How to Do Customer Research in the AI Era: 16 Rules

The Mental Asterisk 

Q: What can AI not do in customer research?

A: It can draft the questions and process the answers. It cannot get a stranger to tell you the truth, or read the feeling behind an answer once you have it. 

AI is everywhere. If you are trying to get away from it, unfortunately for you, everywhere includes this piece. It also includes LinkedIn, where the odds are that the first item you scroll past is about AI, and if it is not about AI, it was written by it. You can usually tell. Something will be quietly compounding, and every paragraph will end on a punchy aphorism that takes a block of text you thought you understood and turns it into a vague, supposedly revolutionary idea. To be fair, sometimes those vague revolutionary ideas are pretty revolutionary. Most of the time they are just confusing.

It's exhausting. When I can tell something was made by AI, I add a mental asterisk to everything I learn from it. Not because the machine is stupid, but because I cannot see what went in. When I use AI myself, I know which documents it read, which numbers I checked, and which ones I waved through because it was late. I don't know that with other people's outputs, so I find myself scrutinizing every word.

So how do we make work that doesn't exhaust the people who are supposed to value it, and that stands up to a reader's AI-induced scrutiny? By understanding your customers better than ever. My name is Lamar, and I do search, content, and strategy research at Kalungi, a marketing consultancy for B2B SaaS companies (which, ironically enough, are about as omnipresent in business literature as AI). I have run customer research at every scale, from Fortune 500s to bootstrapped companies with two employees, a dream, and hopefully, some air conditioning.

The practical argument of this piece is short: AI made everything around research cheap, which makes the parts that are still expensive, the parts where a real person decides whether to tell you the truth, the most valuable skills in marketing. The practices below are about doing those parts well.

But first, some French jazz.

The case of the impossible jazz

I have lately been listening to a great deal of 1940s French jazz. Quirky, I know. I didn't pick it; I'm not that cool. YouTube's algorithm led me there after I let some background music run too long, and I stayed because it is really, really good background music. It is also AI generated. None of the songs are real. The comments are split between people praising the composition and people warning everyone else, and I kept listening anyway, which is its own small embarrassment. I had been led to music I never chose, that does not exist, and worst of all, I love it.

I assumed the premise was as fake as the songs. Jazz, in France, in the 1940s, a country occupied for half the decade and rebuilding from rubble for the rest? I went looking, mostly so I could dunk on the machine. It turns out jazz boomed in occupied France. Django Reinhardt's ensemble sold roughly eight times more records than it had in 1937. A hundred and twenty-five new cabarets opened after the Armistice. A jazz festival in December 1940 sold out, and eighty more concerts followed before the liberation. French fans protected the music by loudly insisting it was a French invention rather than an American one, and the occupiers tolerated it because propaganda goes down easier wrapped in popular songs.

So the machine had not invented an aesthetic. It had found one, muddled between decades of real recordings, real history, and the accumulated wanting of thousands of people who typed some version of “relaxing jazz” into a search bar without knowing what they were reaching for. The mix is fake, but the desire underneath it is real. That underneath layer, the recordings, the history, the wanting, is what we at Kalungi mean by signal, and the entire difference between AI output that works and AI slop is whether there is any of it down there.

(If you're curious about the jazz mix itself, here is the YouTube video).

What AI changed, and what it did not

Q: If AI can run a survey end to end, why learn research skills at all?

A: AI made drafting, cleaning, and summarizing cheap. It did not make anyone more willing to answer honestly, and it cannot tell whether the person answering felt secure, pressured, or checked out, which is often the finding.

The cheap parts first: drafting a survey is quick now, and so is the invitation email, the data cleaning, the coding of open text, and the deck built from all of it. The asking did not get cheaper. Somebody still has to open the email, and finish the survey instead of abandoning it at question fourteen (which is where everyone was abandoning my last one). Somebody still has to take the call, warm up, and decide around minute eight whether to give you the polite answer or the real one. If anything the asking got harder, since cheap production floods every channel through which anyone asks anyone else anything.

I am aware it is convenient for a person whose job is research to conclude that research is the irreplaceable bit, and to be fair, a model writes perfectly reasonable questions. What it cannot do, at least not yet, is read the humanity behind the answer. Does this person feel secure in their job? Are they under pressure from someone in the company? How much energy did they bring to the call? The key word is feel: as long as people buy with emotions, and marketing remains an emotional field, those feelings are the material. A researcher's job is to pick them apart, record them, and use them to supplement the more technical, straightforward information that drives strategy.

7 Rules for Interviews

Q: What makes a good customer interview?

A: In short: ask about things that already happened, in plain words, and feel comfortable to end on silences and awkward notes. Discount compliments, never say a price or a competitor's name first, hunt for whatever contradicts what you walked in believing, and write the synthesis the same day. Aim to let the customer talk four-fifths of the time.

Interviews come before surveys. Surveys look like the cheap option, which is why the order typically gets flipped in practice, usually by somebody staring at a timeline. But closed questions are just hypotheses with answer options bolted on, and if you have not heard a real person describe the problem in their own words, you are asking a hundred strangers to pick between your guesses, which are probably wrong.

One thing before the list: none of these rules are really about questions. They are about the fact that the person across from you is polite, and politeness produces answers structurally identical to real ones. “That would be useful” and “that would be useful” look the same in a transcript, but only one of them means somebody will eventually pay money, and you have to figure out which one it is.

1. Anchor every question in something that already happened. “Would you find this useful” buys you a courtesy. “Walk me through the last time you had to do that” buys you a story, with details attached that you did not know to ask about. People are reasonably accurate about their past and hopelessly optimistic about their future, so ask about the past.
   
