ExpertsFMKnowing What Your Customers Want, All the Time: Listen Labs' Alfred Wahlforss
Training Data · June 2, 2026
5 Selected Lessons:
case · 2 reasoning_pattern · 1 framework · 1 heuristic · 1
Building searchable expert profiles from accumulated interview data
Source
Training Data — Knowing What Your Customers Want, All the Time: Listen Labs' Alfred Wahlforss (June 2, 2026) · at 11:32 ▸
When this applies
when accumulating unstructured conversational or behavioral data over time and wanting to later retrieve the right person/asset for a new need
The takeaway
Capture and persist incidental signal from every interaction into a reusable profile, rather than treating each interaction as disposable.
Why it works
Incidental disclosures across many separate interactions can be tagged and stored so that a person's full profile compounds over time, independent of why they were originally contacted.
One level up
Data compounds in value when captured once and reused across many future contexts, turning past interactions into future targeting capability.
Where it stops holding
This stops working when privacy/regulatory constraints prevent persisting and re-using personal data across unrelated engagements, or when the population is too small/non-recurring for profile depth to accumulate.
Detail
procedure:
- Run interviews on varied topics across a large audience pool
- Capture incidental self-disclosures (e.g. 'I'm a sneakerhead') even when off-topic
- Persist these as structured profile tags on the individual
- When a new, relevant need arises (e.g. Nike product launch), search profiles to surface the right matched person instead of cold outreach
benchmarks: {"goal_scale":"a billion people in audience"}
rationale: Replaces manual email-list spam and high churn-inducing screening (incidence rate) with a pre-qualified, continuously enriched pool.
In their words
So someone might say in a totally unrelated interview, I'm a total sneakerhead. And you can keep that in the database on that person. And then when Nike or what have you is launching their new product line, you can offer that person up.
Asynchronous, low-pressure AI interviews elicit more honesty and participation than human interviews
Source
Training Data — Knowing What Your Customers Want, All the Time: Listen Labs' Alfred Wahlforss (June 2, 2026) · at 9:26 ▸ · at 11:32 ▸
When this applies
when trying to get honest input from people who feel judged, busy, or self-conscious
The takeaway
Reduce the social cost of participation (judgment, scheduling rigidity) to increase both honesty and willingness to engage, even to the point that people will accept lower compensation.
Why it works
Removing a synchronous human judge and allowing flexible timing lowers social pressure, which increases disclosure and lowers the price needed to get participation.
One level up
People optimize for comfort as much as compensation; lowering friction/anxiety is itself a form of payment.
Where it stops holding
This stops working when the topic requires human rapport/trust that an AI can't replicate, or when respondents distrust AI agents more than humans in certain cultural/demographic contexts.
Detail
approach: Design research interactions to be asynchronous, low-pressure, and non-judgmental to maximize honesty and participation.
signals:
- willingness to accept lower pay for AI interviews vs human interviews
- people 'opening up' more to a non-judgmental AI
- ability to interview sensitive populations like kids
rationale: Lower social friction increases both the quantity and the honesty of responses.
In their words
We've also found that people are more honest talking to an AI. We've had people really open up. It's a very therapeutic experience because it's a non-judgmental entity that's really interested in you.
Vertical AI moat comes from proprietary evals that keep getting harder
Source
Training Data — Knowing What Your Customers Want, All the Time: Listen Labs' Alfred Wahlforss (June 2, 2026) · at 28:37 ▸
When this applies
when building any product layered on commodity general-purpose technology and seeking durable differentiation
The takeaway
Create a proprietary, ever-escalating quality benchmark specific to your domain; each time you hit near-ceiling performance, add a harder bar, so the moat keeps renewing.
Why it works
Building and continuously tightening a domain-specific evaluation benchmark forces the product to climb a ladder of increasingly specific quality bars that general models don't optimize for, and each time you clear a bar you raise a harder one, compounding the gap.
