They were pleased with it right up until about four seconds ago, when they stopped looking at their own homepage and started looking at your face instead, and now everybody on the call knows roughly where this is heading without anyone having said a word.
I have found myself in that pause three times in the last month or so, and the situation has been close enough to identical each time that it has stopped feeling like coincidence and started feeling like a category of problem worth writing about. In each case a small team moving quickly has used AI to produce a brand and a website, and what has come out the other end is technically complete, perfectly competent, and completely anonymous. It could belong to any of two hundred companies in the same space, and in a couple of the cases it more or less already does.
The awkwardness of the moment is not really about the design at all. It is that you are being asked, in a fairly indirect way, whether the baby is ugly, and the honest answer, which nobody particularly enjoys giving, is that the baby is not ugly so much as it belongs to somebody else.
The averaging problem
It would be easy to take all of this as evidence that AI is no good at building websites, and I do not think that is right. It is very good at building a website, in the singular and general sense, which is precisely where the difficulty lies, because what you get back is a confident and well-executed composite of everything that has come before it.
You can recognise the anatomy of it from a long way off. There is a gradient hero with a headline about building the future, three feature cards with thin line icons sitting in a neat row, a logo strip, a testimonial carousel that may or may not contain real testimonials, and somewhere around the fold a sentence about being built for teams who move fast. None of it is wrong, exactly, but all of it is the average of the training data, which is another way of saying it is the average of your entire competitive set, delivered at a level of polish that makes it quite hard to push back on.
That is worth sitting with for a moment, because it is not a flaw in the tool so much as the tool doing exactly what it was designed to do. A model trained on everything will hand you the middle of everything, so if the whole point of the exercise is to occupy some ground that nobody else is standing on, then giving the job to an instrument whose fundamental operation is averaging is a slightly odd choice. It will take you unerringly to the centre of the room, and the centre of the room is the one place you cannot afford to be.
Why sensible people end up here
I want to be careful about how this sounds, because in my experience it is almost never a failure of judgement on the founder's part.
The people doing this are usually pre-seed or somewhere around a Series A, with a small team, a smaller budget, and a founder who is doing five different jobs before lunch. From where they are sitting, an agency engagement looks like time they have not got and money they would rather put into engineering, and meanwhile there is a tool sitting right there promising to collapse eight weeks of work into an afternoon. Under those conditions, reaching for it is not naive at all, it is just arithmetic.
The trouble is only that the arithmetic is quietly optimising for the wrong variable, because what these tools reliably produce is something that looks finished, and finished is a genuinely useful property when you are staring at an empty domain and a launch date you have already told people about. Looking finished and looking like you, though, are two quite different things, and the gap between them tends not to become visible until much later, usually at the worst possible moment.
The moment it costs you. That moment is normally the raise.
An investor meeting a company for the first time has two artefacts to go on before they meet an actual human being, which are the deck and the website. They are not assessing either of them as design objects and nobody is going to comment on your typography, but they are absolutely reading them as evidence of how clearly you think, which means they are looking for the positioning, the category, who this is for and why it exists now, and they expect to find all of it inside about ten seconds, mostly from the words.
A generic site fails that test in a specific and expensive way, because it does not say the company is bad, it says the company has not decided yet, and an investor who arrives at that conclusion has just been handed a reason to deprioritise you that has nothing whatsoever to do with your technology, your traction or your team. That is a maddening way to lose a conversation you should have won.
The version I find genuinely hard to watch is when the business underneath is properly differentiated, so the thing they have built is unusual and difficult and defensible, and then none of it survives contact with the homepage, because the homepage was assembled from the average of a lot of companies doing far less interesting work. You cannot be wrong in an interesting way once you have been averaged, and there is no white space to be had in the middle of the room.
Doing it in the right order
None of this is an argument against using AI, and I would not want it read that way. It is an argument about sequence, and the sequence matters a great deal more than the tooling does.
The work that has to come first is the work that cannot be averaged, because it consists entirely of decisions. That means the positioning, so where you sit in the market and against whom, and the narrative, by which I mean the actual story of why this company exists at this particular moment, and then the tone of voice, and a messaging framework so that everybody from the founder to the first sales hire is reaching for the same language, and the brand values, which sound like the soft part right up until you notice they are the thing keeping a hundred later decisions consistent without anybody needing to convene a meeting about it.
Only once all of that exists does the visual identity have anything to respond to, and at that point designing it stops being a matter of taste and becomes a matter of translation, because you now know what the thing is supposed to feel like, who it is meant to appeal to, and what it must not look like. That last constraint is usually the most valuable one in the room.
Where AI earns its place
Which brings me to the part that founders reaching for these tools have got right, just slightly early.
Once the strategy and the identity exist, and once that identity has been built out as a real design system with tokens, a type scale, spacing rules and properly constructed components, AI becomes the most useful thing in the studio. You connect Figma to Claude over MCP, point Claude Code at the system, and from there you can generate production components, page layouts and campaign variations at a speed that still feels faintly unreasonable, all of it on brand, because the model is now working inside a set of constraints that describe you and nobody else.
The model itself has not changed in the slightest. What has changed is what it is averaging over, so before the strategy work it averages the internet and gives you everyone, and after it, it averages your system and gives you more of you. Same tool, opposite result, and the only variable is what you chose to do first.
That, in the end, is the whole argument. Use AI to move quickly in a direction you have already chosen, and do not use it to choose the direction, because it will always choose the middle.
Ian Howarth
July, 05, 2026
A few things designing DX Motorsports taught me about building a brand when you’ve got no history to lean on.
