Yesterday, I posted a short scene every market researcher knows by heart.

A senior insights leader at a major food group described the moment she walks into a meeting, listens to a new concept, and knows — on instinct, built over twenty years — that it will not fly.

The promise is off. The price won’t fly. The benefit isn’t credible. But she has nothing on the table. No KPI. No verbatim. No data. Just her experience. So she pushes back politely, flags a risk, asks for a test — and the project keeps moving for months before a full quant study finally confirms what she already knew on day one.

Her words stuck with me:

“This is precisely where a FlashInsight-type read — on the KPIs everyone is already used to looking at, both qual AND quant — would be incredibly useful.”

A few hours later, Romani Patel 🎯 — Director of Data Science at Microsoft Marketing Research + Insights — reshared a post I could not have scripted better if I had tried.

And here is where it stops being a story about one researcher and one food group.

The message Romani reshared had been published the day before by Takeshi Numoto — Executive Vice President & Chief Marketing Officer at Microsoft. When a CMO of that stature puts his name, in public and on the record, on 90% faster messaging cycles and ~$10M in value, it is no longer a rumor from the lab — it is a position taken at the top of one of the most consequential marketing organizations on the planet.

Kate Weymer — Director of Data Science at Microsoft Marketing Research + Insights and Romani’s manager — echoed it the same day, and Romani closed the loop with a public thank-you to the team that built it.

What Microsoft just said out loud

The post described an internal product called AMA — the AI Messaging Assistant — built on Microsoft Foundry, grounded in customer research and real-world signals. Three sentences from it deserve to be quoted exactly, because they are the entire argument:

“A lot of organizations already have access to customer insight. The harder problem is turning that insight into everyday marketing decisions.”

“Instead of relying primarily on instinct or long feedback cycles, teams can pressure-test ideas in minutes across different audiences before anything goes to market.”

“We’re already seeing measurable impact, including 90% faster messaging cycles and ~$10M in value.”

💡 Read them again. Slowly.

This is not a vendor pitch. It is one of the most sophisticated marketing organizations on the planet telling the rest of us, on the record, that the bottleneck is no longer access to customer voice — it is the time it takes to convert that voice into a decision.

And that the right architecture cuts that time by an order of magnitude. Ninety percent. Ten million dollars. In a single internal pilot.

The full Microsoft customer story is here → Microsoft Foundry case study.

Subscribe now

Why this is not a coincidence

On one end of the spectrum, a senior insights director at a European food company, telling me — over coffee — that she needs a way to put her intuition on the table the moment it fires, in the language of KPIs her organization already trusts.

On the other end, Microsoft, telling the world that it has built exactly that, internally, at scale, and is already booking the value.

Same day. Same signal. From two ends of the maturity curve.

That is not a coincidence. That is a phase transition.

Yes — technology companies move faster. Microsoft has the stack, the data, and the engineering bench to ship Foundry-grounded systems in months. Most CPG, retail, financial services and pharma insights teams do not, and that asymmetry is real.

But the direction is no longer in debate.

The job of the insights professional is being rewritten in real time — from gatekeeper of long studies to architect of always-on, governed customer intelligence.

What changes, concretely

A few patterns are now visible across the conversations I am having every week:

• From quarterly studies to minutes-long pressure tests. Not to replace fieldwork — to pre-qualify what deserves fieldwork.

• From instinct vs. data to instinct plus data. Synthetic and AI-assisted reads do not silence the experienced researcher — they finally give her professional intuition the receipts it has always deserved.

• From insight as a deliverable to insight as an operating system. AMA inside Microsoft is not a report. It is plumbing.

• From “do we have a signal?” to “do we have a governed signal?” Calibration, alignment with established benchmarks — the next-generation insights team is judged on the rigor of its synthetic reads as much as on the rigor of its primary research.

This is the conversation we are also opening with hAppy Insight — the cercle of CMI leaders we are launching in Paris on June 18. Different geography, different industries, same direction of travel.

So — where are you?

The reason I wanted to write this as a newsletter and not as another short post is simple: I genuinely want to hear back.

The signal is loud at the top of the maturity curve. It is starting to ripple. But it is not ripping at the same speed everywhere — and that is the most interesting part of the story right now.

So, three honest questions, and I read every reply:

• 🏭 In your industry — CPG, retail, financial services, pharma, automotive, tech, media, public sector — where are you on the curve? Pilot? Production? Still skeptical? Still waiting?

• 🌍 In your geography — North America, UK, France, DACH, Iberia, Nordics, MEA, APAC, LATAM — what are you seeing accelerate, and what is stuck?

• 🛠️ In your own team — what is the first AMA-equivalent use case you would build if you had two weeks and a clear mandate?

Drop a comment, DM me, or forward this to the colleague you know is wrestling with exactly this question. I will be sharing the patterns back in the next edition, anonymized.

Intuition is still the most underrated asset in market research. Synthetic and AI-assisted reads, done seriously, finally give it the receipts it deserves. Microsoft just put a number on it.

Now it is our turn. 🚀

— “Dr Laurent” , FlashInsight · Co-initiator, hAppy Insight

Subscribe now

Sources & related reads

• Takeshi Numoto’s original post (May 11) — EVP & CMO, Microsoft

• Kate Weymer’s echo (May 12) — Director of Data Science, Microsoft Research + Insights

• Romani Patel’s reshare with team acknowledgment (May 12)

• Microsoft Customer Story — AI Messaging Assistant on Microsoft Foundry → Microsoft

• My post on intuition vs. proof (May 12) → LinkedIn

#MarketResearch #SyntheticPanels #FlashInsight #hAppyInsight #MicrosoftFoundry #CustomerVoice

Written by Dr Laurent Florès · L’Atelier IA · April 25, 2026

Companion note to: What FlashInsight Can and Cannot Measure and Enriching Synthetic Personas

This note began as a methodological observation about synthetic market research panels. It became something larger: a reflection on the difference between knowing about an experience and having lived through it — and what that difference means for how we build AI tools, conduct research, and educate the next generation.

The Precipitating Observation

During a four-wave study using FlashInsight synthetic panels to measure Moroccan perceptions of France, a troubling pattern emerged. The synthetic personas produced highly coherent, internally consistent responses about the impact of France’s 2021–2023 visa restriction policy. 97% reported that it had “strongly degraded” their image of France. The number was crisp, confident, and narratively plausible.

But something was off.

The personas knew the visa crisis happened. They could articulate its effects in fluent, contextually appropriate language. What they could not do — what no LLM-generated persona can do — is tell us how much it weighed against everything else a real Moroccan person carries: the memory of a relative who got excellent care in a French hospital, the admiration for a French professor who changed the direction of a life, the complex pride of speaking French better than most French people. The internal hierarchy of grievances, the emotional calibration of competing experiences — that is not knowledge. It is biography.

This is the distinction that opened the larger question addressed in this note: what is the difference between being something and having lived the thing you are supposed to be?

1. The Epistemological Gap: Representation vs. Experience

Philosophers have circled this problem for centuries. Gilbert Ryle’s distinction between knowing that and knowing how (1949) is perhaps the most cited formulation: propositional knowledge (knowing that Paris is the capital of France) is categorically different from procedural knowledge (knowing how to navigate a Parisian street market without being seen as a tourist). But there is a third category that Ryle touched on less explicitly and that matters enormously here: knowing what it is like — what the philosopher Frank Jackson (1982) called phenomenal knowledge in his famous Mary’s Room thought experiment.

Jackson’s Mary is a scientist who has spent her entire life in a black-and-white room. She knows everything there is to know about the physics and neuroscience of color perception. She can describe wavelengths, cone cells, neural pathways with perfect accuracy. Then she leaves the room and sees a red apple for the first time. Does she learn something new?

Jackson argued yes. And his argument applies directly to LLMs.

A Large Language Model trained on the entirety of human textual production knows about grief, humiliation, the feeling of having a visa refused for your mother’s funeral. It has processed thousands of first-person accounts, journalistic descriptions, sociological analyses, literary representations of these experiences. Its representations are rich, nuanced, and statistically sophisticated. But it has not felt any of them. It has never been in the waiting room. It has never held the refusal letter.

