AI: Asking the Right Questions
by Joseph Galarneau, CivicScience
Every day, CivicScience interacts with more than 10 million people to understand the mind of the American consumer for major customers such as Target, T-Mobile, Bank of America and Microsoft.
Machine learning (ML) has helped us to make sense of that massive flow of data since CivicScience’s founding, when it was fledged from a Carnegie Mellon technology incubator. But even so, the recent pace of progress has been astonishing — breathtaking, honestly — to those of us who’ve been building technology products for decades.
We’ve always had the view that CivicScience’s crown jewel is the data entrusted to us by consumers — four billion answers to polling questions over the past decade — and technology is the way that we amplify its value in ways that no one else can.
While our parcel of very talented PhDs and engineers could handily build foundational AI technologies, we are resolute in our focus as a marketing data technology company, and we’ve increased the pace of development to ensure our products are taking advantage of these innovations. Companies that choose to ignore recent AI advances do so at great peril.
AI is a squishy term that has a grand meaning to the layperson, but to the technologist, it’s mostly an umbrella of different types of ML. Supervised ML, which has been around for a while, marries statistical methods with computing power to predict, classify and find relationships among structured data elements. GPT and Google Bard use a different type of ML — a mix of unsupervised and reinforcement learning — that finds patterns in unstructured data and gets better automatically via feedback loops.
CivicScience is doubling down on both supervised and unsupervised ML in three main areas: making sense of data; automating manual processes; and creating new and richer interactions with customers. While there are dozens of AI touchpoints throughout our products, the most recent success uses GPT to write polling questions.
CivicScience partners with leading national publishers such as Microsoft News, Penske Media, and Nexstar and local news sites such as The Pittsburgh Tribune-Review and WTAE to place polls on news articles. We ask 20,000 unique question/answer sets every day from a database of nearly a half million.
The first question in any poll is used to engage the user by focusing on the article’s content, after which we ask more detailed research questions. Our team of talented editors skim articles to create these engagement questions within hours of the page’s publication.
Powered by GPT, QGen simplifies this by scanning publishers’ sites to detect new articles within minutes, “reading” the article, and automatically writing custom questions and answers. Our editors then review them to ensure they’re safe for publishers’ pages, adjusting for accuracy and tone when needed (as GPT goes off the rails from time to time).
The result has been a 10x improvement in productivity with the initial release, freeing up time for editors to focus on other projects. We have the eventual goal of writing bespoke questions for the million daily unique web pages in the CivicScience network, which would have required us to hire 160 more editors if we didn’t have QGen.
Another GPT application, which will soon be in beta, supports “data democratization” to spread CivicScience data within customer enterprises beyond the users of our core InsightStore SaaS application. A CMO or other executive will be able to ask a complex, natural language question via Slack or other chat platform and receive back a detailed, plain-English response driven by our data, including charts and data extracts.
Using CivicScience’s real-time pulse on consumer sentiment, AI also can predict the future. Our data science and economics team has been converting five million daily polling responses into models that accurately predict economic indicators such as unemployment or housing sales, sometimes months ahead of official statistics releases. We’re also working with customers to do the same with internal metrics that matter to them, such as same-store sales or demand for specific products.
The more quotidian uses of AI fit cleanly into our normal product development process. For more speculative applications, we’ll run a proof-of-concept with a tightly scoped problem, clear success criteria, a small tiger team, and a four to six week window. With minimal investment, this allows us to better understand promising technologies not ready for primetime, as well as to better qualify innovations with near-term business impact that can be fast-tracked into full development.
One beauty of the recent developments in AI is the dramatic increase in accessibility and that we’re only at the beginning of the J-curve. No longer do you need a team of highly skilled engineers and scientists to take advantage of cutting-edge tech. With a bit of education and a passion for exploration, any company can find some way to make AI work for them.