   
2. Ask plainly, in words they already use. If someone has to interpret your question before answering it, part of the answer is about the interpretation, and you will never know which part. This is my most frequent mistake, and it comes from wanting to sound thoughtful.
   
   
3. Let the pause finish. Four seconds of silence and the urge to rephrase becomes almost physical. That pause is the person deciding how honest to be, and filling it decides for them. Count to seven. It feels much longer than it is.
   
   
4. Compliments are not data. What counts is anything that costs something to give: an unprompted story, visible animation about a workaround, an offer to introduce you to someone. When you get a compliment, ask for a specific example instead of writing it down. Sometimes one arrives, sometimes the conversation moves on, and both tell you something.
   
   
5. Never introduce a number or a name. Say a price and every answer after it is calibrated to that price, permanently. Name a competitor and you have put them in the customer's head. If you must prompt, prompt at the end, and record that you prompted, because a name they raised and a name you handed them do not belong in the same column.
   
   
6. Try to leave with something you did not believe walking in. If everything confirmed what you already thought, the likeliest explanation is not that you were right. I run this check hardest on myself, since the interviews that feel best are usually the ones where I got exactly what I wanted to hear.
   
   
7. Write the synthesis within a day. The transcript keeps; the texture does not. What surprised you, which answers came slowly, where the energy went flat. Pull three to five exact quotes while you can still hear how they were said, before memory starts tidying them into better sentences than the person used.

9 Rules for Surveys

Q: What makes a good survey?

A: Every question serves a decision you can name, and gets cut if it cannot. Keep it short, take it yourself on a phone before anyone else does, match the method to the sample you can really reach, choose incentives on purpose, record what every number is measured against, and when the sample is small, report direction rather than magnitude.

A survey is the less forgiving instrument. In an interview you hear a question land badly and repair it on the spot. A survey goes to everyone at once, and if something is wrong with it, you find out afterwards, from data that looks entirely normal.

1. Give every question a decision to serve. Name what you would do differently depending on the answer; if you cannot, cut the question. This removes about a third of most drafts, and everything it removes was costing completion rate in exchange for nothing.
   
   
2. Confirm somebody still wants the answer. This belongs first and gets checked last, if at all. I once built a survey that was designed, reviewed twice, rebuilt in a day when its logic broke, approved, and never fielded, because by then the question it answered was no longer one anybody needed answered. The instrument was fine. Check that the decision is still open, and that whoever owns it will still be in the job when results land.
   
   
3. Length is a data quality decision. Long surveys lose people, and the people who stay answer worse as they go. Free text is the expensive part; a few well-placed open questions beat a dozen, and most of the dozen wanted to be scales anyway.
   
   
4. Take it yourself, then have someone break it. At speed, on a phone, the way a respondent will. You will find the question that makes no sense out of order and the scale whose midpoint means two things. Then hand it to someone uninvolved and ask them to answer it badly on purpose, which takes about ten minutes and is the cheapest QA you will ever run.
   
   
5. Let the achievable sample choose the method. Conjoint and the demanding pricing methods want samples most B2B research cannot reach, and running them anyway produces numbers with decimal points and no meaning, which are worse than no numbers because people believe them.
   
   
6. Incentives are part of the instrument. A prize draw selects for people who care about the topic or like a gamble; a guaranteed payment selects for people pricing their time. Neither is wrong, but the choice shapes who answers, so make it before writing questions rather than in week three.
   
   
7. Record what every number is measured against, when you produce it. The filter, the base, the date range. A 38% completion rate means one thing against people who started and another against people invited, and six weeks later, in a meeting, you will not remember which you calculated. Beside the figure, not in a footnote; footnotes do not survive being copied into decks.
   
   
8. Claim only what the sample supports. With small numbers, report direction, not magnitude, and say so out loud. My personal version: include only what is easily defensible and directionally true given the sample size. The charts come out duller. I have not yet regretted it.
   
   
9. Read the shape before the result. Everything clustered at one end of a scale means the questions were agreeable, not that people agree. Two clusters where you expected one may be two populations wearing the same label. Either way, you are learning about the instrument first.

How to tell whether it worked

Q: What separates a good research outcome from a bad one?

A: Good research changes a decision. If everything confirmed the brief, the instrument probably led the witness. And when two sources disagree, resist reconciling them, because the disagreement is usually the most valuable finding in the program.

The clearest sign the work mattered is that somebody decided something differently than they would have without it. A great deal being learned is not the same thing, and a readout going well is definitely not the same thing, though I do enjoy it when that happens.

The clearest sign it did not is that everything confirmed what the brief assumed. I have produced that result, and it is worth going back through the instrument to work out how, because research that ratifies the plan it was commissioned to test was usually built to do so, and nobody builds it that way on purpose.

The most valuable single result is usually a disagreement: two segments answering differently, or an internal survey and customer interviews contradicting each other about what customers care about. The instinct is to reconcile the gap before the readout, into one number everyone can live with, and it is worth resisting, because the gap is often the one finding in the whole program that could not have been bought anywhere else.

Don't skip the classics

None of this is new. I certainly didn't invent any of it. Most of it predates the current tools by decades, and the versions I use came from people who taught me by watching me get things wrong in front of them.

What changed is the price of skipping it. A badly built survey used to produce a bad slide, and a bad slide had a natural lifespan. Now it produces a bad slide, a summary of the slide, a recommendation resting on the summary, and a model happy to restate all three on request in that confident, faintly encouraging register. The jazz mix works because there are decades of Django Reinhardt underneath it, and whether your next readout works depends on the same question, asked before anything gets generated: what, if anything, is underneath.