Every Monday morning, I spend a couple of hours going through what founders built over the weekend with Lovable or v0. They often send over screens that indeed look impressive, but at first glance. Then I start clicking around, and this is when the illusion breaks.
AI design tools work well for what they do. But the problem I keep seeing is that some founders frequently confuse screen rendering with product design. These two are completely different activities, and there’s a high probability that mixing them can cost teams real money.
How a Vibe-Coding Session Actually Goes
Prompt one: "Create a dashboard for logistics management." The result actually looks good.
Prompts two through five: you refine the layout, add filters, adjust charts, ask for a more unique look.
Then prompts six through twelve arrive: "Why are the filters hidden behind the table? Undo that. And make the table wider." At this stage, you typically stop building and start firefighting.
By the time you've burned through hundreds of tokens chasing your own earlier instructions, the session ends with a result that a junior designer could deliver in half a day. This is what vibe coding actually looks like in practice.
Add it up: $50–200 in Lovable credits, plus 8–12 hours of founder time at a typical rate of $75–150/hour. That's $600–1,800 in real cost for a result that a junior designer would deliver in half a day for roughly $100–200.
The tool that was supposed to save money ended up being the expensive route.
What else is worth mentioning is that burning through tokens is actually a symptom of a broken AI workflow. The root problem is that AI has no long-term vision for your interface architecture, and every prompt is handled locally. When you ask it to move a button, it moves the button. But AI has no way of knowing that the button was visually tied to an element two columns over. You fix one point and break two others. This is the loop.
A designer, in contrast, looks at a screen as a living system where each element is connected to several others. Before moving a single element, they ask: Does this need to exist at all? This question is the foundation of product UX, and it is never asked when you're prompting screen by screen.
Data Visualization Gap
AI reads the data it's given and draws a bar chart or line graph. It doesn't think about what question the user is asking when they look at that data, or which visualization format actually answers that specific question.
The prompt didn't communicate the idea in the specialist's head, let alone what a regular user needs. That's the core difference in how we think. AI looked at the data type (percentages) and picked a visualization to match it. But a designer also looked at the question the user asks when they see this data: "How complete is this?" AI interpreted the question as "what does this consist of?"
So the answers look different and solve different problems.
Cost: anywhere from $0 if the designer catches it in time.
In our case, it was $0, because we caught it ourselves. But that's not always how it goes.
What Gets Lost Between the Prompt and the Screen
AI UI generation is fast. The speed is also what makes it easy to skip the question of whether the screen you just generated actually solves the right problem.
But when you type a prompt into a UI generation tool, you receive pixels that resemble a product. Generative UI gives you the output of a decision process, without the decision process itself.
What you don't get is the layer of thinking that happens before a designer opens their tool. This layer includes:
- Understanding what question the user is actually asking when they open this screen
- Mapping how this screen connects to the flow before and after it
- Identifying which elements are load-bearing for the user's task and which are decoration
- Anticipating how the layout behaves when real data replaces placeholder text
We had a SaaS UX project where AI generated a timeline interface: weekly status summaries at the top, detailed reports below. Summary first, details second, and this actually sounds logical.
Except when real users interacted with it, the logic collapsed. A user would open the screen to understand why last week's metrics dropped. They'd read the status summary at the top, scroll down to find the details, lose track of what the summary said, scroll back up, then back down.
A few people even started opening two browser tabs to compare information side by side. This kind of workaround is common with AI-generated interfaces, as users adapt to the tool's logic instead of the other way around.
When we redesigned the layout so the weekly status was located inside the column header of the detail table, the interface became one coherent tool, and the user could answer their actual question in a single view.
This is what AI UX design misses: the layout looked logical in the prompt, but only a real user session revealed how it actually behaved.
Fortunately, we detected this issue at the design stage, and it took us roughly 4 hours of rework at $50–80/hour.
If the same layout had reached development first, we'd be looking at 2–3 days of backend restructuring, and this is $800–2,000 added to the bill.
AI Can Also Break Down Brand Consistency
A client of mine, a founder with a product company, was traveling for two important sales meetings. The presentation for the first meeting was ready. For the second, he opened Claude on the flight and generated 50 slides.