This matters for market research in a very specific way: the gap between knowing about an experience and having lived it is precisely the gap between structural attitudes and conjunctural salience. Structural attitudes — the value a person places on family, the deep cultural representation of France as a country of order and rigor — can be approximated from textual data. Conjunctural salience — how much this specific event weighs right now, in this person’s emotional economy — cannot.

2. What This Means for LLMs as Knowledge Instruments

Large Language Models are, at their core, extraordinarily sophisticated compression machines for human textual experience. They have ingested more text than any human could read in thousands of lifetimes, and they can retrieve, recombine, and generate from that corpus with remarkable fluency.

But fluency is not understanding. And representation is not experience.

This is not a new observation — it echoes Searle’s Chinese Room argument (1980), which holds that a system that manipulates symbols according to rules produces outputs that are syntactically correct without necessarily being semantically grounded. The LLM generates a response about visa grief that sounds like visa grief. Whether it is visa grief — whether there is something it is like to be that response — is a question the system cannot answer about itself, and we cannot answer about it from the outside.

For practical purposes in research and education, three consequences follow:

Consequence 1: LLMs are excellent at the documentable and poor at the felt.

Anything that leaves a textual trace is potentially accessible to an LLM. Historical facts, scientific findings, published opinions, documented cultural patterns — all of this is within reach, and often handled with impressive accuracy and nuance. What is not accessible is the undocumented interior: the private hierarchy of what matters most, the emotional weight of a specific memory, the way a single experience reshapes a person’s entire interpretive framework. These things exist in the world in abundance. They are simply not in the training data.

Consequence 2: The confidence of LLM output is not calibrated to epistemic depth.

A LLM generates responses about visa humiliation with the same fluency and apparent confidence as responses about photosynthesis. The stylistic markers of certainty do not track the underlying epistemic status. This is perhaps the most dangerous property of these systems for research applications: they do not signal their own ignorance about the phenomenal dimension. The 97% figure on visa impact looks like a finding. It was generated with the same computational process as the 87% optimism figure that validated perfectly against Ipsos. But one is grounded in documentable, slowly-changing sociological reality; the other is grounded in media salience in the training corpus.

Consequence 3: The documentable expands over time; the felt does not transfer.

As more human experience gets documented — through social media, personal essays, oral history projects, qualitative research archives — the LLM’s approximation of the phenomenal improves at the margins. But there is an asymptote. The first-person account of a specific experience, however richly described, is always a representation of that experience, not the experience itself. The gap may narrow; it cannot close.

Subscribe now

3. The Market Research Parallel

Market research has grappled with a version of this problem for decades, under different names.

The distinction between stated and revealed preferences — between what people say they will do and what they actually do — is fundamentally a gap between the documentable and the experiential. Focus group participants describe their purchasing decisions with confident narratives that often bear little relationship to the split-second implicit processes that actually drive them (Zaltman, 2003; Wilson, 2002). They are not lying. They are offering the best representation of their experience that conscious reflection can produce. But that representation is not the experience.

The methodology underlying FlashInsight’s correction of LLM positivity bias, addresses a specific version of this problem: it corrects for the gap between what a persona declares and what its semantic embedding implies. It is a form of revealed preference extraction from language. It works well for structural attitudes precisely because those are the dimensions where the textual record is rich and calibrated.

But this methodology cannot correct for what is not in the textual record at all. The hierarchy of personal grievances, the emotional weight of a specific biographical moment — these are not recoverable from semantic embeddings because they were never fully represented in language in the first place.

This is why the most rigorous path for synthetic research is hybrid: use LLM-generated panels to map structural terrain and generate hypotheses, then validate the conjunctural and phenomenal dimensions with real respondents. Not because real respondents are infallible — they are not — but because they carry something the synthetic panel cannot simulate: the irreducible authority of having been there.

4. The Learning Parallel: Copy-Paste and the Illusion of Knowledge

The same epistemological gap that limits LLM-generated personas operates in the interaction between students and AI tools — and it is producing a learning crisis that most educational institutions have not yet fully named.

When a student uses an LLM to write an essay, they receive a textual artifact that is often stylistically competent, factually informed, and structurally organized. What they do not receive is the process that produces understanding: the struggle with a difficult idea, the moment of confusion that precedes clarity, the cognitive work of building an argument from fragmentary evidence, the experience of being wrong and having to revise.

Dewey (1938) argued that education is not the transmission of information but the reconstruction of experience — that genuine learning requires active engagement with problems that matter to the learner. Piaget’s constructivism (1952) made the same point from a developmental perspective: knowledge is not received, it is constructed through interaction with the environment, including the friction of error and correction.

What LLM-assisted copy-paste produces is the representation of having learned without the experience of learning. The student holds a document that demonstrates knowledge. They do not hold knowledge. The distinction is invisible in the short term — the assignment is submitted, the grade is received — and devastating in the long term, when the student faces a novel problem that requires genuine understanding rather than textual pattern matching.

This is precisely the Mary’s Room problem transposed to education. The student who copies an LLM essay about the phenomenology of grief knows that grief involves specific cognitive and emotional processes. They do not know what it is like to reason through that problem, to confront the limitations of their own framework, to arrive at an insight through their own cognitive labor. They have the output without the process. The output is documentable. The process is experiential.

The troubling asymmetry: the documentable output can be evaluated and graded. The experiential process cannot be assessed from a document alone. Traditional academic assessment was partly designed around the assumption that the process left traces in the output — that you could read a student’s thinking in their writing. LLMs break this assumption. The output can now be generated without the process. Assessment has not caught up.

5. Why Experience Cannot Be Shortcutted — The Neuroscience

This is not a romantic or conservative argument. The claim that experience cannot be replaced by representation has a neurobiological basis.

Learning through direct experience activates multiple memory systems simultaneously: episodic memory (this happened to me, here, now), procedural memory (I learned how to do this by doing it), and semantic memory (this is what it means). The integration of these systems is what produces robust, transferable knowledge — what cognitive scientists call deep learning as opposed to surface encoding (Craik & Lockhart, 1972).

Reading about an experience primarily activates semantic memory. It can prime episodic simulation — the brain’s ability to mentally project itself into described scenarios — but this simulation is metabolically cheaper and neurologically thinner than actual experience. The emotional tagging that makes memories retrievable and actionable under stress (the mechanism studied extensively in the context of trauma and expertise, e.g., Damasio’s somatic marker hypothesis, 1994) is weaker or absent.

This means that knowledge acquired through description alone tends to remain inert — available for recognition and reproduction in familiar contexts, but poorly transferable to novel situations. The student who absorbed an LLM essay can reproduce its arguments in a similar essay. They struggle to apply those arguments to a genuinely different problem. The expert who has built knowledge through years of practice navigates novel situations by pattern recognition grounded in experiential memory. They have built what Kahneman (2011) calls System 1 competence through the repeated activation of System 2 effort.

LLMs produce outputs that look like System 1 competence — fast, fluent, confident — from a system that has no System 2 history at all. There was no struggle. There was no error. There was no revision. There was pattern completion at extraordinary scale.

6. Implications for FlashInsight’s Research Design

Returning to where we started, these reflections have concrete implications for how FlashInsight should position itself and how its outputs should be interpreted.

What FlashInsight does well, and why:

FlashInsight generates synthetic personas that approximate structural attitudes — the slowly-changing, value-rooted orientations that are extensively documented in the textual record. These are the dimensions where the LLM’s representation of human experience is most dense and most calibrated. The convergence between FlashInsight’s finding that 87% of Moroccan personas are optimistic about their country’s future and Ipsos’s real-world finding of 86% is not a coincidence. It reflects a genuine alignment between LLM-represented and empirically measured sociological reality on a stable, well-documented dimension.

