When AI Generation Isn't Worth It

You usually figure out when AI generation isn't worth it after you've already burned half a day on it. The list of these tasks is short, and knowing it ahead of time beats learning ten more prompting tricks.

Showing the real product as it actually is

If you need an honest shot of your own item, a photo wins. Twenty minutes by a window gets you there; I've written up a simple home-shooting setup separately.

Generation earns its place here as a finishing step on a real shot, not as a replacement for shooting one.

Text inside a layout

Headlines, tables, prices, size charts. Numbers get garbled, non-Latin scripts get garbled even worse, and checking the result takes longer than just typing it out.

A layout editor does this in minutes, with no errors.

A diagram with real connections

Flowcharts, structure diagrams, process maps. The model draws something that resembles a diagram, not your actual logic. More on that here.

The mechanics of your product

How the lid opens, how the chair folds down, where a part is supposed to slot in. The model has no idea, and it will just make something up.

When the volume is small

One image a month doesn't pay back the time it takes to learn the tool properly. Ordering a shot or taking it yourself is simpler.

Generation pulls ahead once you're producing at volume, and the gap widens the more you make.

When you need expert judgment

What matters to a buyer, which objection needs addressing, what sets you apart from the competition in search results. A person who actually knows the market decides that.

The model will give you a shot. It won't tell you which shot you need.

When accuracy carries real consequences

Medical devices, food products, anything where a mismatch between the image and the real thing causes trouble beyond a return request.

The pattern behind all of it

The rule it all comes down to is simple: generation does well where you need many variations of a scene built around a clear object.

It does poorly where there's a single correct answer.

If a task has one right result and you already know what that result looks like, go make it directly.

I watched a team generate spec tables and then proofread every number by hand afterward. That took twice as long as just laying the table out in an editor, and nobody noticed for months.

One caveat: when I pointed it out, it turned out they actually liked the process. I didn't argue the point further.

FAQ

Where does AI generation lose to the alternatives? Showing the real product honestly, text inside a layout, diagrams with real connections, and demonstrating how a product's mechanics work.

Why not generate tables and numbers? Numbers get garbled, and checking them afterward takes longer than building the table in an editor from the start.

At what volume does generation start paying off? At scale. One image a month doesn't cover the time spent learning the tool.

Is there a general rule? Generation is strong where you need many variations of a scene, and weak where there's a single correct answer.

How do I tell a task isn't a fit for generation? If it has one right answer and you already know it, make it directly. There's nothing to vary.

Why doesn't generation pay off at low volume? The time spent learning the tool never gets paid back. At ten images a month, ordering or shooting is just simpler.

Sources

1. Flami: home photography without a studio
2. Flami: diagrams with AI
3. Flami: AI generation or a designer