Everyone's feeding interview transcripts to AI but often in a way that throws the best stuff away. Here's what I've learned from a month trying to create the best way to mine internal interviews with AI for brand strategy work.

It's easy to paste an interview transcript into your LLM of choice and ask it for the themes. It hands you a tidy list in about four seconds. It looks useful.  But most of the time it quietly bins the most valuable things you need to build a brand strategy from.

I've spent the last month building a pair of AI prompts to mine internal interview transcripts properly, as one piece of the set of AI tools in Brand Strategy Academy. I say a month, but in reality, it took twenty years, because the prompt is nothing more than the way I've analysed interviews my whole career, captured in a way that AI can use it. Hundreds, if not thousands, of internal interviews for global brands across multiple industries has taught me what matters in a transcript and what's noise. The month was the part where the strategists in our Brand Strategy Academy community tested it on live client projects, told me where it missed something, and got me to fix it.

The prompts live inside Brand Strategy Academy, and they took too long to create to give away in a blog post. But the reason to keep reading is the thinking behind them, and I'll give you plenty of that below, because the thinking is where prompts like these go wrong. You can use it whatever you're building and whatever tool you reach for.

You have to know what you're looking for before you write a single word

This is the one thing nobody wants to hear. A prompt is only ever as good as the strategist sitting behind it. If you don't know what you to extract to help you create a brand strategy, the AI will give you a generic theme list and you'll nod along, because you have nothing to measure it against.

I know exactly what I'm hunting for in a transcript, because I've always got an eye on the prize - the four questions you've got to answer in any brand strategy and the fodder you need to extract from each stage of research that you built the answers from.  Every instruction in the prompt is built to extract the fodder that feeds the strategy. The AI isn't doing the strategy. It's doing the donkey work of finding the raw material, fast, at a scale I couldn't manage by hand. The judgment about what to look for is mine, and it has to be yours. No tool can supply it for you.

Sort everything into buckets of 'fodder'

Once you know you're feeding those four questions, the next question is where to look. Over twenty years and hundreds of interviews I've landed on 13 buckets of themes: the areas that, time and again, give you the richest fodder to build a strategy from. Things like what people are proudest of and would fight to protect, the job they believe they do for customers, how the organisation works when it's at its best, and the stories that prove a value instead of just naming it.

To give you a feel for how specific each one has to be, here's one of them word for word:

BUCKET 11: WHERE WE NEED TO CHANGE OR GROW (the forward-looking signal) This bucket is different from the others. It captures where people indicate something needs to shift, improve or be added for the organisation to reach its goals. Look for: frustrations with how things work now, behaviours people say are holding things back, capabilities or mindsets people say are missing, and things leadership describes as necessary for the next stage that don't exist yet. Listen for language like "we need to get better at", "the challenge is", "what's missing is", "as we grow we need to", or "the old way was". This is about honest tension, not aspiration statements. It tells you what the strategy needs to stretch toward, not just reflect back.

Notice how much is packed into a single bucket: what to look for, the exact phrases to listen for, and a clear line on what doesn't belong in it. Vague prompts get vague output. All 13 are tuned like this, and the tuning is most of the twenty years.

Build the prompt backwards from what wins in the boardroom

Ask yourself what lands a strategy. It's never just a list of common themes. It's often a real quote, in the employee's own words, and the specific story that proves out why a particular recommendation is right better than any adjective can. "We're caring" lands flat. Explaining the time someone drove three hours on a Sunday to fix a client's problem brings life to the idea.  Finding phrases like 'Be a cereal entrepreneur' (as Airbnb have in their values - no this is not a typo) creates a much more ownable and distinctive strategy.

So the prompt has to be obsessive about real language. Capture the exact words, never tidy them up, attribute them properly, and never let the tool invent or round up a quote to make a point land more cleanly. One fabricated quote in a client deck costs you all your credibility. If you build your prompt backwards from the quotes and stories you'll stand up and use as well as forwards from "summarise this," you end up with something far more useful.

