Can AI protect taste?

A lot of brand-led founders distrust growth advice, and honestly they've earned the right to—they've sat through the deck, nodded along at the promise, and then watched some well-meaning operator try to trade the actual soul of the thing for a slightly better click-through rate, which is the sort of thing that teaches you to keep one hand on your wallet the moment anyone says the word "optimise". So when AI wanders into that room the worry sort of writes itself, which is, well: is this just the same trade again, only faster now?

It can be. Bad AI will quietly sand every edge off the work and hand you back something smoother and emptier and call it a win because the metric moved, and I've watched it happen. But it can also do the opposite, which is the part I find interesting—it can protect taste by making the brand's own language easier to see, easier to actually use, and a lot harder to flatten when someone in the room decides it'd convert better as a beige version of itself.

Taste is mostly pattern recognition

Taste gets treated like some kind of mystery, and sometimes it genuinely is, but most of what we call taste is just pattern recognition that's been quietly running in the background for years—the founder who knows on sight which line feels wrong, the creative lead who can feel when the brand is trying too hard, the customer who can tell in about half a second which phrase sounds like a real person and which one sounds like a brochure got promoted. Same instinct really, just pointed at different distances.

And the thing AI turns out to be good at is collecting those patterns from the places the language actually lives, which is the reviews and the DMs and the support threads and the offhand community comments where someone describes the problem the way they'd describe it to a mate. That's the language people reach for when nobody's asking them to be impressive, and it's worth more than almost anything a workshop will ever produce—which I say having sat through a few workshops.

That's voice-of-customer work done properly. Not the sterile version where a chart announces that women 25-34 like convenience, but the useful version where you get to read the exact words people reach for when they're describing the hesitation, or the relief, or the small delight, or the thing they can't stop telling people about.

A brand's defence is easier with evidence

If your job is protecting a brand then evidence helps—not because data should get to overrule taste, it shouldn't, but because taste holds up a lot better in a room when it can point to receipts. A Voice-of-Customer Engine mines the reviews and DMs into themes, personas and message angles, and then tests those themes against order history and segments and campaign performance, so you can actually see which customer language maps to real paying customers rather than to a persona somebody named after a colour from a slide deck. I've done this across fourteen-odd brands and the words that move money are almost never the words the brand thought it was selling on.

And what that really buys the operator, when you get down to it, is a better argument. "Our customers don't talk like that." "This angle works, but that one phrase kills the feeling." "We can lean harder on this benefit without sounding like everyone else in the category." Those are much easier sentences to say out loud when there's something underneath them.

The job is to extend the human ear

The point was never to have AI invent the brand. The brand already exists—it's in the founder's taste, and the product, and the customers, and the weird internal rules nobody can quite justify, and the one sentence someone flatly refuses to change because it's somehow load-bearing. None of that needs inventing. It just needs listening to, at a scale a single human ear honestly can't manage on its own, so the team gets to see the patterns without ever once mistaking the patterns for the whole truth.

In practice that's a short list of pretty unglamorous jobs:

• mine reviews for repeated phrases and objections
• cluster DMs into real buying moments
• turn customer language into ad angles
• check whether a draft sounds like the brand or like a committee wrote it
• give the team a searchable memory of what customers actually said

The human still makes the call. All the machine is doing is making sure the call is a well-informed one—which is the only version of this I'm interested in building, partly because I've watched the other version go wrong. I once let a broad-match test off the leash on my own account and it cheerfully went and bought traffic I'd never have chosen in a hundred years. The machine is brilliant at scale. It has no taste at all, and you don't really want it making the calls that taste is for.

This is the part I actually care about. Performance marketing has this habit of grinding every brand down until they all sound the same, and it doesn't have to—used properly, AI just gives the operator more ways to hold onto what's distinct while still doing the unglamorous work that growth quietly demands of you anyway. A brand with a real voice should be able to carry that voice across four thousand product descriptions without it thinning out somewhere around the eight hundredth one. That's exactly the sort of job a human ear can't sit through and a machine can. The trick is that the human stays the one deciding what good sounds like.