A marketing lead showed me a blog cover last year that she was about to publish. Clean composition, good lighting, a team of four around a laptop. On the shirt of the person nearest the camera was a logo. Not her company's logo. It was close enough to a real software brand that I recognised it, and the model had invented it on its own.
Nobody in the review chain had read the image. They had looked at it, which is different.
That is the whole story with AI image generation for B2B teams. It is genuinely useful, it is cheap, and it will confidently hand you something broken with a straight face. The teams that do well with it treat it as a production tool with a known failure rate, and they check for those failures every time.
Where AI images genuinely work
Three categories, and in these the results are good enough to stop debating.
Abstract and conceptual covers. Blog headers, social cards, section dividers, slide backgrounds. Anything where the image carries a mood rather than information. Generation is better than stock here, because stock gives you an image that four competitors also licensed and generation gives you something nobody has seen.
Backgrounds and texture. Gradients, geometric patterns, out-of-focus environments behind a headline. These almost never fail, because there is no anatomy and no text to get wrong.
Iterating on a concept before you spend money. This one is underrated. Before booking a photographer or briefing a designer, generate eight versions of the idea in twenty minutes and find out which direction you actually want. You are using the model to think, then handing the winning direction to a human. That is the highest-value use I see.
There is a fourth: internal material nobody outside the company sees. Deck backgrounds, workshop slides, internal wiki headers. Lower stakes, so a lighter check is fine.
The pattern is the same one that applies to generated writing. Volume is easy now and the check is the expensive part, which is why quality control on AI output is the piece worth building properly before you scale anything.
Where they fail badly
The failures are predictable, which is good news, because predictable failures can be checked for.
People, especially hands. Six fingers, four fingers, two left feet in the same photo, a wrist bending in a direction wrists do not bend. Models have improved a lot and they still produce these regularly. Faces fail differently: technically correct, subtly wrong, and readers register it as fake without being able to say why.
Text inside the image. Short words often survive. Anything longer degrades into shapes that resemble writing. Worse is invented branding, which is what caught out that marketing lead. A model will put a logo on a shirt, a sign on a wall, or a name on a laptop lid because logos appear in that position in its training data, and it has no idea it just put a competitor's mark on your page.
Anything specific. Your product's actual interface. Your building. Your team. A generated screenshot of software that does not look like your software confuses buyers and undermines the page it sits on.
Anything implying real past work. This one is serious enough to deserve its own section.
The claim problem nobody talks about
A photo on a case study page is a factual statement. It says: this is work we did.
Put a generated image there and you have published something untrue about your own business. It is a small lie and it does real damage, because the buyer who eventually sees the actual site, the actual product, or the actual facility notices the gap. What they take from that is not that your designer used a shortcut. What they take is that your evidence might be decorative.
Same rule for team photos, office shots, and anything showing your product in use. Use real photography, even when the real photography is a phone picture with imperfect light. An honest photo of the real thing beats a beautiful image of something that never happened. If you are building out case studies that buyers actually read, the images have to carry the same standard of truth as the numbers do.
The line is simple. If a viewer would reasonably assume a camera was present, a camera should have been present.
Keeping a generated set on brand
Generate twelve images across three months and you will have twelve images from twelve different companies. Model updates shift output, and prompts written from memory drift.
Fix it with a written style block you paste into every prompt. Not a vague mood word. Specific: the lighting, the colour range, the camera distance, the level of detail, what is never in frame. Something like "wide shot, cool blue and slate grey palette, soft overcast lighting, shallow depth of field, no people, no text." Reuse that block word for word.
Then generate the set in one sitting rather than one image at a time over a quarter. And before shipping, put all of them on one screen together. Inconsistency is invisible one image at a time and obvious side by side.
What it actually saves, and what it does not
The pitch for generated images is speed and cost. Both are real, and both are smaller than people expect once the work is done properly.
What you genuinely save is the search. Finding a usable stock photo for an abstract concept can take longer than the article did, and the result is usually an image three competitors are also using. Generating gets you something specific to your point in a few minutes.
What you do not save is the iteration. The first output is rarely the one you ship. You adjust the prompt, regenerate, decide the composition is wrong, and go again. Teams that report big time savings are usually accepting the first result, and it shows on the page.
Then there is the work after generation, which nobody counts. Cropping to the aspect ratio your template needs. Converting to WebP and compressing so the page still loads fast, because a generated image arrives as a multi-megabyte file and shipping it raw undoes any performance work you have done. Writing genuine alt text. Checking it against the brand.
Budget the same production time you would give any other image and treat the speed gain as removing the search, rather than removing the job. Teams that plan it that way stay happy with the results. Teams that planned for images in two minutes end up shipping whatever came out first.
The review step before anything ships
One person, every image, same list. It takes about a minute.
- Count the limbs and the fingers. Look at every hand. Confirm a pair of feet is one left and one right.
- Read every word in the frame. Signs, screens, labels, clothing, packaging. If it is not real language you chose, regenerate.
- Hunt for logos nobody asked for. Shirts, laptop lids, walls, mugs, equipment.
- Ask what the image claims. Does it imply a place, a team, a client, or a piece of work that does not exist?
- Zoom to full size. Failures hide at thumbnail size and appear on a desktop screen.
Name the reviewer. When review is "everyone", everyone assumes someone else did it, and that is exactly how the invented logo reached a publish button.
Disclosure and what buyers care about
Buyers do not care that your blog header was generated. Nobody has ever chosen a vendor based on the provenance of a background image, and labelling decorative art draws attention to something irrelevant.
They care when the image is evidence. A generated photo presented as your team, your office, your product, or a client's result crosses from decoration into a claim. Label it or replace it.
Regulation is moving in the same direction, so the practical position is to build the habit now: decorative goes unlabelled, documentary is either real or disclosed. That rule will still be right when the rules catch up.
The takeaway
AI image generation is a good tool with a known failure list. Abstract, conceptual, and background work is solved. People, text, specifics, and evidence are not.
Write one style block and reuse it. Name one reviewer. Read every image before it ships instead of looking at it. Do those three things and generation saves you real money without ever putting a stranger's logo on your homepage.