Where the limits are, and why they are real:

The 97% visa impact figure is a different kind of output. It reflects not calibrated sociological measurement but the media salience of a politically charged event in the LLM’s training data. The feeling of having been humiliated by an administrative policy — the specific weight that carries in a specific person’s biography — is not in the training data. It is in the lives of real Moroccan people, accessible only through real fieldwork.

The research design implication:

FlashInsight is most valuable as a structural terrain mapper and hypothesis generator, used upstream of fieldwork rather than as a replacement for it. The structural map it produces — which value profiles are most receptive to which messages, which psychographic fault lines exist in the target population, which communication frames resonate with which identity orientations — is genuine intelligence that reduces the cost and improves the quality of subsequent real research.

The conjunctural measurement — what weighs on people right now, and how much — requires the irreplaceable instrument of real human testimony.

7. A Broader Provocation

The most consequential long-term risk of LLM proliferation in knowledge work is not the risk of misinformation, though that is real. It is the risk of a gradual atrophying of the cognitive and experiential processes through which genuine understanding is built.

If students consistently shortcut the struggle of learning, they do not merely fail to acquire specific knowledge. They fail to develop the cognitive architecture through which novel problems can be approached. The ability to sit with ambiguity, to work through confusion, to construct an argument from partial evidence, to revise one’s thinking in light of new information — these are not skills that can be acquired by watching an LLM perform them. They are acquired by doing them, repeatedly, in contexts where the stakes are real and the consequences of getting it wrong are felt.

The same applies to research. A market researcher who consistently delegates the interpretive work to AI tools does not merely risk misreading a specific dataset. They risk losing the practiced intuition — built through years of sitting with data, being surprised, being wrong, revising — that makes expert judgment possible.

This is the deepest reason why being something and having lived through what that thing is are not equivalent. The LLM that simulates a Moroccan persona has the representation of the experience. The researcher who has spent years in fieldwork, listening to real respondents in their real contexts, has something the LLM cannot compress and cannot generate: the calibrated judgment that comes from having been there, been wrong, and learned from it.

That judgment is not documented anywhere. It cannot be retrieved from a corpus. It cannot be injected into a prompt.

It has to be lived.

References

• Craik, F. I. M., & Lockhart, R. S. (1972). Levels of processing: A framework for memory research. Journal of Verbal Learning and Verbal Behavior, 11(6), 671–684.

• Damasio, A. R. (1994). Descartes’ Error: Emotion, Reason, and the Human Brain. Putnam.

• Dewey, J. (1938). Experience and Education. Kappa Delta Pi.

• Jackson, F. (1982). Epiphenomenal qualia. The Philosophical Quarterly, 32(127), 127–136.

• Kahneman, D. (2011). Thinking, Fast and Slow. Farrar, Straus and Giroux.

• Piaget, J. (1952). The Origins of Intelligence in Children. International Universities Press.

• Ryle, G. (1949). The Concept of Mind. Hutchinson.

• Searle, J. R. (1980). Minds, brains, and programs. Behavioral and Brain Sciences, 3(3), 417–424.

• Wilson, T. D. (2002). Strangers to Ourselves: Discovering the Adaptive Unconscious. Belknap Press.

• Zaltman, G. (2003). How Customers Think: Essential Insights into the Mind of the Market. Harvard Business School Press.

• Pichardo, V., et al. (2026). Cross-provider SSR independence in synthetic panel research. arXiv:2602.13862.

L’Atelier IA — latelier-ia.com | laurent@latelier-ia.com

Method over Magic

First: What Is a Synthetic Persona?

If you’re new to this topic, let me save you the jargon.

A synthetic persona is an AI-generated virtual respondent — built from real data, behavioral models, and language systems — that can answer questions, react to concepts, simulate preferences, and engage in dialogue as if it were a real consumer.

Not a survey bot. Not a chatbot. A structured simulation of human response patterns.

The simplest analogy: a flight simulator. It’s not a real aircraft. But it trains real pilots — because the feedback loops are real enough to matter.

The debate in market research isn’t really about whether synthetic personas work. It’s about what they’re for — and whether we’re using the right mental model to evaluate them.

The mistake we keep making is evaluating synthetic personas as measurement tools. They’re not. They’re simulation systems. And that distinction changes everything.

The Timeline Nobody Talks About

Here’s the thing about the Qualtrics prediction: at the time, most professionals didn’t believe it.

I was there. I was sharing that number in workshops and presentations throughout 2024 and early 2025 — and the reaction was skepticism at best, polite dismissal at worst.

“Too fast.”

“Not rigorous enough.”

“It’ll never replace real respondents.”

Fast-forward 18 months:

• Qualtrics has validated a fine-tuned synthetic model that reduces fielding costs by up to 50% and produces results nearly identical to human respondents

• Toluna has extended synthetic personas to 15 markets as of October 2025

• Ipsos France published a formal taxonomy distinguishing four categories of synthetic data: Data Boosting, Data Imputation, Persona Bots, and Digital Twins

• The global synthetic data market — valued at ~$267M in 2023 — is projected to reach $4.6 billion by 2032

• Stanford and Google DeepMind have demonstrated that digital twins built from a few hours of interviews reproduce real responses with 85% accuracy — comparable to how consistent humans are with themselves two weeks apart

• More than half of market researchers now use synthetic data to expand scope and accelerate insights, according to Forbes (Sept. 2025)

This is no longer a forecast. It’s a status report.

One year ago, nobody reacted when you shared the Qualtrics numbers. Today, those same numbers generate serious debate. That’s the sign that the gap has been crossed.

What the Debate Is Actually About

The pushback on synthetic personas is real — and some of it is legitimate.

Critics argue that synthetic respondents can’t produce genuine surprise. That they reflect training data biases. That the “human costume” — the names, ages, simulated hesitations — creates an illusion of listening when no real listening has happened.

These are valid concerns. And they all point to the same underlying issue: literacy.

A flight simulator is only dangerous if the pilot doesn’t know they’re in one.

Synthetic personas are only misleading if the practitioner treats them as evidence rather than as a structured surface for exploration.

The real shift — the one that’s still underway — is methodological, not technological.

It’s not about whether the tools work. It’s about when to use them, what questions to ask, and how to interpret the outputs responsibly.

This is also why the vocabulary matters so much. When we call synthetic respondents a “panel”, we import all the expectations of a panel — representativeness, longitudinal tracking, human unpredictability. We set the tools up to fail at something they were never designed to do.

Renaming is not semantics. It’s a strategic act. The language we use shapes the mental model — and the mental model shapes the evaluation criteria.

What’s Actually Being Built Right Now

I don’t just write about this. I build with it.

Over the past year, I’ve been developing and iterating on research frameworks using synthetic personas — not in theory, but in practice. Running concept tests. Testing messages. Exploring consumer reactions across segments I couldn’t reach cost-effectively with traditional methods.

Recently, working with Codex (rather than Claude Code — token limitations are real), I’ve been testing the portability of these frameworks. I started with French consumer data. After a conversation with a friend, I extended the same framework to Morocco. Imperfect? Yes. But it worked — and that alone tells you something important about where this is going.

The value isn’t in replacing a field study. It’s in expanding the space of possible questions before you invest in answers.

• Stress-test a concept in 2 hours, not 2 weeks

• Identify which hypotheses are worth validating with real respondents

• Test portability across markets before committing budget

• Run sensitivity analyses on message framing before campaign deployment

This is what I mean by walk the walk. Not talking about synthetic personas as a trend — but using them as an active part of a research workflow. Learning from what breaks. Iterating.

The gap between those building and those debating is widening fast. And the companies filling that gap are the ones that will define the next generation of insights work.

Three Things Worth Watching in 2026

1. The validation race

The critical question for 2026 isn’t adoption — it’s validation. Which use cases produce reliable, actionable outputs? Under what conditions does synthetic data diverge from human responses? The players investing in rigorous validation frameworks will separate from those selling hype.

2. The hybrid model as default

Synthetic exploration upstream → real data validation downstream. This isn’t a compromise; it’s an optimized workflow. Faster, cheaper, and — when done well — more rigorous than running expensive studies on unvetted hypotheses.

