I have spent most of my working life trying to get the right people on the phone.
At Nokia (opens in a new tab), at IBM (opens in a new tab) and at MIT (opens in a new tab), user and customer research was part of my job: understanding how people actually used a product, what they needed and what they would pay for.
Then as a consultant at McKinsey (opens in a new tab), as a founder, and later as an investor, I worked with more than a hundred businesses. Different industries, different countries, different stages. The one thing they all had in common was a question nobody could answer from a spreadsheet: what do the customers actually think?
And every single time, the answer was hiding behind the same bottleneck. Finding the people to ask.
The bottleneck nobody puts in the project plan
If you have never had to source interviews, here is how it usually goes. You write a screener. You ask around your network. You post in groups, beg colleagues for introductions, offer gift cards, and chase no-shows. Two weeks later you have spoken to six people, three of whom were not quite the customer you needed, and one of whom mostly wanted to talk about something else.
Meanwhile, the project has not stood still. A team is waiting on the findings. A product decision is penciled in. An investment committee wants an answer by Thursday. So the research gets squeezed, the sample gets smaller, and the decision gets made on the strength of whoever happened to pick up the phone.
I watched this happen in boardrooms, in startups and in due diligence rooms. The process was not just slow. It was the bottleneck. Everything else in a modern business can move in days. Customer understanding still moved in weeks.
The real cost is bigger than the invoice
The direct cost is easy to see: recruitment fees, incentives, research agencies, travel, transcription. A traditional focus group, depending on who you ask, costs thousands of dollars per session before anyone has analyzed a word.
The indirect cost is the one that kept me up at night.
- Decisions made without customers. When research is expensive, teams skip it. They build on assumptions and find out months later.
- Only testing the favorite idea. When you can afford one study, you test the idea you already believe in, not the three alternatives that might be better.
- Senior time burned on logistics. I have seen partners, founders and CEOs spend hours chasing interviewees instead of thinking about what they heard.
- Learning that arrives too late. By the time the insight lands, the budget is committed and the roadmap is locked.
Add those up across a hundred businesses and the number is outrageous. Not because anyone was careless, but because the tools for customer understanding had not changed in decades.
An expert network user, and an expert
I know the expert network world from both sides. I have been a client of GLG (opens in a new tab), paying to speak with people who know an industry. And I am a GLG expert myself, the person on the other end of those calls.
GLG is the best known of a whole industry of expert networks, alongside firms like AlphaSights (opens in a new tab), Guidepoint (opens in a new tab), Third Bridge (opens in a new tab) and Coleman Research (opens in a new tab). Expert networks do something valuable: they connect you with people whose experience you could never reach on your own. But the mechanics of it feel like the 1980s. A request goes in. Someone searches a database and makes calls. Profiles come back. You schedule around time zones. You pay by the hour. You get one person's view, at one moment, shaped by how much they want to be asked back.
The expert side is no happier. As an expert, I spend a surprising amount of time on manual, repetitive back and forth: answering the same screening questions again and again, filling in compliance forms, chasing scheduling emails, often only to learn I did not qualify for the project. Most experts I talk to share that frustration, and many feel underpaid and underappreciated by the networks that sell their time.
Sitting on both ends of that call, I kept thinking the same thing: there has to be a way to get the pattern across many people, fast, without waiting for calendars to line up.
Why synthetic users became possible
For a long time, the honest answer was "there isn't." Then the research changed. Studies like Out of One, Many showed that language models, given rich backstories, could reproduce the answer patterns of real groups of people. Later work found that agents built from real interviews could predict people's own answers almost as well as those people predicted themselves two weeks later. You can read the whole trail of evidence in Cast Diaries.
That research did not make human research obsolete. It made a new first step possible: a fast, cheap, honest first lap of customer learning, run before you spend weeks and budget on the second lap with real people. That is the job we built Cast for at Enhance Ventures (opens in a new tab), and it is the tool we now use every day to make our own decisions.
The problem with one perfect persona
Our first experiments taught us something uncomfortable. If you describe your ideal customer precisely and then ask an AI to play her twelve times, you do not get twelve interviews. You get one interview, twelve times. The answers are polished, consistent and almost useless, because real markets are not made of one person.
Real customers who match a profile are similar, not identical. One is a little older. One has a tighter budget. One was burned by a vendor last year. One loves new technology and one tolerates it. The insight lives in that spread: in where people agree despite their differences, and in where they split.
So we needed a way to cast the people around a profile, close to it but never exactly it. We called it Fuzziness Technology™, or simply Fuzziness.
What Fuzziness is
Fuzziness is how far each AI user is allowed to drift from the customer profile you describe. Once Cast understands a profile, the Casting Agent generates people in its neighborhood: real-feeling individuals who share the essentials of the profile but differ, deliberately and measurably, along the dimensions that matter.
