A Non-Technical Guide to AI Models for Insights Professionals
What you need to know to capitalize on the opportunity
When people think AI, they think ChatGPT. Fair enough, it's the celebrity tool of artificial intelligence right now. But that celebrity status has led a lot of folks (especially in the business world) to think that all AI is a chatbot. Or worse, that all AI needs to be a chatbot.
Let’s dispel that myth.
Imagine a fast food drive-thru. You talk to a speaker, place your order, maybe have a brief, awkward back-and-forth. That speaker? That’s your large language model (LLM), like ChatGPT, the chatbot we all recognize. It’s a smart, capable interface. But as we all know it’s not the kitchen. It’s not flipping your burger or pouring your drink. Behind the scenes, a lot of specialized systems are working together to get your order right. Same goes for AI. When you ask ChatGPT to generate an image the chatbot doesn’t generate the image, a different model behind the scenes handles that, this is the root of the confusion; like the drive thru speaker, the LLM is a window into the world of AI.
Large language models might be how you interact with AI, and the fact that LLMs are fueled with the knowledge of the internet makes them extremely smart and mind blowing, but the real magic and some of the coolest AI applications are often found behind that metaphorical kitchen door.
I put this document together to provide some guidance on the major technologies fueling the AI revolution so those in the insights industry don’t end up designing a restaurant and forget to hire a chef. Admittedly, we all love to eat at restaurants but have little interest in doing the cooking so while this topic can go deep I've broken this article in two to make it as digestible as possible (okay, enough food metaphors).
The super easy to understand version - below,
The more nuanced yet still approachable version - part 2 article,
What’s next in AI - part 2 article.
Each part is intended to be easy to understand, for that reason I may leave some big things out, but I will endeavor to provide some links to useful well structured videos for those interested in learning more.
To get ahead of the inevitable question, why does this matter? Some of you may think I’ll just write a check to MSFT, Google, OpenAI and use what they give me and I’m done. And for some of you that might be enough. But if you plan to bring AI into your insights pipeline, want to do so at a reasonable cost, and have to navigate the inevitable requirement of big customers (i.e. keeping their data secure), you may end up needing a team(s) to look at how to bring this tech in house, hopefully the info below will help you in some small way from keeping them talking over your head.
The Three Big Use Cases for AI
Before diving into the models, let's quickly set the stage by outlining the primary ways AI is transforming the insights industry. Broadly speaking, we can categorize these transformations into three major groups.
Storytelling Enhancers
First up, we have storytelling enhancements, focused on elevating how we share insights with end users. AI steps in here by interpreting data, crafting insightful reports, or even creating engaging podcast-like narratives to clearly communicate findings. Beyond reports, AI helps craft targeted, impactful content, like ads or persuasive copy, making insights directly actionable in business contexts. This is a world where you can chat directly with your data as if you're talking with someone from the audience you've measured. Say goodbye to crosstabs.
Methodological Advancements
The second group covers methodological advancements, the area where AI reshapes what's possible. AI-driven moderation of focus groups, computer vision for coding visual data (e.g. receipts, photos) or maps of intricate networks of influence are just a few benefits. We’re already glimpsing a future where insights become increasingly predictive and prescriptive rather than simply reporting historical data. Soon, AI will unlock entirely new possibilities for interactive feedback, leveraging immersive experiences through VR and AR, fundamentally shifting research from passive observation to active participation.
Operational Delivery
Finally, there's operational delivery, the essential behind-the-scenes work that insight consumers typically never see. Here, AI makes its mark by streamlining processes such as respondent selection, dynamically crafting better surveys, efficiently managing fraud detection, and automating tedious data processing tasks. It can also continuously monitor data for anomalies, spotting issues long before humans might notice. While these enhancements stay mostly out of sight, their impact is profound, resulting in savings and enabling teams to spend more time on strategic thinking.
With these use cases as the backstory let’s dig into the models.
Big AI Models in a Nutshell
Someone will fight me on this but I’m going to keep my list to the 10 big technologies I see as powering the most interesting advancements in AI and here they are:
Transformer Models (GPT):
Acts like a smart assistant that can read a book, write a summary, and tell you what’s likely to happen in chapter 12. Or better yet, that best friend who knows you so well they can complete your sentences for you. This is how ChatGPT knows what to say, they’re really good at predicting the structure of sentences.