One level up
Durable advantage in a commoditizing technology stack comes from an internal ratchet of self-imposed, domain-specific quality standards, not from the base technology itself.
Where it stops holding
This stops working when the underlying general models improve fast enough to leapfrog your narrow eval anyway, or when your eval doesn't actually track what customers value.
Detail
rule: Keep pushing: build a proprietary eval, climb it, then create a harder eval once the old one is mastered.
threshold: moved an eval from 20% to 85% pass rate, then introduced a new harder eval that reset to ~20%
rationale: This repeated escalation is what vertical AI companies can use as sustainable competitive advantage over general-purpose model providers.
In their words
And that's your advantage as a vertical AI company. Keep pushing forward, better data, harder problems, better data, repeat.
Interviews outperform behavioral/credit-card data for training a simulation persona
Source
Training Data — Knowing What Your Customers Want, All the Time: Listen Labs' Alfred Wahlforss (June 2, 2026) · at 23:13 ▸ · at 25:41 ▸
When this applies
when choosing which data source best captures a person's underlying reasoning rather than just their past actions
The takeaway
When modeling a person's future decisions, prioritize data that captures reasoning and context over data that only captures past outcomes.
Why it works
Behavioral logs (e.g. purchases) show what happened but not why; open-ended interviews let the model capture reasoning, tangents, and context that behavioral data structurally can't contain.
One level up
To predict behavior, you need the generative process (how someone thinks), not just its past outputs (what someone did).
Where it stops holding
This stops working when the interview itself is poorly designed (leading questions, wrong methodology), in which case even rich conversational data degrades in value — as Alfred notes elsewhere, interview design quality matters a lot.
Detail
situation: Testing which data source best predicts an individual's future answers for Listen's simulation feature.
decision: Compared behavioral/purchase data against structured interview data as model inputs.
outcome: Interview data performed best because it captures reasoning and allows tangents; behavioral data alone was less predictive.
lesson: Rich qualitative data that captures 'why' beats quantitative behavioral logs for predicting future individual preference, provided the interview is well-designed.
In their words
But what performs much better is we tried credit card spend, kind of behavioral data, purchasing behavior. But what we found was the best data set is interviews because it's more kind of allows you to go off tangents.
Niche-trained simulation outperformed general ChatGPT on a real decision (talk title)
Source
Training Data — Knowing What Your Customers Want, All the Time: Listen Labs' Alfred Wahlforss (June 2, 2026) · at 20:05 ▸ · at 23:13 ▸
When this applies
when a general-purpose AI model is asked to predict preferences of a specific niche audience rather than the 'average person'
The takeaway
A model fine-tuned/grounded on real data from your specific target niche will out-predict a general model prompted to 'act like' that niche, because grounding beats persona role-play.
Why it works
General models are trained on the average person's distribution, so their predictions regress toward generic preference; a model conditioned on real interviews from the specific niche captures that niche's distinct preference distribution.
One level up
Simulated specificity from a prompt is not the same as learned specificity from real population data.
Where it stops holding
This stops working when the niche sample size backing the simulation is too small or stale to represent the true population, or when the decision is novel enough that no amount of niche grounding applies (true out-of-distribution questions).
Detail
situation: Needed to pick the most effective of 100 candidate talk titles, and later to pick between two actual past talks (one successful, one not).
decision: Ran both through Listen's customer-panel-trained simulation and separately through vanilla ChatGPT.
outcome: Listen's simulation picked the correct/better title in both cases; ChatGPT picked the wrong one in the head-to-head test; the simulation's top pick scored twice as well as the next.
lesson: Niche-grounded simulation beat general-purpose LLM prediction for a real low-stakes decision, in Alfred's own early, informal test.
In their words
and actually had one. I had another talk I did that was not so successful. And I inputted a competitor or another talk that was more successful. And I showed both of them to ChatGPT and both of them to our simulation. And in ChatGPT, it picked the wrong one. And in our simulation, it picked the right