What AI can and can't do in brand strategy

A founder sends a brand deck over before our first call. Forty slides, beautifully formatted. It takes about ninety seconds to work out that a model wrote most of it - not because anything in it is wrong, but because nothing in it is theirs.

The positioning statement could be pasted onto a competitor's deck without changing a word. The three brand values are the three brand values: some arrangement of authenticity, innovation and community. The tone of voice section says confident but approachable. The customer persona has a name, an age, a job title and no opinions.

None of it is a mistake. That is what makes it difficult to talk about. The deck is internally consistent, competently argued, and completely interchangeable - and the founder often cannot see it, because they have been reading their own vision into sentences that do not actually contain it.

So I want to be precise about where AI helps in brand work, because "use AI for speed and keep humans for creativity" is a slogan, not an operating model. It tells you nothing about what to do on a Tuesday.

The three layers of a brand

Every brand problem I have worked on lives in one of three places.

Vision is what the founder believes the brand is, means and promises. It lives in their head, in internal documents, and in the way they describe the product when nobody is taking notes.

Expression is how that vision becomes language, visuals, product decisions, pricing and channel choices. Website, copy, campaigns, packaging, ads.

Perception is what the customer actually receives. Reviews, comments, what they tell a friend, what they think they are buying.

The direction of travel is vision, then expression, then perception. Almost every brand problem shows up in perception and is caused in expression. That is why separating the layers is worth the trouble: it tells you where to look when something is wrong, rather than redesigning a logo because sales are soft.

It also tells you where a model can help, because the line runs along these layers, and it runs in one specific place.

What this looks like on a real brand

Vibes Shades sells fashion eyewear, which is close to the most commoditised category there is. When they came to us the brand was running on paid traffic, discounts and one-time buyers. The conversion rate was the symptom everyone was looking at.

The instinct in that situation is to write a new brand story. We did the opposite, because the vision layer is not something you author on someone else's behalf. We went reading instead - the old site, line by line, hunting for anything that did not sound like every other eyewear label.

One line did. Buried about three scrolls down a page nobody scrolled, it said that sunglasses should not have to carry the markup that retail shelves and distributors add.

That was the whole brand, already written, in the founder's own words. A second line sat in the homepage banner - Milan to Tokyo, Miami to Barcelona - which was there to say the brand ships internationally. Read differently, it described the person wearing the glasses rather than the logistics behind them.

Put together, that gave the position: high-performance eyewear without the retail markup, for a high-velocity life. The promise underneath it: same precision, same quality, no retail tax. Neither of those is a sentence a model would have produced, because neither of them is the average sentence for the category. Both of them were already in the building.

What changed was everything downstream. The markup argument moved to the top of the page and was repeated at the moments where price is actually decided. The navigation stopped using the category's default words: Shop became Access, Collections became Series, About became the Manifesto, the FAQ became a Signal Check. The specifications stopped being a table at the bottom of the page and started being the argument.

Notice what did not happen. Nobody invented a new vision. The founder's belief was already there and already specific. The job was finding it, deciding it was the one worth betting on, and then building an expression layer that carried it - and only that middle step, the deciding, is the part that cannot be handed over.

Vision: why this stays human

The first step in any diagnostic we run is a Decode - extracting the founder's vision in their own words, before showing them any data at all. The sequencing is not a formality. Once a founder has seen their own analytics, they start explaining their brand in terms of what the numbers seem to want, and the uncontaminated version is gone.

Here is what that produced with Mike Chu, who runs Tahan Outdoors, a Malaysian outdoor gear brand. Asked what he wanted the brand to become, he said:

> "I want to pull away from the conversation that is always about price, price, price, and go to the conversation about value, value, value - and I'm willing to pay this price because of the value that they give."

Then the last question we ask in any Decode, and we always ask it last: how would your dream customer describe this product to a friend, in their own words?

> "You should use this product because they really put thought in their design and their service is very reliable and the product quality is good. Don't go for the cheap stuff because you will end up buying a new one anyway. Might as well just invest in a better one and you get a better experience over the long run."

That is the brand. Not a positioning statement, not a value set - a real sentence a real person would actually say to another person in a car park. Invest in better. Put thought in their design. A better experience over the long run.

Then we went and read the website. The About page said the mission was making great outdoor products accessible and affordable. Elsewhere, fairly priced. The pop-up offered money off before a visitor had read a single line about the product.

Not one phrase from his answer appeared anywhere on the site.