3. The nomenclature shift

The field will need to develop a stable vocabulary. Synthetic personas vs synthetic panels. Simulation vs measurement. Exploration vs validation. The companies and researchers who get the language right will also get the use cases right.

Where Do You Stand?

If you’re still in the “watching from the sidelines” phase, this is a reasonable moment to reconsider.

Not because synthetic personas are a silver bullet — they’re not. But because the organizations learning to use them now are building a compounding advantage in speed, cost, and hypothesis quality that will be very hard to close in two years.

The prediction was made in October 2024. We’re 18 months in. The next 18 will move faster.

Simulation upstream. Validation downstream. That’s the workflow. Everything else is commentary.

Want to Go Deeper?

If you’re curious about how synthetic personas work in practice — what they can and can’t do, how to build a responsible framework, and where the real value is — I’m building resources around exactly this at L’Atelier IA.

You can also explore the fundamentals directly: I’ve put together a base of resources on synthetic personas — concepts, case studies, and methodological frameworks — that you can dive into and discuss.

👉 Explore the Synthetic Persona Base — and if you have questions, let’s talk.

Sources

• Qualtrics (Oct. 2024). 2025 Market Research Trends Report — 71% of researchers predict synthetic majority within 3 years

• Qualtrics (Oct. 2025). The Synthetic Research Breakthrough — Fine-tuned models outperform general AI

• EY & Evidenza (2024). Synthetic Data Validation Study — 95% correlation with traditional methods

• Stanford HAI & Google DeepMind (2024). Digital Twins: 85% accuracy study

• Forbes Tech Council (Sept. 2025). The Promising Rise of Synthetic Personas in Market Research

• Ipsos France — Mathilde Guinaudeau (Jan. 2026). Tribune on synthetic data taxonomy. MR News

• Research Live (Jan. 2026). Preview of 2026: Synthetic Data — industry perspectives

• Christophersilvestri.com (July 2025). State of Synthetic Research in 2025

• Columbia Business School — Olivier Toubia’s research on digital twins and synthetic respondents

Laurent Florès is a professor at Paris-Panthéon-Assas University, former ESOMAR President, and founder of L’Atelier IA. He writes at the intersection of AI, market research, and decision intelligence.

“The old equilibrium—where take-home work could reliably measure understanding—is dead. Gone. Kaput.”

This brutal assessment from Panos Ipeirotis, professor at NYU Stern, captures the predicament facing every educator in higher education today. With students armed with ChatGPT, Claude, and Gemini, the question is no longer whether they’re using AI—it’s how we evaluate what they actually know.

The trigger: suspiciously perfect submissions

It started with a red flag. In their “AI/ML Product Management” course, co-taught with Konstantinos Rizakos, the pre-case submissions were too good. Not “strong student” good. More like “McKinsey memo after three rounds of editing” good.

So they started cold-calling students in class.

The results were illuminating—and damning. Many students who had submitted thoughtful, well-structured work couldn’t explain their own choices after two follow-up questions. Some couldn’t participate at all.

The principle is simple: if you can’t defend your work in real-time, then the written artifact isn’t measuring what you think it’s measuring.

The oral exam problem: a logistical nightmare

The obvious solution? Return to oral exams. After all, it’s the format most resistant to AI cheating. You can’t have ChatGPT whisper in your ear when an examiner is looking you in the eyes.

Except oral exams have a brutal problem: they don’t scale.

Let’s do the math:

• 36 students × 25 minutes × 2 evaluators = 30 hours of faculty time

• For 100 students? 83 hours

• For 200 students? Impossible

And that’s before accounting for accommodation requests—travel, other finals, family events. All legitimate. All turning exam periods into scheduling nightmares.

This is the bottleneck the NYU experiment attempted to break.

The architecture: specialized AI agents

The platform they used was ElevenLabs Conversational AI, which bundles speech-to-text, text-to-speech, turn-taking, and interruption handling. But the real innovation was in the pedagogical design.

Rather than a single “chatty” agent trying to do everything, they built a workflow of specialized agents:

Authentication agent

Verifies student identity before proceeding. Simple but essential.

Project discussion agent

Receives the context of each student’s group project and asks personalized questions. This is where the “LLM did my homework” strategy dies—you have to explain specific choices, justify trade-offs, respond to probing follow-ups.

Case discussion agent

Selects a case study from the course and tests conceptual understanding in real-time.

This modular approach prevents the system from drifting into unbounded conversation and makes debugging possible.

The numbers: shocking economics

The results deserve attention:

Metric Value Students examined 36 over 9 days Average duration 25 minutes (range: 9-64) Messages per conversation 65 average Total cost $15 ($0.42/student) AI-human grade agreement 89% within ±1 point

Compare this to the human alternative: 30 hours at TA rates (~$25/hour) = $750. That’s a 50:1 cost ratio.

But the real value isn’t the cost savings.

What broke (and what it teaches us)

The first version had problems—instructive ones for anyone designing AI-assisted learning systems.

The voice was intimidating

They had cloned a colleague’s voice. Students complained it was too intense. One wrote:

“I had prepared thoroughly and felt confident, but the intensity of the interviewer’s voice unexpectedly heightened my anxiety and affected my performance.”

Another: “The agent was shouting at me.”

Lesson: Voice selection isn’t a technical detail. It’s a pedagogical decision that directly impacts student performance.

The agent stacked questions

The system would ask: “Explain your metric choice, and also tell me what baselines you tried, and why you didn’t use X, and what you’d do next.”

That’s not one question. That’s four questions in a trench coat. Cognitive overload guaranteed.

Lesson: One question at a time. If you want depth, chain across turns. This applies to human examiners too.

The agent didn’t let students think

Students would pause to reflect. The agent interpreted silence as confusion and jumped in with follow-ups.

Lesson: Thinking time is part of assessment. Respect it; don’t fill it.

“Random” selection wasn’t random

They asked the agent to “randomly select” a case study. From December 12-18, the agent picked Zillow 88% of the time. After removing Zillow, it latched onto Predictive Policing for 16 out of 21 exams.

Lesson: LLMs aren’t random. They have implicit preferences and ordering biases. Asking an LLM to “pick randomly” is like asking a human to “think of a number between 1 and 10”—you’ll get a lot of 7s.

The grading council: when three AIs deliberate

This might be the most fascinating aspect of the experiment.

The grading used a “council of LLMs“ approach, borrowed from Andrej Karpathy (former AI director at Tesla, OpenAI co-founder). Three models—Claude, Gemini, ChatGPT—evaluated each transcript independently. Then they saw each other’s assessments and revised. Finally, Claude (as “chair”) synthesized the final grade with evidence.

Round 1: disagreement

When models graded independently, agreement was poor: 0% identical grades, only 23% within ±2 points.

The kicker? Gemini was lenient (average: 17/20). Claude was rigorous (average: 13.4/20). A 3.6-point gap—the difference between a B+ and a B-.

Round 2: convergence

After seeing each other’s assessments and evidence, agreement improved dramatically between Round 1 & 2:

• Perfect agreement: 0% → 21%

• Within ±1 point: 0% → 62%

• Within ±2 points: 23% → 85%

• Mean max difference: 3.93 pts → 1.41 pts

Gemini lowered its grades by 2 points on average after seeing Claude’s more rigorous assessments. It couldn’t justify giving 17s when Claude was pointing to specific gaps.

Crucial insight: The disagreement wasn’t random. “Problem Framing” and “Metrics” had 100% agreement within ±1 point. “Experimentation”? Only 57%. When students give clear answers, evaluators agree. When answers are vague, partial credit becomes subjective—for AIs and humans alike.

The uncomfortable mirror: what AI revealed about teaching

Here’s where it gets painful.

Analyzing results by topic, one bar stuck out: Experimentation. Mean score: 1.94 out of 4. Compare that to Problem Framing at 3.39.

The breakdown was brutal:

• 3 students (8%) scored 0—couldn’t discuss it at all

• 7 students (19%) scored 1—superficial understanding

• 15 students (42%) scored 2—basic understanding

• 0 students scored 4—no one demonstrated mastery

Performance by topic — the Experimentation bar stands out :

The natural reflex would be to blame the students. But the professors made a more painful realization: they had rushed through A/B testing methodology in class, pressed by the calendar.

