Sophia Toh

We're facing more and more questions about the use of synthetic data in market research. I'd recommend this blog post, which really resonated with me. So much of the conversation is "how closely can this replicate human data?" But there are so many other considerations researchers should be thinking about. https://lnkd.in/eSQfSTPj

31

3 Comments

Another great blog post from TRC Insights on segmentation and consumer insights. A nice example that utilizes proper survey questions to elicit differentiation, machine learning techniques for the analysis, and thoughtful synthesis of the results. https://lnkd.in/ere7YVfd

5

From Strategy to Activation: Getting Your Data AI-Ready Continuing the conversation from the Snowflake World Tour, let's build off the foundational strategy with data activation. Most companies want to activate AI, but their data just isn’t ready for it. Dirty, decentralized, and inoperable data stops innovation before it starts. If your AI strategy is set, it’s time to turn to your data strategy. Here’s how to get your data ready to fuel innovation and growth: Clean it - Audit what data exists and what’s missing. - Eliminate duplicates and outdated records. - Set up continuous hygiene processes so bad data doesn’t creep back in. Centralize it - Consolidate your data into a unified, governed environment. - Define clear ownership, permissions, and metadata standards. - Eliminate silos so teams can operate from a single source of truth. Activate it - Connect clean, centralized data to the AI workflows that drive value. - Use it to train models, personalize experiences, and surface insights faster. - Track outcomes, refine what works and retire what doesn’t. You can’t build an intelligent organization on an unintelligent data foundation. AI readiness isn’t about more data, it’s about better data. If AI is the engine, is your data clean enough to fuel it? #DataStrategy #AI #Snowflake #DigitalTransformation #DataGovernance #Analytics #EmergingTech

9

2 Comments

A great post from my colleague Erin Davison Medeiros and Vision Insights' head of Data Science. We're constantly hearing about data providers leveraging synthetic data or data imputation to fill holes and gaps in data sets or finding ways to bolster sample sizes without having to actually survey more people. But at what cost to the integrity of your data? A couple of quick takeaways from this blog that really stuck with me related to the use of these types of data in market research: 1️⃣ "The point is to listen to people, and that requires … well, listening to people." If your market research about humans isn't actually asking humans, is it really market research? 2️⃣ "LLMs are trained on past data, whereas the goal of a survey should be to listen to people now. Even if an LLM has historical data aligning with our question, it is outdated as soon as it has been trained" In sports, if your team goes on a massive playoff run or a new partner makes a splash with a big partnership, how will a data set reflect current and future success if it's being based off of past performance or historical data? 3️⃣ On the topic of hard-to-reach respondents, (minorities, executives, etc.)...even when subgroup bias has been addressed in some way, there exists a "butterfly effect, that small changes to algorithms, training data, or prompts can lead to substantial changes in the output of LLM models that magnify the effect of biases." If the data provider already struggles to provide enough sample for unique and hard-to-reach audiences, how accurate and reliable will look-alike models be?

14

I'm pleased to share a recent Paradigm piece on strategies for elevating data quality while broadening audience reach. These objectives are often seen as competing, but with the right approach, they can be mutually reinforcing. https://lnkd.in/eFi7CT_W

28

Thought-provoking research here. Great perspective from Rob Kaiser, Ph.D. on the growing conversation around synthetic audiences and the Columbia University Digital Twins research. As someone working in advanced analytics within the insights industry, I see real potential for synthetic respondents as part of the research toolkit. But this work highlights an important reminder: the most valuable insights often come from the unpredictability of real people. That element of surprise is still something models struggle to replicate. #MarketResearch #AdvancedAnalytics