Almost all slides looked this way:
When he landed and showed a designer, they went through the deck slide by slide, and almost every slide had the same problem: it looked like a generic corporate presentation pulled from a template library. The colors, fonts, and visual structure had no connection to the product's brand. A person in that meeting room would have had no way of knowing this came from the same company as the website, the product UI, or the materials from the previous meeting.
We rebuilt the entire presentation before the second meeting. Here’s the new design we came up with:
Ultimately, the client did win new partners from that trip. Whether a polished, on-brand presentation has made the outcome better, we can't say for certain. But we know what visual consistency does: it creates trust. When a company's materials look cohesive, people assume the company knows what it's doing.
AI had no access to the product's visual personality, its tone, its audience's expectations, or the specific fonts and colors those users had seen hundreds of times. Interface consistency is usually developed gradually, across every touchpoint, but AI starts from zero each time.
Visual familiarity builds trust because of a well-documented cognitive pattern in UX called the mere-exposure effect: people trust what they recognize. An interface that works correctly but looks visually foreign creates a discomfort that users feel but can't explain. This is one of the most overlooked problems in AI UX today.
The Investment Question
Some founders plan to skip design investment early and add it later when there's more budget. But the cost doesn't disappear. It just moves to a stage where fixing it is harder.
Reworking a vibe-coded interface that's already in development means going into a live system, understanding what every piece depends on, and changing the structure while the product is running. That's a different category of work from building clean from the start, and it's a recurring pattern in AI interface design projects. And engineering hours spent untangling AI-generated layout decisions can be some of the most expensive hours in a product budget.
Before computers, accountants did all their work by hand. When computers arrived, they didn't disappear. They got faster. UX designers are at that same point now. AI is the calculator. The judgment about what to calculate, and why, still has to come from a person.
And this is the gap AI product design tools haven't closed yet. The calculator got faster. The judgment still has to come from a person.
Budget spent on design before development is an investment. But the same budget spent after development is a repair bill. And this sequence is usually the same across every AI development workflow.
Where AI Genuinely Belongs in the Design Process
AI still has a real place in design work. It's good at one specific thing: executing a clear specification fast. For example:
- When a designer needs to explore five different layout directions for a component, AI can produce these starting points in minutes. The designer picks, adjusts, decides. That's a genuinely faster way to work.
- AI also handles repetitive UI work, like content variants, iterations on an established component pattern, and microcopy for buttons and labels. When the creative decisions are already made, AI is a good production tool.
- For founders, AI is useful for making a rough idea visible before a designer is involved. Used this way, AI prototyping shortens the briefing conversation without replacing the design work. Turning a vague concept into something you can point at and say "more like this, less like that" makes the briefing conversation much shorter.
- Data visualization: Ask AI to chart a dataset, and it picks a format based on the data type: percentages get a pie chart, trends get a line graph.
A Practical Approach If You're Already Using AI Tools
If AI is part of your process, here's how to avoid the expensive version of these mistakes:
- Define the user's actual question before prompting. What task are they completing? What decision are they making? Write it down.
- Treat AI output as a reference. Have a designer review the flow before it reaches a developer. Even a short review catches the critical mismatches between what the screen shows and what the user actually needs to do. This review step is the part teams often tend to drop from their UX workflows when they switch to AI-assisted design.
- Include your brand guidelines in the conversation. If you have them, attach them. If you don't have them yet, that's worth solving before generating 50 slides for a sales meeting.
- Track your iteration count. If you're on prompt 15 and still adding complexity, stop. That's a signal to step back and reconsider the structure. More pro mpt engineering won't fix a layout at this point.
Final Say
Designers used to hear "we don't need UX, just build it." Now, founders say, "I'll generate the design with AI and hand it straight to the developer." The excuse is different. But the result is still the same: product thinking is skipped, and the cost shows up later.
AI is a real tool, and it speeds up parts of the design process that used to take hours. But faster screen generation doesn't replace the thinking that impacts what these screens should actually do. That part still has to happen, and the earlier it happens, the cheaper the whole project will be.
Top comments (0)