Fuzziness Technology™ is proprietary to Cast, and it is what keeps Cast working close to the truth. With Fuzziness, you make better decisions based on the collective wishes and wants of an audience, not the opinion of one imagined customer.
The result is that your findings come from a small population rather than from one voice repeated. When eight out of twelve varied people raise the same objection, that objection is far more likely to be real than if one idealized persona said it twelve times.
We did not get this right on the first try. Fuzziness came out of research and countless trials: too little and every answer sounds the same, too much and you are no longer interviewing your customer. Finding the useful middle, and making it adjustable, took longer than anything else in Cast.
The default dimensions
Every audience in Cast starts with eleven dimensions of Fuzziness. Each can be set to Low, Medium or High, and the whole audience starts at Medium, which we found gives the most useful balance of focus and variety.
- Age: from within three years of the profile, up to any working age.
- Gender: from as described, to a mix weighted to the market, to an even mix.
- Nationality and background: from the most common for the role to the full local and expat mix. This matters enormously in markets like the Gulf.
- Location: from named cities only, to nearby cities, to across the country.
- Industry: from the exact industry, to adjacent ones, to any that fits the problem.
- Years in the industry: from typical tenure to newcomers and veterans.
- Seniority: from the exact title level to anyone from analyst to owner.
- Company size: from as described to startup through enterprise.
- Marital status and family: from most common for the age to the full range.
- Income level: from the market median for the role to the full range.
- Attitude to new technology: from pragmatists, to skeptics and fans, to the hostile and the evangelists.
The optional dimensions
Some studies need more. You can add optional dimensions when they matter to your decision:
- Buying authority: from no say at all to sole signer. Essential for B2B.
- Languages spoken: from the main local language to any language in the market.
- Education: from typical for the role to school leavers through postgraduates.
- Years at current employer: from new joiners to lifers.
- Tools and software used: from the standard stack to paper or custom systems.
- Risk appetite: from very cautious to very bold.
- Children: from none to a large family.
- Commute and car ownership: from typical for the city to the full range.
You can also add your own. If "uses a competitor today" or "has been through a failed implementation" is what really separates your customers, make it a dimension.
What stays fixed
Fuzziness never touches the essentials. If your profile says fleet operations managers in Riyadh, every AI user is someone who could credibly hold that job, in that market. And whatever the settings, every audience includes skeptics, people with a realistic reason to say no, because a study where everyone loves your idea is telling you about the study, not the market.
How to choose a setting
- Low when you know exactly who your customer is and want depth on one tight segment.
- Medium for most first studies. This is the default for a reason.
- High when you are not sure who your customer is yet, and want to discover segments you did not know existed.
- Mixed when one dimension is the question: keep everything at Medium and push only income, or only attitude to technology, to High.
My favorite way to use it is to run the same questions twice, once at Low and once at High. Where the answers hold, you have something robust. Where they change, you have just found the boundary of your market.
What I would tell my younger self
For years I accepted that understanding customers had to be slow, expensive and limited to whoever I could get on the phone. It does not. You should still talk to real customers, always. But you should arrive at those conversations having already explored the space, tested your alternatives and found the objections worth probing.
At Enhance Ventures (opens in a new tab) we now run Cast on every decision. More and more often, a Cast study on its own is enough to fortify a thesis: the patterns are clear, the objections are specific, and the evidence holds up under challenge.
I want to be equally clear about what Cast is not. It is not here to replace real customer contact. What it does is make that contact far better. Cast shows us our blind spots and tells us where to focus the limited time we have with real, and expensive, people. We walk into those conversations knowing which questions matter, which segments to prioritize and which assumptions are still shaky.
That is what Cast is for, and Fuzziness is the reason it works.
What is Fuzziness in synthetic user research?
Fuzziness is Cast's proprietary technology for keeping synthetic research close to the truth. It sets how far each AI user is allowed to drift from the customer profile you describe. It turns one profile into a small, realistic population, so findings reflect patterns across varied people rather than one persona repeated.
What is the default Fuzziness in Cast?
Every audience starts at Medium across eleven default dimensions, from age and location to income and attitude to new technology. You can set each dimension to Low, Medium or High.
Can I add my own Fuzziness dimensions?
Yes. Alongside eight optional dimensions such as buying authority, languages and risk appetite, you can add any dimension that matters to your decision.
Does high Fuzziness mean less accurate results?
Not less accurate, but broader. High Fuzziness is for discovery; Low Fuzziness is for depth on a known segment. Comparing the two is one of the best ways to find the edges of your market.
Read next: skeptics by design, and why your cast never all agrees.