The external evaluator, objective and relentless, reflected back the image of their own teaching. Like a mirror, it doesn’t lie. Student weaknesses revealed course weaknesses.

The grading output became a mirror reflecting our own weaknesses as instructors. Ooof.

This might be the most unexpected benefit of the system: it diagnoses teaching as much as it evaluates learners.

A counterintuitive finding: duration ≠ quality

Exam duration had zero correlation with score (r = -0.03).

The shortest exam—9 minutes—earned the highest score (19/20). The longest—64 minutes—scored 12/20.

Taking longer doesn’t mean knowing more. Often it signals struggling to articulate. Confidence is efficient.

What students said: accept the concept, reject the execution

The professors surveyed students before releasing grades:

• Only 13% preferred the AI oral format. 57% wanted traditional written exams. 83% found it more stressful.

• But: 70% agreed it tested their actual understanding—the highest-rated item.

• They almost universally liked the flexibility of taking the exam at their own place and time.

Student survey results :

The verdict: they accept the assessment principle but reject the current execution.

The questions this experiment doesn’t answer

Technical enthusiasm shouldn’t eclipse fundamental questions.

Are we solving the right problem?

If students can’t defend their work, is it a cheating problem or a never-taught articulation skills problem? AI oral exams detect the symptom, not the cause.

What are we actually measuring?

Does the ability to respond under pressure to a voice agent measure deep understanding or stress tolerance and verbal fluency? Those skills matter. But are they the primary pedagogical objective?

Anxiety as confounding variable

83% of students found the format more stressful. Anxiety can mask real comprehension. A brilliant but socially anxious student might fail not from lack of knowledge but from paralysis facing the examiner’s voice.

The access equity question

Students take the exam “at home.” But not everyone has access to a quiet environment, stable internet, or quality audio equipment. Does inequality in material conditions become inequality in evaluation?

The arms race continues

If AI oral exams become standard, how long before students use an LLM whispering answers through an earpiece? Webcam monitoring won’t prevent creative workarounds.

The real stakes: rethinking what we assess

The NYU experiment is a symptom of a deeper transformation. Take-home exams are dead. Returning to pen-and-paper in the classroom feels like regression. But is automating oral assessment the solution or a technical bandage on a pedagogical wound?

Maybe the real question isn’t “How do we prevent students from cheating?” but “What do we actually want them to learn?“

If the goal is critical thinking, decision-making, and real-time reasoning, then yes, oral exams make sense. But if the goal is deep understanding and the ability to produce quality work, perhaps we need to rethink entirely what an “exam” is.

Ipeirotis concludes: “Fight fire with fire.”

We could also stop playing firefighter and redesign the building.

Sources

• Ipeirotis, P. (2025). Fighting Fire with Fire: Scalable Personalized Oral Exams with an ElevenLabs Voice AI Agent. A Computer Scientist in a Business School.

• Jabarian, B. (2024). AI vs. Human Interviewers: Evidence from a Randomized Field Experiment. SSRN Working Paper.

• Karpathy, A. (2024). Council of LLMs. GitHub.

• ElevenLabs. (2025). Conversational AI Platform Documentation.

Laurent Florès is a professor at Paris-Panthéon-Assas University, former ESOMAR President, and founder of L’Atelier IA. He writes about AI, research, and the human factor.

2025 wasn’t the year of progressive adoption. It was the year of fracture.

The GRIT Business & Innovation Report, published in November 2025, put it bluntly: “we’re not going through a temporary crisis, we’re witnessing a deep and lasting restructuring of the industry.”

The market isn’t slowing down. It’s splitting. And this bifurcation is already redrawing our sector’s map for 2026 and beyond.

The Great Schism: Tech-Led vs Service-Led

Imagine two competing companies. Same sector, same starting expertise, same positioning five years ago. In 2025, their trajectories have completely diverged.

→ Company A, “tech-led” model: Invested massively in AI. Automated its analysis processes. Processes millions of data points in seconds. Revenue growing year after year.

→ Company B, “service-led” model: Bet on traditional human expertise, personalized consulting, client relationships. Revenue stagnating. Sometimes even declining.

This isn’t a theoretical scenario. It’s the reality documented by GRIT based on thousands of responses from sector professionals (insight industry).

This fracture isn’t a cyclical accident or temporary slowdown. It’s a permanent reconfiguration of the competitive landscape.

Consulting Has Crossed the Inflection Point

While market research debates whether to adopt AI, consulting firms have crossed the critical threshold.

The numbers are eloquent:

→ McKinsey: 12,000 AI agents deployed to assist 40,000 employees. Over 75% use Lilli, the internal AI platform, which handles 500,000 requests per month.

→ Accenture: 12,000 layoffs in three months. Official reason: “people unable to retrain for AI.” Training of 550,000 employees on generative AI.

→ BCG: $2.7 billion in AI-related revenue in 2024, or 20% of total. Target: 40% by 2026.

Kate Smaje, McKinsey’s global technology head, summarizes the situation bluntly: “Do we need armies of business analysts creating PowerPoints? No, the technology could do that.”

The consulting profession isn’t dead. It’s reinventing itself around what AI doesn’t do: weighing options in a company’s cultural context, anticipating human resistance, adapting a recommendation to power dynamics.

The machine digests data. The consultant orchestrates the decision.

From the Fourth to the Fifth Generation

We take questionnaires designed for humans and submit them to language models. Then we’re surprised the results are sometimes disappointing. We digitize the old instead of inventing the new.

Generative AI calls for a conceptual leap on a radical scale.

We need a fifth generation that’s a true methodological rupture:

→ Continuous hypothesis generation and filtering

→ Modeling possible behaviors

→ Rapid comparative scenario testing

→ Stress-tests of risks and tension zones

→ Hybrid combination of observed and simulated data

Research adopting AI-native approaches tells an entirely different story: Columbia Business School achieves 85% accuracy with calibrated personas, Park et al. reduce certain biases by 40%, innovative protocols achieve correlations close to 0.9 with human data.

Already Viable Use Cases in 2025

Rather than asking “Can they replace humans?”, it’s more useful to ask: “For what types of decisions are the signals produced already good enough?”

→ Ideation and initial exploration

→ Pre-tests and concept selection

→ Scenario stress-tests

→ Analytical complement to human data

The key question isn’t “is this true to the hundredth?” but “is this reliable enough for the type of decision I want to make now?”

The Elephant in the Room: The Panel Quality Crisis

The GRIT Report 2025 confirms it: concerns about data quality have increased by 40% in one year.

Fraudulent respondents, bots posing as humans, disengaged participants, compromised samples: this is the reality of our “human” data.

Criticizing synthetic respondents in the name of data quality while ignoring the progressive collapse of traditional panels amounts to comparing a theoretical ideal to a severely degraded reality.

The question is no longer choosing between a perfect human and an imperfect synthetic, but knowing how to combine imperfect sources to produce better, more robust, and faster decisions.

2026-2027: Perspectives and Predictions

→ Partial adoption convergence, but not revenue convergence

→ Industrialization of synthetic data

→ Market polarization: massive automation at the bottom, ultra-qualified consulting at the top

→ Standardization via the ESOMAR Task Force on AI

→ Generalization of the 5th generation

→ Consolidation through acquisitions

→ Emergence of “Insight Engineers”

The window of opportunity to position yourself is 18 to 24 months.

What to Do Now? Three Viable Strategies

1. Become Tech-Led: Invest massively in AI capabilities. Automate processes. Develop fifth-generation native methodologies.

2. Become a Quality Guarantor: Position yourself as a methodological rigor expert facing the data confidence crisis. Audit AI solutions. Validate synthetic data.

3. Become an Orchestrator: Don’t do everything yourself, but master the ecosystem. Know how to choose the right partners and arbitrate according to context.

The only losing strategy? Waiting and seeing.