How to use for insights?
Excellent for interactive data exploration, allows users to easily dig through complex datasets through natural conversation rather than complicated interfaces. Effective for simulating realistic consumer personas for deeper qualitative insights. These models can automate reasoning processes, automatically generating robust insights reports, and improving qualitative research engagements. Additionally, they help innovate traditional survey methods and efficiently generate compelling, consumer-targeted advertising copy.Causal Inference/Counterfactual AI:
Causal inference is like a coach watching a basketball game replay to figure out if the star player’s three-point shots really helped the team win, or if other things, like great defense or the other team’s mistakes, were the real reason. Counterfactual AI is like rewinding the basketball game in your head and imagining what would’ve happened if the star player took a dunk instead of a three-point shot. It’s like the computer guessing if the team would’ve still won or lost, based on how the game went, to help plan better plays next time. This is how Amazon targets you with advertising that's designed to get the highest conversion rates.
How to use for insights?Ideal for strategic, prescriptive insights by not only identifying what’s happening, but deeply understanding the reasons behind observed phenomena. They excel at scenario planning, determining outcomes under hypothetical conditions. Great for strategic marketing effectiveness and optimize future ad spend (Mix Modeling). Also useful in isolating the precise impacts of specific product features, targeting consumers with tailored marketing, diagnosing root causes of performance shifts, and creating sophisticated media planning scenarios that drive desired business outcomes.
Convolutional Neural Networks (CNN):
Kinda like someone that can look at a puzzle piece and know exactly what part of a larger image it belongs to. Really great at understanding how the small parts of an image is related to larger images. This is how CNNs help Tesla's drive themselves.
How to use for insights?
Outstanding at interpreting visual data, especially for evaluating and isolating elements within advertising creative to predict performance. They can also efficiently monitor brand presence and context in media content and product reviews, significantly improving marketing responsiveness. CNNs provide real-time facial and emotional analysis during consumer interactions, accurately capture shopping behaviors via video feeds, and streamline receipt analysis for market insights and tracking consumer purchase behaviors.Generative Adversarial Networks (GANs):
Imagine one kid tries to draw fake Pokémon cards and another kid tries to spot which ones are fake. Over time, the drawing kid gets better until even the best Pokémon expert can’t tell the difference. This is how GANs create pictures of people that aren't real.
How to use for insights?
Great for identifying fraudulent activities by spotting data anomalies that deviate from expected patterns. They can generate realistic synthetic data to boost samples, and can enable sharing of sensitive datasets without exposing actual consumer data. GANs support predictive modeling in marketing mix and financial analyses, enabling scenario testing of potential media strategies and outcomes.Diffusion Models:
Like a sculptor. They know what they want and start with a block of stone in front of them. They slowly remove the excess and eventually what once looked like a formless shape is the Statue of David. This is what a diffusion model does for image generation starts with a fuzzy representation and gradually improves it until it generates the image you’re looking for.
How to use for insights?
Effective in stable synthetic data creation with greater diversity and realism compared to other methods, useful for augmenting limited or sparse datasets. Ideal for producing hyper-realistic visual product mock-ups and advertisements for early-stage consumer research and testing. These models can create consumer avatars grounded in extensive insights data, allowing for deeper empathy in customer strategy. Additionally, they can efficiently perform advanced data imputation, accurately filling in missing or incomplete datasets.Reinforcement Learning (RL):
Just like giving your dog a treat when they sit on command and time-outs for when they're bad. The rewards given to the AI eventually get it closer to being able to figure out how to do something. This is how you teach your Nest Thermostat what to set the temperature to in your home, every time it gets it right you ignore it (positive reinforcement) and every time it gets it wrong you change the temp (negative reinforcement).
How to use for insights?
Useful for dynamically pairing research respondents with the most relevant survey tasks, improving response quality. RL can create highly personalized respondent experiences, increasing engagement and data accuracy. Ideal for continuously optimized testing scenarios such as pricing strategies and creative effectiveness. They help automate ongoing insights generation, adaptively improving based on feedback. They can provide media allocation recommendations to maximize real-time campaign performance.Self-Supervised Learning (SSL):
Imagine a curious explorer in an unlabeled museum filled with artifacts. Instead of a guide labeling each item (“this is a vase”), the explorer studies the objects, grouping similar ones (e.g., noticing that shiny, curved items are often pottery) and predicting what missing pieces might look like. Through this self-guided exploration, they build a deep understanding of the museum’s patterns. Later, when tasked with identifying a specific artifact or category, they can quickly apply their knowledge with minimal extra help. This is how Siri is able to understand multiple accents and pronunciations.|
How to use for insights?