Build in checks and balances, because the AI won't

AI has predictable ways of going wrong on a job like this, and a serious prompt has to defend against every one of them. Left alone, a model over-weights whatever's vivid or whatever it read last, it tends to treat every point as equally important, and it'll happily round a half-remembered quote up into something cleaner than what was said.

So you have to build the guards in. My favourite is the raw frequency pass. Before the tool is allowed to sort anything into buckets, it has to read every transcript once and simply count what recurs, in people's own words. Why first? Because the moment you start bucketing, you bury the themes spread thinly across many topics. A point made once in five different conversations is one of the strongest signals in the whole set, but it never tops any single bucket, so sorting alone makes it disappear. The frequency pass catches it before that can happen.

There are more of these guards: a weighting step so you can see at a glance what's loud versus what's niche, a hard rule that the tool can't keep a story unless it can quote the exact line it came from, and a final flag on anything that's internal belief dressed up as outside fact. These were painstaking to create and came from me working backwards from great presentations and making sure that the tool captured the insights that made those presentations land. "Nope -you missed this idea so what do I need to build into the prompt so you don't miss things like this...." Lots of chats back and forth. All of them are the difference between output you can trust in front of a client and output that just looks tidy.

Split the capture from the analysis. That's why there are two prompts.

Here's the structural decision that took the most testing to get right. If you ask one prompt to both extract everything and analyse it, it gets worse at both. It rushes the extraction to reach the clever-sounding synthesis, it flattens the detail on the way, and by the time it's "analysing" it's really just re-reading its own summary instead of going back to what people said.

So I split it in two. One prompt captures and sorts everything, carefully, and then stops. A human reads it. A second prompt does the pattern-finding. The gap in the middle matters as much as the prompts do. It's where you catch the misattributed quote before it hardens into a false finding, and it's where you, the strategist, put your own eyes on the raw material before any conclusions get drawn from it. It's where you need to compare the notes you took, the gut feeling you have, the themes you thought use to the top with what the prompt is saying. This is exactly how you'd work through the process before AI: interview first, analyse second. Good tools respect that order. They don't collapse it.

It's a check, not a replacement

As a strategist we're paid for our thinking and our judgment. Not to run AI and hand over whatever it produces. The tool doesn't sit face to face with people doing those interviews. It doesn't clock the hesitant pause before someone answers, or the thing they almost didn't say, or feel the vibe change when you hit a nerve.  It doesn't capture the passionate telling of a story vs the reciting of one that's just old hat. The nuance, the feeling, your gut on what might be an interesting area to explore.  You've been honing that for years. Don't throw it away now.

The most useful way to use any of this is as a second set of eyes on work you've already done in your own head. Do your own read first.  After each interview capture the things that really struck you.  Rant into an AI tool to do that capture if it's easier. But don't just sit back, read the questions like a robot and not really think and listen because you know you have transcripts and prompts to use. Then let the tool catch what you missed and challenge where you landed.

And here's the part most people have backwards: the value is in where it disagrees with you, not where it agrees. Don't scan the output hunting for confirmation. Go straight to the places where it weighted something differently than you did, and work out who's right. Sometimes it caught something you missed. Sometimes it flattened a nuance you heard live. Either way, that's telling, and that's the bit only you can do.

The other very important check to do right at the start is the permission check. Here’s a clause you can add into your contracts to make sure you have client permission to use AI in this way in your brand strategy work.

In summary: how to use AI to really help in brand strategy work

A good prompt isn't a clever paragraph you can lift off the internet. It's a proven process, encoded. The two interview-mining prompts, the way they weight what's loudest, the discipline that stops the tool making things up, and the rest of the AI toolkit I'm building into Brand Strategy Academy all sit on twenty years of doing this work the slow way, and getting better at it over time, plus a month of a smart community stress-testing it on real client work. 

If you want to do brand strategy with AI properly, that's the part worth investing your time and money in. Not the prompt itself, but the twenty years of knowing what to ask it to do, and what to leave out. And what you do with it next to create a strong brand strategy. That's what I teach inside Brand Strategy Academy.