2

I tried Anthropic’s new Claude Interviewer demo released on 12/4/2025. It’s a quick adaptive AI-led interview that follows up on your answers in real time. Not available for fielding your own research yet, but it’s a clear look at the future of data collection. What caught my attention is the gap between how useful people say AI is vs. how they feel about it. People love the time savings. They don’t love the stigma, the trust issues, or the uncertainty about what this means for their future. Scientists in particular questioned whether AI is actually a net time saver once you verify everything. This aligns with what we’re seeing across the industry. The value is there. The comfort level isn’t always. I also pulled the key findings into a quick infographic below, created in NotebookLM. My demo run wasn’t perfect (Claude froze when I asked it for example answers), but even with bugs, the direction is obvious. AI is getting close to acting like a qualitative researcher: asking follow-ups, probing, adjusting, and synthesizing themes at scale. Would you rather give feedback to an adaptive AI interviewer or a static programmed survey? Demo: https://lnkd.in/ezbSAdXE Study Results: https://lnkd.in/egB7qerT Dataset: https://lnkd.in/eayhQdgv #AI #Anthropic #UserResearch #MarketResearch #Insights

27

This breakdown of generative AI vs. predictive AI is a timely reminder: while both are powerful, they serve very different purposes. Predictive AI is about forecasting outcomes based on patterns. Generative AI creates new content, ideas, or assets based on inputs. One helps you see what’s likely to happen. The other helps you imagine what could happen. At Propelis, we’re exploring both through a strategic lens, asking how these technologies can support better decision-making, faster execution, and more meaningful brand experiences, without sacrificing trust or transparency. AI should never be applied blindly. It should be used with purpose, guided by strategy, and grounded in ethics. #AI #GenerativeAI #PredictiveAI

15

2 Comments

Most conversations about end-to-end AI and model training still focus on scale. But scale alone doesn’t get you to production AI. What actually matters across the full model training lifecycle is signal: high-quality labeled data, clear data provenance, human-in-the-loop evaluation, and retraining pipelines supported by strong MLOps, deployment, and monitoring. At Shutterstock, we offer solutions to AI teams across the entire training lifecycle from licensed training data, custom training datasets, and multimodal data curation to model training, fine-tuning, model evaluation, benchmarking, and continuous model improvement. We operate as an AI lifecycle partner, not a point solution, supporting enterprise AI teams with data sourcing and creation, refinement, human preference signals, and qualitative evaluation that models can reliably learn from. The outcome isn’t just better benchmarks. It’s production-ready and always improving models with stronger model alignment, reduced regressions, and greater confidence at deployment. Get in touch with our team to learn more: https://lnkd.in/guCmrqxm #ProductionAI #EnterpriseAI #ModelTraining #Shutterstock #FuelGreatWork

3

Is AI coming for UX research? Angel Tang, UX Researcher at Merkle, weighs in on how AI transforms the field—and what it can’t replicate. Watch this insight-packed Fast Currents clip:

7

1 Comment

In the consumer world, we want a single interface that can write a poem, plan a vacation, and debug code. In that open world, scale is the only strategy that works. But enterprise workflows don’t live in an open world. Most B2B problems like parsing invoices, routing tickets, classifying clauses operate in closed systems. They have well-defined inputs, explicit outputs, and hard failure modes. In my latest article, I argue that the future of enterprise AI isn't always about getting bigger, it's often about model specialization. The most effective architectures I’m seeing are cascading. They use SLMs for routine volume, and escalate to LLMs only when deep reasoning is required. Read the full breakdown below. https://lnkd.in/gkyJbFWb

57

6 Comments

91% of data leaders say culture is the biggest barrier to AI. Not tech. That number hit me hard. AI isn’t just a test run anymore. It’s quickly becoming part of daily life for managers. Work is changing fast, and leaders are expected to keep up even faster. Here’s the bottom line. AI is handling the busywork so managers can focus on what matters most, the human side of work that truly drives results. We have to keep in mind this is the end goal we all want! Here are the top five takeaways you should know. 1. Managers are looking for real relief. 55% believe AI will make scheduling easier next year, and half expect to spend less time on admin tasks. 2. People are already saving time. Leaders such as Danielle Spires at Asana have gained back four to five hours each week because AI now handles prep and synthesis. 3. The pressure around communication is changing. AI will track who has received and understood updates, so managers can spend more time coaching and solving problems instead of following up. 4. Managers still need to oversee AI. Leaders will check AI’s output, guide decisions, and make sure it supports the organization’s goals. 5. The amount of work isn’t shrinking, it’s just changing. AI will handle repetitive tasks, while managers focus on strategy, communication, and people leadership. The focus is shifting from tasks to true leadership. This is the change I’ve been hoping to see. AI is removing distractions and showing us what strong leadership really means. Managers who focus on clarity, trust, and human connection will do well. These were the type of projects my team and I were working on at Paramount, and we were having great success! Basically, those who stick to old habits will quickly fall behind. #AILeadership #AITransformationLeadership #FutureOfWork #ManagerInsights #WorkplaceTrends #LeadershipShift #AI2026 #TeamPerformance #HumanCenteredLeadership #DigitalWorkflows #PeopleAndCulture #LeadershipMindset #WorkforceInnovation #SmartManagement #AIInTheWorkplace #NextGenLeaders #ProductivityRevolution #BusinessIntelligence #ModernManagement #AITransformation #LeadershipExcellence https://lnkd.in/es_MBsHY