Conclusion

2025 will go down in our profession’s history as the year everything started to tip.

The real danger isn’t choosing the “wrong” strategy. It’s not choosing at all. It’s staying in the middle ground that’s disappearing.

The real question isn’t “will AI replace humans?” It’s: “what does the human bring that AI cannot?” And then: “do I have the skills to bring this value in 2026?”

The GRIT report’s message is unambiguous: inaction is synonymous with obsolescence.

The time to choose is now.

Sources:

• GRIT Business & Innovation Report 2025 — Greenbook

• McKinsey: Meet Lilli, Our Generative AI Tool

• BCG Annual Report 2024

• Accenture Annual Report 2024

• Park, J. S., et al. (2024). Generative Agent Simulations of 1,000 People. arXiv.

• Columbia Business School (2025). Digital Twins Research

This article was originally published in French on L’Atelier IA.

Laurent Florès is a professor at Paris-Panthéon-Assas University, former ESOMAR President, and founder of L’Atelier IA. He writes about AI, research, and the human factor.

I’m still curious.

In fact, I’m more curious now than I’ve ever been.

If you found this newsletter through my piece on how AI is transforming consulting, you might be wondering: who is this guy?

Fair question. Before we go further, I owe you some context.

Because the ideas don’t exist in a vacuum. They come from somewhere.

Three Revolutions, One Thread

I’ve had the unlikely luck of witnessing — and working through — three technological revolutions.

The PC Revolution (1980s)

I studied computer science when “artificial intelligence” meant something completely different. Expert systems. Rule-based logic. The promise was already there, but the technology wasn’t.

What I learned: every revolution starts with inflated expectations, then disappointment, then quiet, transformative adoption.

The Internet Revolution (1990s-2000s)

In 1998, I moved to the United States to help build Ipsos’s digital measurement business. The internet was being born, and I was in the room.

I was so convinced this was transformative that I left the corporate world to start my own company — CRM Metrix — which I grew across the US, Europe, and Asia before it was acquired.

During those same years, while running a transatlantic business, I completed a PhD. My dissertation introduced the Brand Delphi method, a new approach to research that embraced the unique characteristics of online environments. It was later published in Marketing Science.

What I learned: new technology doesn’t just accelerate old methods. It demands new thinking.

The AI Revolution (Now)

And now we’re here. Generative AI. Large language models. A rupture more profound than anything I’ve seen.

The difference this time? I’m 59. I have more perspective. But I also feel the same fire I felt in 1998 — maybe even stronger.

Because this one changes everything.

Where I Come From

I grew up in a modest family where work was the central value. Not work as drudgery — work as dignity. The idea that nothing valuable comes free. That respect is earned. That if you want something, you build it.

This stayed with me.

When I struggled to find my path after university, I worked through it. Marketing studies. Applied statistics. Eventually, market research — a field I entered in 1991 and never really left.

I worked at Research International in Brussels. Millward Brown in England. Ipsos in France. I specialized in advertising effectiveness, running qualitative and quantitative studies across countries and cultures.

Then the US. Then entrepreneurship. Then academia.

Since 2008, I’ve been a researcher and professor at Paris-Panthéon-Assas University. I’ve also taught at Fordham University in New York, at EADA in Barcelona — in English, in French, in Spanish. To executives, to students, to anyone willing to learn.

Teaching became my way of giving back.

The Sea, the Wind, and Staying Curious

There was a moment, after years of running at full speed, when I needed to step back.

I became a sailing instructor at Les Glénans a famous sailing school in France. I crossed the Atlantic. Twice.

I discovered kitesurfing — and it became a metaphor for how I want to live: reading the wind, adjusting constantly, finding balance between power and control.

Then kitefoiling. Then paragliding.

These aren’t hobbies. They’re how I stay curious outside of work. They remind me that learning never stops. That the body and mind are connected. That humility comes from confronting forces bigger than yourself.

Why “One Curious Thing”?

I’m constantly amazed by how many people have stopped being curious.

They let themselves be carried. They consume without questioning. They wait for answers instead of seeking them.

We live in an era of unprecedented access to knowledge, to tools, to possibilities. And yet so many people don’t see it. Don’t use it. Don’t wonder.

This newsletter is my small rebellion against that.

Each issue, I’ll bring you one curious thing — an idea, a study, a question, a pattern I’ve noticed at the intersection of AI, research, work, and human understanding.

Not hype. Not doom. Just curiosity, applied with pragmatism.

What You Can Expect

Here’s what I’ll be exploring in the coming months:

AI and the transformation of work

How artificial intelligence is reshaping knowledge professions — consulting, research, marketing, education. Not replacing humans, but redistributing value. The “inflection point” concept I’ve been developing.

The future of insights and research

My home field for 35 years. Synthetic respondents. Digital twins. The fifth generation of market research. What’s real, what’s hype, and what actually works.

The human factor

What remains irreducibly human in an age of automation. Judgment. Context. Empathy. Change management. The skills that appreciate as AI commoditizes execution.

Lessons from three revolutions

Patterns I’ve seen repeat. Mistakes I’ve made. What history teaches us about navigating technological rupture.

Curiosity itself

How to stay curious. How to learn continuously. How to avoid the trap of thinking you already know enough.

My Commitments to You

I won’t pretend to have all the answers. I’m exploring too. This newsletter is my thinking out loud, not my lecturing.

I won’t chase hype. I’ve seen too many cycles to get swept up in the excitement. Pragmatism first.

I’ll cite my sources. My academic training won’t let me do otherwise. When I share data, you’ll know where it comes from.

I’ll stay curious. That’s the whole point.

A Word About Value

I don’t believe in free.

Not because I’m against generosity — I’ve spent years doing pro bono tutoring in my neighborhood, serving on nonprofit boards, mentoring young researchers.

But because I believe value must be exchanged. Attention is valuable. Time is valuable. If you’re giving me yours, I owe you something worth reading.

That’s the deal.

Who This Is For

If you’re a knowledge worker wondering what AI means for your career — this is for you.

If you’re in market research, insights, consulting, or marketing — this is definitely for you.

If you’re curious about technology but skeptical of hype — welcome.

If you believe that human judgment still matters — let’s explore together.

If you’ve ever crossed an ocean, literal or metaphorical, and found yourself changed — you’ll understand where I’m coming from.

Let’s Begin

I’ve been on this journey for over 35 years. Three revolutions. Multiple countries. Countless students. Two Atlantic crossings.

And I’m still curious.

That’s the only credential that matters.

Welcome to One Curious Thing.

Laurent Florès is a professor at Paris-Panthéon-Assas University, former ESOMAR President, and founder of L’Atelier IA. He writes about AI, research, and the human factor.

System1, the leading platform for creative effectiveness measurement, awarded the spot a 5.0-star rating—the highest score ever recorded for a French advertisement since they began tracking.

But here’s what caught my attention: one week after launch, a qualitative study conducted on a panel of synthetic personas had already identified, with remarkable precision, the deep reasons behind this success.

Are we talking about “prediction”? Not quite. A week after launch, the buzz was already visible. Let’s talk instead about anticipation capacity—the ability to understand why something works, well before post-hoc analyses are published.

And that’s where it gets fascinating.

A Question Worth Asking

Before diving in, let’s pause on something important:

Is it easier to anticipate success or failure?

The Intermarché case is spectacular, but it raises a fundamental methodological question. When something works powerfully, the signals are often clear. The verbatims are enthusiastic, positive emotions dominate, and engagement is palpable.

But what about creative work that will fail? Are the signals equally readable? Can synthetic research detect flaws before they manifest in the market?

We’ll explore that soon with another case study—perhaps on a less successful creative, or on quantitative synthetic approaches. Stay tuned.

The Story of a Wolf Who Conquered the World

Let me briefly recap the film. An uncle gives his nephew, who’s afraid of wolves, a wolf puppet. The father then tells the story of this wolf, rejected by all the forest inhabitants because he is... a wolf. Following advice from a philosophical hedgehog, our protagonist decides to change his diet, learns to cook vegetables, and ends up being accepted by the community during a shared meal.