Perfect for unifying diverse, inconsistent survey datasets into coherent datasets without extensive manual effort. Excellent for advanced segmentation, identifying naturally occurring consumer groups and trends within complex raw data. These models enrich raw data by automatically adding labels and context, greatly improving the effectiveness of downstream AI tasks.Graph Neural Networks (GNN):
Imagine a GNN as a gossip network in a small town. Each person (node) knows a bit about themselves (their features) and talks to their friends (edges). By sharing and combining gossip with their neighbors, everyone learns more about the town’s social scene (e.g., who’s popular or who might get along). After a few rounds of chatting, the GNN can predict things like who’ll be invited to a party (node task) or whether two people will become friends (edge task), based on the town’s connections. This is how Instagram is able to make eerily accurate recommendations of people to connect with.
How to use for insights?
Powerful for analyzing interconnected datasets, delivering advanced segmentation by identifying groups defined by behaviors, preferences, and social interactions. GNNs reveal novel insights through deep understanding of subtle yet impactful data relationships, enhancing traditional analysis methods. Ideal for insights with multiple data streams. Use to track influence networks and identify key individuals or products driving market trends or purchases.Multimodal Models:
A single mode model is like seeing a recipe in a magazine whereas a multimodal model is like having a talented chef who can combine ingredients like meats/vegetables/spices, innovative cooking techniques, and world class presentation to create a delicious dish. By tasting and adjusting each part while considering how they work together, the chef produces a meal that’s perfectly balanced, just like a multimodal model blends different data types to give smart answers or create new content. This is how Google can look at a photo and tell you the breed of a Dog in the photo.
How to use for insights?
Best suited for nuanced sentiment analysis, accurately interpreting complex emotional cues across video/images and text including sarcasm, irony, memes, and other non-traditional expressions. Multimodal models enable sophisticated contextual ethnographic research, seamlessly integrating qualitative video/photo data with quantitative metrics. These models significantly enhance customer experience tracking, combining analytics data, audio from customer interactions, and user-generated content into comprehensive insights. They also excel in analyzing advertising and product performance by simultaneously considering visuals, audio, and consumer reactions.Neural Radiance Fields (NeRF):
Imagine taking a ton of photos of your bedroom and giving them to a computer than uses the photo to build a version you can navigate in VR. This is how Google improves street view into immersive experiences. The latest season of Black Mirror on NetFlix has an interesting take on this tech.
How to use for insights?
Currently a niche application in insights, NeRF models are ideal for creating immersive, realistic 3D product models from simple 2D images, greatly enhancing product engagement in consumer research. Perfect for virtual reality research applications, enabling realistic and highly interactive environments for testing store layouts, shelf arrangements, or product placements.
Again, Why Does This Matter?
In the grand scheme of things, if you’re not building insights applications or organizing your company’s data, this mostly doesn’t matter. That is, other than in the context that you should take away from this that there’s more to AI systems than “GenAI”.
Worst case scenario you now know more than 99.9% of business leaders leaning on their teams to build new AI applications. Best case scenario you realize that these new AI models are the tip of the iceberg and the business of the future will likely take advantage not just of one of these technologies but many of them.
There’s a whole menu of AI tools beyond chat. Some of the most powerful applications aren’t conversational at all. They're engines working quietly in the background to:
Predict outcomes
Simulate possibilities
Spot patterns
Make better decisions
Personalize experiences
Reconstruct environments
Understanding these building blocks helps you contribute to building an AI strategy that fits your business. Sometimes the best new feature is a shiny experience for customers (e.g. chat with data), but in the long run dollars will flow to solutions that positively impact the P&L (e.g. more efficient marketing spend).
So yeah, ChatGPT might be the drive-thru window, but AI’s real value might just be what’s happening in the kitchen.
To dig in a little deeper (still accessible) and learn a little about what’s on the horizon check out part 2.