AudienceMatchAI: Using LLMs for Marketing Audience Segmentation
In a COLLIDE 2024 session, Softcrylic showed AudienceMatchAI, an early-stage tool that runs unsupervised customer segmentation and lets marketers question the segments through a language model. The team said limiting the data the model sees reduces hallucinations and makes answers easy to check.
By The AI Enterprise Magazine Team4 min read
Key takeaways
- Segments become personas. A language model turns heat-map statistics from clustering into short, readable personas.
- Less data, fewer hallucinations. Giving the model only segment-level data makes answers easier to spot-check against tables.
- Unsupervised by design. Softcrylic chose clustering so the data, not analyst assumptions, defines the audiences.
- Addressable audiences. Segments built on first-party data contain real customers who can be pushed to marketing platforms.
- Augment, not replace. Presenters framed the tool as a time saver for marketers, with performance data and image generation on the road map.
At COLLIDE 2024, the COLLIDE Data & AI Conference in Atlanta in October 2024, marketing analytics consultancy Softcrylic demonstrated AudienceMatchAI, an early-stage generative AI tool that sits on top of machine learning customer segments. In a session sponsored by Softcrylic, Brad Kossmann, Senior Director of Data Science & Analytics, and colleague Morgan Depetro showed how marketers can question their segments in plain language, generate personas and draft campaign ideas. The pair said at the time that the product was "very much in the early phases" of development.
What is AudienceMatchAI?
Softcrylic describes itself as a marketing analytics consultancy with three practices: data engineering for marketing data, analytics, and marketing technology that connects a client's first-party data to the wider martech ecosystem. AudienceMatchAI sits near the end of what Kossmann called the "data journey," where a company defines the audiences it wants to reach.
Kossmann said marketing data science rests on two pillars, audience and attribution, and the session covered audience. Companies have collected transactional, digital and marketing data in warehouses for decades, he said. Turning that data into one-to-one marketing is hard because the data sets are large, rarely clean and difficult to connect to the channels where customers are targeted.
How does the tool turn customer data into audience segments?
The workflow Kossmann outlined has four steps. A user loads a customer data set, such as transactions, digital behavior and demographics. A proprietary automated machine learning layer then runs the segmentation with one click. The segment outputs are passed to a large language model. Finally, the user exports the segments as addressable audiences.
Softcrylic chose unsupervised learning for the segmentation. "We like to let the data speak for itself," Kossmann said, so the team does not introduce its own biases into the model.
Because the segments are built from first-party data with customer identifiers attached, each one is made up of real customers, he said. That makes it possible to push a segment into a marketing platform or build lookalike audiences from it.
How does a language model help with personas and strategy?
A clustering run produces descriptive statistics, often hundreds of numbers in a heat map. Kossmann said that is hard to use for marketing, where teams need shorthand. The language model reads those figures and turns them into short, human-readable personas.
In the demo, Depetro showed segments with names such as "young budget conscious regulars" and "affluent mid agers," each with a description, quantitative metrics and top categories. She then asked questions in a chat window. One request asked for a concise summary of all the segments that could go to an executive leadership team. Another asked which segments would be most receptive to a goal of increasing in-store foot traffic, and why.
Kossmann said that building personas by hand, with a team looking at a heat map, can take days. He said the segmentation and preprocessing steps alone represent a few weeks of manual work. Teams that already hold psychographic profiles can feed those in so the model maps them onto the data-driven segments, he added.
How can marketers reduce AI hallucinations?
Depetro said the way the data is exposed to the model "is meant to reduce hallucinations." Kossmann called this the main takeaway. The less data the model is given, he said, the less likely it is to hallucinate, and the easier its answers are to check.
The demo showed the cross-check. Depetro asked for the average purchase amount of one segment, and the answer matched the figure in the segment description and the metrics panel. Kossmann said a tool that is known to fail or give false information loses organizational buy-in. He added that narrow niche use cases, with retrieval-augmented generation tuned to the task, worked better than broad analytics tools.
What did the demo show about campaigns and diagnostics?
Depetro asked the tool for marketing copy aimed at the budget-conscious segment, then for creative imagery ideas. The first copy answer was long, so she asked for shorter versions. Kossmann said that kind of iteration gets a team past the first hurdle of having ideas on the page to refine.
She also tried to "trick" the tool. She said a sundress campaign aimed at the hoodie-buying segment had performed poorly and asked why. The tool pointed to a product mismatch.
On the road map, Kossmann said the team planned to add campaign performance data, so a user could ask why an email's click-through rate fell for a segment. It also planned to add reasoning models that show their thinking and multimodal output that creates images rather than describing them. He described the product as a time saver for marketers, "a UI for your data" rather than a replacement for them. In his words, it opens the data to more users who do not write SQL.
COLLIDE 2027 returns to Atlanta
COLLIDE 2027, the largest AI conference in the Southern U.S., will bring 1,000+ enterprise data and AI leaders to the Sandy Springs Performing Arts Center in Atlanta on Wednesday, October 13, 2027. Details are on the COLLIDE event page, and registration is coming soon.
Frequently asked questions
What is AudienceMatchAI?
A Softcrylic tool, in early development at the time, that runs machine learning segmentation on customer data and lets marketers query the segments with a language model.
How does AudienceMatchAI reduce hallucinations?
By feeding the model only segment-level data, which makes hallucinations less likely and lets users cross-check answers against the tabular metrics.
Why use unsupervised learning for audience segmentation?
Softcrylic said it lets the data speak for itself and avoids introducing the modelers' own biases.
Can the segments be used for campaigns?
Yes. Because they come from first-party customer data, the presenters said each segment is made of real customers who can be exported to marketing platforms.
This article is based on remarks by Brad Kossmann, Sr. Director of Data Science & Analytics, Softcrylic; Morgan Depetro, Softcrylic (title not given) at COLLIDE Data + AI Conference 2024 on October 10, 2024. More coverage: COLLIDE AI Conference.
Filed under: AI Use Cases & Applications, Observability & Data Quality