The whole thing is carried by sublime 3D animation, produced by the Montpellier-based studio Illogic Studios, and accompanied by Claude François’s song “Le Mal-Aimé” (The Unloved One).

Watch the spot on YouTube

The result? An emotional tidal wave that transcended borders, generations, and cultures.

Two Perspectives, One Striking Convergence

What makes this case particularly interesting for research professionals is the ability to compare two approaches:

1. The Post-Hoc Analysis (Perplexity)

A few weeks after launch, I asked Perplexity to synthesize the reasons for the spot’s success. The AI compiled available analyses and identified three key factors:

1️⃣ Format & duration are not limitations: accepting 2.5 minutes rather than 30 seconds creates a narrative space where emotion can develop organically.

2️⃣ Music is a viral multiplier: the association with a culturally significant song amplifies virality well beyond the initial video.

3️⃣ Creative humanity triumphs: in the age of generative AI, visible, conscious, patient creative craftsmanship constitutes an irresistible emotional differentiator.

2. Synthetic Qualitative Studies (one week after launch)

I had access to comprehensive results from a study conducted by Airpanel on a panel of synthetic personas representing French household shoppers aged 18 to 60. Fourteen complete interviews, with detailed verbatims.

Meanwhile, Toluna—with their “HarmonAIze Personas” (over one million synthetic personas) and their ACT Instant AI advertising testing solution—also offers this type of analysis, with results 20 times faster than traditional methods.

What’s striking? The same success drivers were identified by these different approaches—with a depth of analysis and nuances that the post-hoc analysis hadn’t captured.

💡 Transparency note: I’m reporting here the results I had access to, completely independently. The fact that multiple market players arrive at the same conclusions about this spot reinforces the face validity of the synthetic approach for advertising testing.

The Three Pillars of Success: What Synthetic Research Anticipated

1. The Universal Archetype of the “Unloved One”

What the synthetic study revealed:

“We feel a bit sorry for the wolf, that ‘unloved’ aspect. We’ve all experienced that a bit, haven’t we? Being left out.”

— Pierre L., 35 years old (synthetic persona)

The study demonstrated that the wolf functions as a powerful projection screen. Panelists, regardless of age or profile, could identify with this feeling of exclusion. The resolution through an act of generosity (cooking for others) offers a satisfying catharsis, typical of Christmas tales.

What the Perplexity analysis confirms post-hoc: the viewer becomes “complicit in this fable”—exactly what the synthetic verbatims had captured.

Verdict: ✅ Perfectly anticipated

2. Redefining “Eating Better” as Social Connection

What the synthetic study revealed:

“It says they’re not just concerned with selling products, but also with... with the connection between people. ‘Eating better’ is that—it’s sharing a good meal, taking care of your loved ones.”

— Marcel L., 45 years old (synthetic persona)

The study identified the major semantic shift operated by the film: “eating better” moves from an individual health injunction to an instrument of socialization. Food becomes a language, a bridge between individuals.

Verdict: ✅ Perfectly anticipated

3. Resonance with New Food Consciousnesses

What the synthetic study revealed:

“Bravo for daring to do something other than the traditional family meal with meat at the center of the table. [...] It’s a message of hope, in a way.”

— Lucie V., 22 years old (synthetic persona)

The study captured a dimension that many mainstream analyses underestimated: the strong resonance of the film among vegetarians and vegans. For these audiences, the wolf switching to vegetables is perceived as a strong signal of modernity from the brand.

Verdict: ✅ Anticipated with nuance, with the study capturing the particular intensity of these profiles’ reactions.

What Synthetic Research Captured Beyond

Beyond validating success factors, the synthetic qualitative study provided insights not available in post-hoc analyses:

The Complete Emotional Palette

The synthetic interviews revealed an emotional journey in multiple phases:

• Initial skepticism (classic family scene)

• Curiosity when switching to animation

• Empathy for the rejected wolf

• Amusement with the hedgehog character

• Admiration for technical quality

• Satisfaction and warmth at the resolution

The “Stop” Effect of the Animation Switch

“I didn’t look at my phone once, which is quite an achievement during a commercial!”

— Elodie F., 28 years old (synthetic persona)

The study identified that the shift from live-action to animation acts as an attention trigger—a valuable operational insight for creatives.

The Identified Weakness: Length

“It’s a bit long. I’ll admit that in the middle of the song, I checked the time.”

— Louis H., 50 years old (synthetic persona)

The study also captured the only recurring criticism: the nearly three-minute format represents a risk for multiple exposures. The recommendation to prepare shorter formats (60s, 30s) proved relevant.

What Was Less Predictable

Let’s be honest: certain elements of the global success were difficult to anticipate:

1. The Scale of International Virality

The study focused on a French panel. It couldn’t anticipate that the film would resonate so strongly beyond francophone borders. The universality of the “unloved one” archetype and the absence of dialogue facilitated this internationalization.

2. The “Anti-AI” Effect as Differentiator

The release context (just after Coca-Cola’s AI-generated ads, criticized for their coldness) created a contrast effect that amplified positive reception. This timing element was difficult to model.

3. Secondary Musical Virality

The song “Le Mal-Aimé” experienced a renaissance on streaming platforms. This phenomenon of musical virality, which extends the campaign’s lifespan, wasn’t in the study’s scope.

The Real Advantage of Synthetic: The Cost/Speed Ratio

Let’s be clear: a traditional qualitative study could have been conducted in a week. An experienced “qual researcher” can recruit, conduct interviews, and deliver an analysis in 5-7 business days.

So the real question isn’t “is it doable traditionally?” but rather: at what cost?

💰 The cost ratio of synthetic vs. traditional: depending on configurations, we’re talking about a factor of 1:10 to 1:20. A traditional qualitative study of 14 in-depth interviews (as run here) with complete analysis may cost between €5,000 and €15,000 or more. An equivalent synthetic study? A few hundred euros.

This ratio fundamentally changes the equation:

• More iterations possible: test multiple versions, multiple angles

• Democratization of insight: accessible to SMEs, startups, teams with limited budgets

• Agility: possibility to launch exploratory research before validating traditionally

What This Teaches Us About Synthetic Studies

The Opportunities

⚡Speed of execution: Actionable insights in hours, not weeks.

🎯Accessible qualitative depth: Complete interviews with rich verbatims, at a fraction of the cost.

🔄 Iteration capacity: Test, adjust, re-test—synthetic enables agile exploration.

The Limitations

🌍 Cultural bias: A “French” synthetic panel doesn’t capture international reactions.

📊 Absence of behavioral data: The study measures stated intentions and emotions, not actual behaviors.

⏰ Context effects difficult to model: Timing, competition, external events remain uncontrolled variables.

Recommendations for Using Synthetic Qualitative Research Well

1. Use Synthetic for the “Why,” Not the “How Many”

Synthetic panels excel at exploring deep motivations and emotional resonances. They’re less suited to predicting precise quantitative metrics.

2. Diversify Persona Profiles

Richness comes from variety. The study benefited from profile diversity (ages, professions, food sensitivities) that enabled capturing important nuances.

3. Treat Results as Hypotheses

Synthetic verbatims are plausible and coherent, but they remain simulations. They should be treated as hypotheses to explore, not absolute truths.

4. Combine with Other Sources

The ideal is to use synthetic studies as a complement: social listening, quantitative studies, traditional panels for critical validations.

Conclusion: A Methodological Revolution in Motion

The Intermarché Wolf case perfectly illustrates the potential of qualitative studies on synthetic panels. By capturing key success factors one week after launch, the study demonstrates a remarkable anticipation capacity—at a cost and speed that changes everything.

Synthetic doesn’t replace traditional. It augments it. It enables rapid hypothesis exploration, creative pre-testing, and understanding emotional reactions—all with qualitative depth that was previously reserved for large budgets.

At a time when generative AI is reshaping creative professions, it’s ironic to note that it’s also AI that allows us to better understand why human creations move us.

What’s Next?

