Why AI packaging briefs often miss the point

In short: AI-generated packaging briefs tend to specify tolerances and construction detail nobody asked for, while leaving out the budget, market and sales channels a supplier actually needs to recommend the right route. AI can help organise a brief. It can't be trusted to specify how something gets built, and trusting it over expert pushback is where most of the real damage happens.

I've been doing this a long time now, long enough that I can usually tell within the first couple of lines of a brief who's actually thought it through and who's still working it out as they go. That used to be the whole skill in reading one. These days it's a bit different, because more of what lands in our inbox hasn't really been thought through by a person at all, not in the way it looks like it has.

We get sent a lot of packaging briefs now that have clearly been put together with AI, and the honest truth is most of them have a problem somewhere. Some of those problems are minor and would take five minutes to sort out, if the person on the other end is open to hearing it. Most aren't minor.

Most arrive dressed in enough confidence and detail that it looks like the thinking has already been done, when really none of it has been checked by anyone at all.

That's not to say AI is useless here. It isn't. Used properly, it's a genuinely good tool for organising your own thoughts, taking scattered information and turning it into something linear and easy to follow, provided you already have the knowledge behind it.

You wouldn't use AI to design a car from scratch. For that you need a mechanical engineer, someone who understands how the thing actually has to work, and AI can help that engineer structure their thinking, but it isn't the engineer. Packaging works the same way. AI can help you organise what you already know about your product. It isn't, and isn't close to being, the authority on how that product actually gets built.

What does a packaging brief actually need to give a supplier?

Every brief we've ever had a good working relationship start from has the same thing in common, and it's got nothing to do with dimensions.

A packaging brief needs to explain what the product is, its budget, its market and how it will be sold, not specify how it should be manufactured. That's the information we actually need, because it's what lets us recommend the right route. We've written before about how to actually put a brief together, and none of it involves knowing anything technical.

What AI tends to generate instead is the opposite: enormous technical precision on things nobody asked it to specify, and comparatively little on the things that actually matter. That happens because a spec full of numbers looks thorough in a way that "here's our budget and who we're selling to" doesn't. AI is good at producing the appearance of rigour, and precise-looking numbers are the easiest way to manufacture that appearance, whether or not the numbers mean anything.

So a brief comes to us stating exact tolerances and dimensions to a fraction of a millimetre for details that were never the client's to decide in the first place, while leaving out the one or two things that would actually let us give useful advice.

Why does over-specifying a packaging brief cause problems?

It sounds like it should be the opposite problem, more detail helping rather than hurting, but it rarely works out that way in practice.

Over-specifying a brief backfires because it assumes any quote matching the stated parameters must be correct, when the parameters themselves are often wrong. A supplier can either quote to exactly what's been asked for, or quote something different, and either way it's entirely possible neither is the right solution, because the parameters were never sound to begin with.

What should happen is a genuine back-and-forth, where someone who actually does structural design for a living can say "that's not quite right, here's why, here's what would work better." A hyper-detailed AI brief tends to remove the space for that conversation before it's even started, because it presents itself as already decided.

We get asked to be someone's "packaging partner" often enough that I suspect it's become a bit of a buzzword. What's odd is how often that's followed by the opposite of what a partner relationship actually looks like. Instead of asking their partner for advice on the things they don't know, people arrive having already stipulated the technical parameters themselves, despite having no real knowledge of what they're specifying. You could hand over a child's drawing of the box you want, with all its lovely imaginative flourishes, and it would tell us about as much. At least nobody mistakes a drawing like that for an engineering document.

Does AI actually know what packaging should look like?

You'd think, at the very least, that a picture couldn't lie. It can, just as easily as a spec sheet can.

AI does not reliably know what packaging should look like, and can generate visuals that combine features from different construction types in ways that don't work as one real object. It isn't just construction AI gets wrong, in other words. The look of the thing is just as unreliable.

Ask it to generate a type of packaging from a name or a description, and it can produce something that borrows the right visual language from several different directions at once, a closure feature from one kind of box, a mechanism from another, a proportion from somewhere else again, none of it reasoned through as a single thing that has to open, close and hold together. It looks like packaging in the way a plausible-sounding sentence in an unfamiliar language sounds like it means something. It doesn't have to be right to look convincing.

That's the same problem as everything else in this post, just showing up before a spec has even been written. Someone who knows packaging looks at a generated visual and can usually see within seconds that the pieces don't actually belong together. Someone who doesn't have that experience has no way to know, which is exactly why it matters. The reference material for a brief should come from packaging that's genuinely been produced, not from something generated to order.

Why do people trust AI over an expert who disagrees with it?

If I'm honest, I don't think this is really an AI problem underneath it all. I think it's an older one that AI has just made easier to fall into.

People increasingly trust AI's confidence over expert pushback because AI never hesitates, and admitting you don't know something has quietly become something to avoid. Fewer people seem willing to ask the obvious question anymore, the one that risks looking like you don't already know.

I've never minded being the person in the room who asks it. I'd rather understand how something actually works than look like I already did, and more times than I can count, someone's come up to me afterwards and admitted they were glad I asked, because they didn't understand it either. They just hadn't wanted to be the one to say so.

AI makes that easier to avoid altogether. You don't have to ask the room, or admit you don't know, you just ask AI and get an answer that sounds sure of itself either way. That's really the mistake underneath all of this. It isn't that AI got something wrong. It's that nobody was willing to be the one to ask whether it might have.

What's the one question worth asking AI before you brief a supplier?

There is a way to catch most of this before it ever reaches someone like us, and it takes one sentence.

Ask AI: "Which parts of this brief are facts I have supplied, which parts are assumptions you have made, and which technical details need checking by a packaging specialist before I ask for prices?"

That one question does more to catch a bad brief than anything else in this post, because it forces a separation between what you actually know and what's been quietly filled in on your behalf. It won't turn AI into a packaging engineer. It will stop you handing a supplier a document that sounds finished when it isn't.

AI has a genuinely useful place in putting a brief together, organising ideas, tidying up messaging, getting a document into some kind of shape. It has no reliable place in deciding how something is actually constructed, what it can be made from, or whether the numbers on the page are physically possible together. That part still needs someone who has built the thing before.

Skip that step, and you're not saving time. You're just moving the point where the problem gets found, later, more expensively, and usually by someone else.