The founder had the value argument fully formed and had been having the price conversation with his own customers for years, through his own expression layer, without noticing. That is a language gap, and it took forty minutes and one well-sequenced question to find. It did not take a model.

Three reasons the Decode is the part I will not delegate.

A model does not know what to be surprised by. The useful moment in a Decode is when the founder's voice changes. They get animated about a detail that is not on any slide. They get defensive about a decision nobody questioned. Their language stops sounding like marketing and starts sounding like a person. Catching that requires a baseline for this founder, built in the first ten minutes of listening to them. A model has no baseline, so nothing stands out.

A model normalises. Ask one to summarise a founder interview and it returns something tidy. Tidy is precisely wrong here. What you need is the odd phrase, the half-finished sentence, the bit that does not fit the category - because that is the part nobody else can say. Mike's answer would have come back as "emphasises quality and long-term value", which is true, useless, and would have fit any competitor on the shelf.

A model asks the average question. Its questions come from the middle of everything ever written about brand discovery, and so do its follow-ups. The Golden Question only works because the forty minutes before it were responsive to this particular person.

Transcription is fine. Interpretation is not.

Expression: where AI genuinely earns its place

This is the part where I am not precious at all, because the gains here are real and I use them daily.

Volume is the obvious one. To find three subject lines worth testing you want to look at thirty, and generating thirty is exactly the kind of work a model does well and a person does slowly. The same applies to ad variants, headline options, alt text across a catalogue, and turning one asset into six channel-specific versions.

Then there is structure. A first-draft outline, a competitor scan across twenty sites, a long document reduced to the four things that matter - all faster, all low-risk, because a human is choosing what survives.

But there is a condition, and it is the whole ballgame. AI works in the expression layer when it is handed constraints that came from the vision layer: a glossary of the words this brand uses and the words it refuses, a register, a list of claims that have actually been verified, and a banned list.

For Tahan that list is short and it is decisive. Use: invest in better, built for, designed for, trip after trip. Avoid: affordable, budget-friendly, save, don't miss out. Two of those banned words were the brand's own mission statement. Give a model that constraint and it becomes a fast, useful expression engine. Give it nothing and it returns the category average - which is the forty-slide deck at the top of this piece.

(Why the average is what you get, mechanically, is its own subject. I will take it properly in the next note.)

Perception: collected by machine, read by human

This layer is a clean split.

Models are very good at reading several hundred reviews, comments and survey responses, pulling out verbatim phrases and grouping them by theme. That used to be most of a day's work. It is now something you can run before lunch, and running it more often is straightforwardly better.

What a model will not do is tell you what the grouping means.

It will tell you that a particular word keeps appearing in your reviews. It will not tell you that the word belongs to a customer you did not design for, that your ad account has been recruiting more of them for eighteen months, and that the fix is therefore in acquisition rather than in copy. That is a diagnosis, and diagnoses have consequences - they decide what you stop doing.

Naming the gap is a judgment. Judgment is the job.

The working split

Read the human row across and you will notice it is all the same kind of work: deciding what matters. Read the machine row across and it is all production. That is the actual line, and it is worth saying plainly because the usual framing - creative humans, fast machines - puts it in the wrong place. Plenty of the human work here is not creative at all. It is just judgment, and judgment does not delegate.

The test

Here is the version you can run on yourself.

If every AI tool you use disappeared tomorrow, would your brand still know what it means?

If yes, you are using it in the expression layer. Carry on, and use more of it.

If no - if the only written articulation of what your brand stands for is a document a model produced - then the vision has been outsourced and you have a translation problem you cannot see yet, because there is nothing to translate from. Every piece of expression after that point is a copy of a copy, and it will surface in perception six months later as customers describing you in words you never chose.

The fix is cheap and slightly uncomfortable. Close everything and answer these in your own words, badly:

• What problem does this product solve, in your own mind?
• Who did you build it for - describe them like a real person, not a persona?
• What do you wish more people understood about it?
• What do competitors get wrong about this category?
• What have you tried that felt off-brand, and why did it feel that way?

Then put your answers next to your homepage. Tahan's founder had the answer to all five. His website had none of them.

We are currently building the full version of this as something you can download: the Decode prompt, the complete question set as a form you can actually sit down and work through, and the Gap Map sheet that turns your answers into a diagnosis rather than a pile of notes. The whole package, so a founder can run it on their own brand without hiring anyone to do it. It is not finished yet. If you want it when it is, leave your email below and I will send it.

If you ran those five questions on yourself this afternoon, which one would you struggle with most?