3

Running experiments often feels like gambling. Should you put more volume behind variant A, or give variant B another chance? Traditional A/B testing splits traffic and waits - but what if you could continuously adapt, maximizing gains as you learn? Enter Multi-Armed Bandits: an elegant blend of probability, statistics, and decision-making that turns experiments into dynamic optimization engines. Just like choosing the right slot machine at a casino, Multi-Armed Bandits help you decide which option deserves your next coin - except here, the coin is traffic, impressions, or user attention. Let’s explore how they work, why they beat static testing, and how we’ve applied them in Databricks. [...] Final Thoughts If anomaly detection is about spotting trouble fast, Multi-Armed Bandits are about finding upside faster. They turn experimentation from a fixed split into a living system that learns where your attention pays off. In practice, the math is modest, the operational footprint is small, and the payoff - fewer wasted sends, quicker convergence, clearer confidence - compounds over time. Simple tools, thoughtfully combined, still win. https://lnkd.in/dW-emEwC

30

4 Comments

Dr. Ramla Jarrar, an MMM expert, shares an important perspective every marketer using MMM should read. The customer journey is complex, and while MMM is a powerful tool for understanding how spend impacts that journey, its limitations are often overlooked and nuances brushed over once results are packaged in polished decks. The tension I often see: marketers need one strong KPI to track large marketing investments, but the complexity of measurement means no single framework provides that holy grail metric. Measurement methodologies are sophisticated and technically robust, yet each has constraints. Triangulating across multiple frameworks—MMM, incrementality testing, attribution, brand tracking—provides better insights, though interpretation remains as much art as science. This requires partnership. Marketers need to understand what each methodology can and cannot deliver. Analytics partners, whether internal teams or external consultancies, have the responsibility to explain assumptions, acknowledge limitations, and help interpret results responsibly. When both sides engage with this complexity honestly, we make better decisions. A question for marketers: where do you find the biggest gaps between marketing performance measurement and the questions you need answered? #MarketingMixModeling #MMM #MarketingAnalytics #MarketingMeasurement #MarketingROI #DataDrivenMarketing #MarketingStrategy

1

Another important movement in advertising is getting back to the basics of marketing... Marketing fundamentals like human attention, memory and persuasion that has made consumer experiences better, and businesses more successful in establishing a connection with consumers. thank you Joseph Meehan for calling out these critical signals is driving performance!!

6

Everyone is talking about #AgenticAI. Here is some news you can use. Personally I love the agent GPA idea. "Innovations from Snowflake’s AI Research Team make Snowflake Intelligence up to three times faster on text-to-SQL queries, delivering real-time answers with the same trusted accuracy. To increase the trustworthiness and accuracy of responses, the team also pioneered a novel evaluation method coined the Agent GPA (Goal, Plan, Action) framework that catches up to 95% of errors when tested on standard datasets, achieving near-human levels of error detection." https://lnkd.in/eRqVaxBy #Agentic #AI

