Jobs to Be Done & VP Design

26. October 2026

What are synthetic users good for? My write-up of the Produktwerker episode

Tim Klein has invited me onto Die Produktwerker for the third time. This time it was about a question that is on a lot of product people’s minds right now: are AI and synthetic users any good for interviews, including Jobs to Be Done interviews? You can listen to the whole episode, Was taugen synthetische Nutzer? (“What are synthetic users good for?”), on the Die Produktwerker podcast (in German), or on Apple Podcasts. Since the episode is in German, the quotes below are my own translation.

AI helps. At almost every step.

Tim’s starting point was the effort involved: good interviews go deep, but the analysis takes a lot of time. So my first answer was not a warning:

“AI is a fantastic tool, or in my view a fantastic opportunity, to get help at all sorts of different points. So from how do I conduct an interview all the way to who do I actually need to talk to?”

That runs from preparation through to analysis. And the analysis is the part where most teams get stuck. In my experience, very few have a process for it. Those who do still need time.

Sounds plausible, but too smooth

What is a synthetic user? Put simply: an AI that answers on the basis of a candidate profile as if it were that person. If you interview a candidate like that yourself, you notice:

“At first you hear something that sounds totally plausible. […] Still, something smells a bit off. Something about it is just too smooth. […] the answer reads as if it had been read out of some marketing brochure. […] And that breeds mistrust.”

If you have little interviewing experience, you will find these answers plausible. If you have conducted lots of real interviews, you do a double take.

Forces yes, triggers no

For me, this is the most important point of the episode. Synthetic interviews are not worthless. They capture some things well, and one thing not at all:

“[…] with synthetic interviews we get at pains and gains, so when we talk about forces, what people like, what they don’t like, you can get at that relatively well. But where you can’t get at all is the events, the trigger events, the triggers for a change in behaviour patterns.”

A model knows what people like and dislike – but not the moment they actually switch. You need that moment if you want to understand why someone buys. And why they didn’t for so long.

The edges count, not the middle

When we select interviewees, we deliberately look for the extremes: “So you need the outliers; the middle is what you need least. You have to understand the outliers, at the top and at the bottom.”

There are patterns in triggers too, life events for instance: moving house, a divorce, a child. The Jobs to Be Done interview about the height-adjustable desk that we published is one such case: “A child comes into the family and suddenly he buys himself a height-adjustable desk. Why then? Even though he’d been thinking about it for years.”

Still, the pattern alone won’t get you any further: “So there are clear patterns, and they’re really important for providers to know, too, it’s just that without the specific context, that in itself is worth nothing.”

Without the context of the individual person, a trigger remains a textbook phrase. This is where synthetic interviews get no further:

“With synthetic interviews you get more or less generic patterns. They’re universally valid, because you can only draw on existing data. You can’t really discover anything new with them that the models and the machines don’t already know.”

You can’t predict behaviour well with them. What a model doesn’t already know, you will only find by talking to real people.

The synthetic user always answers

A synthetic user always has something to say: “[…] it always gives you an answer. But that in turn makes you suspicious”.

“Clients are suspicious too.” Synthetic results are valued and used far less than real interviews, no matter how good they are. The reason:

“Because this empathy with a human being, with a real person you can hear, who has genuinely lived through it themselves, is obviously still worth much more than when some computer has calculated it, and you think it’s only telling me what I want to hear anyway.”

Cheaper? Not necessarily.

Tim asked about the break-even point. My answer: “I don’t think it’s cheaper.”

AI’s biggest lever lies in analysing real interviews. Pattern analysis across many conversations is much faster and cheaper today than it used to be.

That has a downside too: “If I then just binge on it, it isn’t worth anything any more […]. That’s simply the downside of these automated analyses.”

