it saves on that one task for two weeks. If it saves real time, expand it. If it doesn't, you have lost nothing but two weeks and you know that tool wasn't the answer.
For the practical side of picking tools once you are past that stage, the AI tools for small business post has the shortlist by function, with actual prices. And if generative AI turns out to be the smaller half of what you need, because the real bottleneck is moving data between systems and not generating text, the AI automation guide covers that other half.
According to Stanford's AI Index, business adoption of generative AI has grown fast since 2023, but usage maturity, meaning teams that actually measure the ROI and not just the seat count, lags well behind adoption. That gap is exactly where the €5,000 mistakes happen. Skip it by measuring from week one.
For the bigger economic picture, McKinsey's research on generative AI's economic potential is worth a read, and NIST's AI Risk Management Framework is the closest thing to an official checklist for the "someone has to own the output" problem, if you want the formal version of what I just said above in plain language.
Frequently asked questions
What are some real examples of generative AI in business?
First-draft ad copy and product descriptions, meeting summaries, job postings, first-response customer emails, image and mockup generation, and boilerplate code are the ones I see used correctly, day to day, in small and mid-market businesses. Notice none of them are "final output." That is the pattern that matters more than the list.
What is the difference between generative AI and AI automation?
Generative AI creates new content (text, images, code, audio) from a prompt. AI automation moves information and triggers actions between systems, with or without generative content involved. A tool that writes your email is generative. A tool that reads a form and files it in the right folder is automation. They often get bundled together, but the jobs they do are different, and the ROI math is different too.
Can generative AI replace a marketing team?
No, and treat any pitch that says otherwise with suspicion. It replaces the first-draft stage: the blank page, the fifteenth variation of the same headline. It does not replace the person who decides what the brand should stand for, reads the room on a campaign that is about to age badly, or takes responsibility when the messaging misses. That is still a human job, and it is the harder half of marketing.
What can generative AI not do well?
Strategic judgment under ambiguity, anything requiring current or verified facts without a human check, and any output where being wrong has real legal or financial consequences. It also cannot tell you it is unsure. It will produce a confident, well-formatted, entirely wrong answer with exactly the same tone as a correct one. That is the part people underestimate.
Is ChatGPT considered generative AI?
Yes. ChatGPT, Claude, and Gemini are all generative AI, specifically large language models trained to generate text. Midjourney and DALL-E generate images. GitHub Copilot generates code. The "generative" label refers to any AI model that creates new content, as opposed to AI that classifies, predicts, or extracts information from existing content.
How much does it cost to start using generative AI in a small business?
Nothing, to start. Most of the tools above have usable free tiers or cost €20–30/month per seat at the paid level. The mistake I see most often is skipping the free trial stage and going straight to an enterprise contract before anyone has confirmed the tool actually fits the workflow. Prove it on the free tier first. Pay once you know exactly what you are paying for.
What industries use generative AI the most?
Marketing and media, software development, and customer service show the earliest and heaviest adoption, because the output in those fields is text, code, or images, exactly what generative models produce. Healthcare and financial services use it more cautiously, for drafting rather than deciding, because the accountability requirements are higher.
Do I need a developer to use generative AI in my business?
No, not for the examples in this post. Writing a good prompt, reviewing the output, and knowing when to ignore it is a skill any team member can build in an afternoon. You need a developer when you want to wire a generative model into your own systems via API: a different, more technical project with a different price tag.
Tijdo Koster
Automation consultant since 2009. 100–200 projects. Still answers his own emails.
Someone asked me last month whether generative AI would take their job. I told them it would take the part of the job they already hated, which is most of what a first draft is. They seemed relieved. Then they asked ChatGPT to double-check my answer. Cheeky, but fair.
There is more on the blog if you want to keep reading. And if you want the shortlist of tools worth actually paying for, the products page has the opinionated version.
Some links in this post may be affiliate links. Read the disclosure.