This case opens fascinating questions:

• Does synthetic detect failures as well as successes?

• How do synthetic personas behave with more polarizing creative work?

• What about quantitative synthetic?

I’ll have more case studies to share on these topics soon.

In the meantime, if you have questions, cases to discuss, or want to explore integrating synthetic into your research processes—don’t hesitate to reach out directly.

This article was written by Dr Laurent Florès based on Perplexity’s analysis of the Intermarché spot’s success and comprehensive results from the synthetic qualitative study conducted by Airpanel on December 13, 2025

Laurent Florès is a professor at Paris-Panthéon-Assas University, former ESOMAR President, and founder of L’Atelier IA. He writes about AI, research, and the human factor. More about him here.

McKinsey just went from 45,000 to 40,000 employees. In the same breath, they deployed 12,000 AI agents.

Accenture laid off 12,000 people in three months — those deemed “unable to retrain for AI.”

BCG now makes 20% of its revenue from AI. Target for 2026? 40%.

These aren’t projections. These are facts from the past twelve months.

And if you think this only concerns consultants, I have news for you.

I’ve been watching the intersection of Tech and professional work for over two decades. What’s happening right now in consulting is a preview of what’s coming for every knowledge worker.

But here’s what most people get wrong: this isn’t about replacement. It’s about redistribution.

The consulting profession isn’t dying. It’s being surgically reorganized around what AI can’t do — yet.

Let me show you what that looks like.

The Three Phases — A Model That Became Reality

A few months ago, I wrote about three phases in AI adoption: productivity, inflection point, substitution.

I thought we had time.

We didn’t.

Phase 1: Productivity (2022-2023)

ChatGPT arrives. Consulting firms quickly build internal tools. McKinsey launches Lilli, trained on decades of institutional knowledge.

The goal? Help consultants work faster. The human stays in control.

Gains were real but modest: 10-15% time saved on specific tasks.

Phase 2: The Inflection Point (2024-early 2025)

This is where everything changed.

The “inflection point” is the moment AI stops assisting and starts replacing entire functions.

Three signals tell you it’s happening:

→ AI performance matches or exceeds humans on specific tasks

→ Cost savings justify restructuring

→ Required skills change radically

All three signals fired in 2024.

Phase 3: Substitution (Now)

We’re here.

AI no longer complements. It takes over.

Some functions disappear. Others become more valuable. The question isn’t “will AI change my job?” — it’s “which parts of my job are already being automated while I’m not looking?”

What AI Actually Replaces

Let’s be specific.

Over 75% of McKinsey employees now use Lilli. The platform handles 500,000 requests per month.

What does it do?

→ Research synthesis

→ First-draft presentations

→ Data analysis

→ Competitive benchmarking

→ Initial recommendations

These tasks used to be a junior consultant’s entire job.

A diagnostic phase that took 2-3 weeks? Now 3-5 days.

And here’s the twist: quality went up. AI accesses a larger knowledge base than any human could, identifies patterns we’d miss, produces structured output faster.

The Skill Equalization Effect

A Stanford study of 5,000 customer support agents revealed something fascinating:

AI improves junior performance by 35%.

Senior performance? Only 2%.

Why?

Because AI compensates for what juniors lack: knowing which questions to ask, which frameworks to apply, which patterns to recognize.

Seniors already have these instincts. AI saves them time but doesn’t transform their approach.

This “equalization from below” has massive implications.

The classic consulting pyramid — many juniors supporting few seniors — is inverting. You need fewer juniors (AI does their work) but the same number of seniors (judgment can’t be automated).

Kate Smaje, McKinsey’s global technology head, put it bluntly:

“Do we need armies of business analysts creating PowerPoints? No, the technology could do that.”

What AI Can’t Do (Yet)

So what remains?

Contextual decision-making.

AI synthesizes and analyzes. Humans decide.

That distinction is everything.

Deciding isn’t choosing the optimal option on paper. It’s weighing options within a company’s cultural context. Anticipating human resistance. Adapting recommendations to power dynamics.

An AI can calculate that closing three factories improves margins.

It can’t evaluate whether a CEO can politically survive that decision. It can’t anticipate union reactions. It can’t craft the communication strategy that makes implementation possible.

“Political” intelligence.

Consulting isn’t just finding the right answer. It’s identifying the acceptable answer in a specific context, then convincing people to adopt it.

This requires empathy. Reading between lines. Detecting what’s not said in meetings.

Change management.

Here’s a sobering stat: 70% of transformations fail. Not because the strategy was wrong — because organizations resisted change.

AI can’t spend three months on-site helping a team adopt new practices.

Human presence, trust-building, individual convincing — these remain irreducibly human.

The Numbers Behind the Shift

Let’s look at how the big three are responding:

McKinsey

• 12,000 AI agents deployed

• Staff reduced from 45,000 to 40,000

• 70%+ trained on Lilli

• 500,000 AI requests monthly

Accenture

• 12,000 layoffs in Q1 2024

• 550,000 employees trained on generative AI

• AI specialists tripled from 40,000 to 77,000

• Revenue from AI-related services surging

BCG

• $2.7B in AI-related revenue (20% of total)

• Target: 40% by 2026

• BCG X division: 3,000+ AI specialists

• Partnerships with Anthropic, OpenAI, Google, Microsoft

The pattern is clear: massive training, strategic restructuring, repositioning around AI-human collaboration.

Who Should Be Paying Attention

This isn’t just about consultants.

Every profession built on collecting, processing, and presenting information is on the same trajectory:

Translation → Human translators now focus on cultural nuance and creative work. Technical translation is automated.

Accounting → Less time on entries and verification (automated), more on strategic advisory.

Legal research → AI scans thousands of precedents in seconds. Lawyers focus on argumentation and client relationships.

Market research → Data collection and processing automate. Analysts specialize in contextual interpretation and strategic recommendations.

Teaching → AI handles knowledge transmission and personalized paths. Teachers focus on motivation and critical thinking.

The through-line: execution automates, judgment appreciates.

What This Means For You

The window to adapt is closing.

A recent Stanford study confirms junior roles are most affected by AI, with demand declining 16% since late 2022.

Accenture’s CEO Julie Sweet was explicit:

“We’re parting ways, in a very short timeframe, with people we believe won’t be able to acquire the skills we need.”

So what skills matter now?

1. AI fluency — Not coding. Understanding what AI can and can’t do. Knowing how to prompt effectively, evaluate outputs critically, iterate intelligently.

2. Synthesis capacity — AI produces raw analysis. You need to contextualize it, translate it into actionable recommendations.

3. Emotional intelligence — Trust-building, conflict management, reading rooms, supporting humans through change. These become the core of the job.

4. Deep expertise — In a world where AI levels access to general knowledge, specialized expertise becomes your moat.

5. Learning velocity — The tech changes monthly. Cognitive agility isn’t optional.

The Real Question

Here’s what I’ve come to believe after watching this unfold:

The inflection point isn’t coming. It’s here.

The question isn’t “will AI transform my profession?”

It’s: “When will I decide to transform with it?”

Consulting in 2025 was a laboratory. It shows us what intellectual work looks like when AI takes over execution: less production, more synthesis. Less analysis, more judgment. Less solo work, more orchestration.

This evolution isn’t good or bad. It’s happening.

The question is how you — individually, intentionally — choose to respond.

The train is moving. Are you getting on board?

Sources:

• Stanford University (2023), “The Impact of AI on Productivity: Evidence from 5,000 Customer Support Agents”

• Stanford HAI (2025), “AI Hiring Slows; Junior Developer Jobs Change Dramatically”

• McKinsey (2024), “Meet Lilli, Our Generative AI Tool”

• McKinsey (2023), “The Irrational Side of Change Management”

• Boston Consulting Group (2024), Annual Report

• Accenture (2024), Annual Report

• The Conversation (2025), “Salariés, chefs d’entreprises : l’IA va tout changer”

Laurent Florès is a professor at Paris-Panthéon-Assas University, former ESOMAR President, and founder of L’Atelier IA. He writes about AI, research, and the human factor. More about him here.