So we put a choice to clients: “Do you want to internalise this? […] Or do you just want to see a result so that you can make the next decision better?” If you want the team to internalise it, the team does the pattern analysis itself. “But then they’ve really absorbed it, and nothing can replace that. You can’t replace this empathy by having an automated analysis shown to you that you skim, because it’s a lot of text, too. You can’t just boil that down to one or two charts in PowerPoint and then believe you know what makes your customers tick. I think that’s a hallucination on the human side.”

A result helps you with the next decision. You only know what makes your customers tick once you have worked your way through the patterns yourself.

What synthetic users are good for

At the end, Tim wanted the short version. My first case: “when you can’t find candidates, when you have an area where you think it’s incredibly hard to get to the people.” Even if you do reach candidates with a lot of effort, there remain “gaps at the edges”, and that is exactly where synthetic candidates can help.

The second case is a small budget: “Or if you have little budget, a small business, and just want to get started in the first place, then it’s still better than nothing”. With one condition: “knowing full well that you should definitely come back to it later and that this is only a start for now.”

Synthetic users are an honest start if you know you’ll have to follow up later. And for anyone who dismisses the topic by reflex: “so healthy scepticism is good, […] So then have a closer look at it. Try to understand it.”

My recommendation

Finally, Tim asked about the first step:

“[…] make sure you stay in touch with customers yourselves […]. So keep trying to do interviews yourselves and get help with the analysis. That’s the laborious part, the annoying part that hardly anyone can be bothered with, and AI is very good at that […]. And that moves you forward, because it means you’re producing your own raw data.”

Stay in touch with customers yourself and get help with the analysis. If there is nothing at all and the interviewing experience is missing too, it can be worth buying in synthetic interviews. And before that: “So first clarify the goal. Why do I even want to do this? What is supposed to be different afterwards, once I have this information?” If that isn’t clear, you’re better off leaving it.

Postscript: what people buy is a compromise

I’m adding one thought because it was missing from the episode. People never buy the best product, only ever the least bad compromise in a specific situation. They want more than they get, and they give up one thing to achieve another. What wins in the end is not decided on average but in that situation: often by a person they trust, or by a piece of evidence that puts a worry to rest. Exactly how this compromise is formed is what makes the difference. That is precisely what synthetic users can’t yet do. Only a conversation about a real purchase shows you that – and from it follows what you should promote and what you should leave out.

What you take away from this

Synthetic users show you what people like and dislike, but not what moves them to switch. You find the trigger and its context by talking to real people, ideally at the edges of your target group. Conduct your interviews yourself, get help with the analysis and clarify beforehand what should be different afterwards.

If you want support

Ideally, you conduct an interview yourself and then have someone who documents and prepares it properly for you. For me at least, that was like a dream come true. In the episode I mentioned that this is exactly what we now offer: “if you’ve conducted interviews and don’t want the stress of analysing them, you can just send them our way […].”

This way of analysing interviews is the result of at least six years of work. This AI-assisted analysis, which we use for ourselves and also offer to our clients as a service, really is the best you can get out of an interview – at least in my view – in terms of usefulness and usability.

Afterwards you can see what carried the purchase and what was given up for it: what was paid for, what was put up with and what remained only a wish, plus, as a hypothesis, what that suggests for your decision. Every statement is linked to the point in the conversation where it was made, and if there is a recording, every quote can be checked right back to the original audio. And you get feedback on how you ran the interview, because: “There’s no such thing as the perfect interview.” It is the conversation that is assessed, not you.

If you would rather internalise it, you can learn to do the analysis yourselves as a team: with a real interview, live, and without skipping the manual sorting, because empathy only comes from getting your hands dirty with the data.

One question to finish

When did a customer last surprise you by doing something different from what they had said in the conversation?

The answer is yours. If you have interviews about a purchase or a switch that nobody has analysed yet, there is the Buying-Reason Analysis of your own interviews. And if you want your team to be able to do it themselves, there is Mastering JTBD, the live online workshop for you and your team.

You can find the whole episode with Tim Klein on Die Produktwerker under Was taugen synthetische Nutzer? (in German) or on Apple Podcasts.

Related topics: AI · Practical Example