13

🎯 𝗖𝗮𝘂𝘀𝗮𝘁𝗶𝗼𝗻 > 𝗖𝗼𝗿𝗿𝗲𝗹𝗮𝘁𝗶𝗼𝗻 (10 and final) 𝗧𝗵𝗲 𝗢𝗽𝘁𝗶𝗺𝗶𝘇𝗮𝘁𝗶𝗼𝗻 𝗠𝗶𝗿𝗮𝗴𝗲 Everyone is excited about “autonomous marketing.” LLMs writing copy. Agents orchestrating journeys. AI deciding what every customer sees. It feels like the future. Most of it is optimizing inside a hallucination. I call it The Optimization Mirage. 𝗧𝗵𝗲 𝗳𝗶𝗿𝘀𝘁 𝗺𝗶𝘀𝘁𝗮𝗸𝗲: 𝗰𝗼𝗻𝗳𝘂𝘀𝗶𝗻𝗴 𝗱𝗲𝘀𝗰𝗿𝗶𝗽𝘁𝗶𝗼𝗻 𝘄𝗶𝘁𝗵 𝗱𝗲𝗰𝗶𝘀𝗶𝗼𝗻Churn scores. LTV models. Propensity rankings. These are descriptive models. They contain zero intelligence about what will happen if you intervene. In lifecycle marketing, their role must be narrow: • As surrogate outcomes for experimentation when true LTV is delayed • As context features inside a real decisioning engine They are inputs. They cannot be policies. Using descriptive models as decision engines is like trying to fly a plane by reading yesterday’s weather report. 𝗧𝗵𝗲 𝘀𝗲𝗰𝗼𝗻𝗱 𝗺𝗶𝘀𝘁𝗮𝗸𝗲: 𝗹𝗲𝘁𝘁𝗶𝗻𝗴 𝗟𝗟𝗠𝘀 “𝗱𝗲𝗰𝗶𝗱𝗲” The natural extension! An LLM can: • Generate treatments • Embed context • Summarize history It can tell you what usually happens together. But it cannot reason about why a treatment caused which outcome. Correlation does not translate to policy. 𝗧𝗵𝗲 𝘁𝗵𝗶𝗿𝗱 𝗺𝗶𝘀𝘁𝗮𝗸𝗲: 𝗧𝗵𝗲 𝗛𝗼𝗿𝗶𝘇𝗼𝗻 𝗧𝗿𝗮𝗽 Even the most state-of-the-art reinforcement learning systems must initialize somewhere. If you let an LLM initialize treatments based on correlational patterns in your warehouse, you fall into: The Horizon Trap. It looks like intelligence because it reflects past patterns. But those patterns are not causal. They are status-quo artifacts. The decisioning will converge. But to what? A local maximum defined by correlational guesses. You can dynamically allocate traffic across five bad ideas and still lose money. The algorithm didn’t fail. Your initialization horizon was wrong. 𝗪𝗵𝗮𝘁 𝗿𝗲𝗮𝗹 𝗱𝗲𝗰𝗶𝘀𝗶𝗼𝗻𝗶𝗻𝗴 𝗿𝗲𝗾𝘂𝗶𝗿𝗲𝘀 If you want to avoid: • Boomerang effects • Slow convergence • Local maxima traps You need causal priors, and decisioning belongs in constrained, auditable, outcome-driven systems. Have you seen LLM-driven “autonomous” systems produce real incremental lift — or just elegant automation?

19

3 Comments

AI frontier vision models delivered a breakthrough I didn’t see coming. I’ve been prototyping a Databricks native AI pipeline to explore how AI frontier vision models interpret creative assets. Using UC Volumes and multiple model families (GPT‑5, Gemini 2.5, Claude, Llama, Gemma), the workflow converts visuals into structured creative attributes. The most interesting part has been comparing how different models interpret the same asset and what that means for creative intelligence and pattern analysis. Seeing the differences across models surfaced an insight that reshaped how I think about this space. Early signals point to meaningful opportunities for future analytics workflows: ✅ More consistent creative metadata ✅ More interpretable model outputs ✅ More opportunities for pattern discovery Curious what others are seeing as they push deeper into AI vision work. #analytics #ai #databricks #visionmodels #creativeintelligence #machinelearning

6

Data quality isn’t negotiable at Ipsos. AI can accelerate insights - but unmanaged, it can threaten research integrity. We use AI responsibly to protect clients’ studies from AI-driven risks and keep results trustworthy. Outcome: clean, reliable data, delivered fast. Check out the carrousel to see how we are safeguarding data quality as AI evolves. 👇 #MarketResearch #DataQuality

1