Predictive research and DIY: What we've learned building in this space
Editor's note: This article is an automated speech-to-text transcription, edited lightly for clarity. For the full session, please watch the recording.
EyeSee has spent years developing a successful AI-powered packaging research solution. A process that has taught the organization a lot about building in this space.
During the Quirk’s Virtual Sessions – DIY series on September 24, 2026, Iva Tasovac, research capability director, and Marija Djordjevic, product portfolio director at EyeSee shared some of what they learned during the development of PackSee.AI.
Tasovac and Djordjevic start from the beginning of the development journey, sharing a bit about the process of construction, testing and client feedback.
Session transcript
Joe Rydholm
Hi everybody, and welcome to our session, “Predictive research and DIY: What we've learned building in this space.” I'm Quirks Editor, Joe Rydholm. Thanks for joining us.
I just wanted to quickly go over the ways you can participate in today’s discussion. You can use the Chat tab to interact with other attendees. And you can use the Q&A tab to submit questions to the presenters and they will answer them offline via email after we are through.
Our session today is presented by EyeSee. Enjoy the program!
Iva Tasovac
Hello and welcome. Thank you for joining us.
We are Marija Djordjevic and Iva Tasovac from EyeSee. Marija looks after the development of our research portfolio, while I work closely on the development of our predictive and do-it-yourself solutions.
At EyeSee, we've spent years combining behavioral research with technology, and that work has recently included PackSee.AI, our new addition to the product portfolio. PackSee.AI is a predictive pack screening solution built on more than 12 years of behavioral expertise and 15+ million of human data points. EyeSee was also ranked fifth among technology providers in 2025 with most innovative suppliers report.
So today, we want to use some of what we've learned firsthand and share our knowledge with you.
Marija Djordjevic
Thanks, Iva. And thanks everyone for joining us. We're going to look at how predictive research has evolved over the past couple of years, what we have learned from building in this space and where DIY fits into that picture.
A couple of years ago, predictive AI and research was still something the industry was trying to get its head around. There was a lot of excitement around what could it make faster and cheaper, but there was just as much caution. Clients wanted to know what was actually being predicted, what the models had learned from, how reliable the outputs were and whether a result that looked convincing on a dashboard would hold up when it was tied to a real business decision.
That caution made sense. Research is not only about producing an answer, it is about being able to explain why you trust that answer, what it can support and where its limits are.
So, although predictive tools were clearly going to become part of the research landscape, the industry was not yet at the point where everyone was ready to use them without asking a lot of questions first.
At the same time, the attraction was obvious.
If you could use years of existing behavioral data to predict how a new pack, communication or visual would perform, you could screen more options, learn earlier and avoid spending a full research budget on every small decision.
That created a logical link between predictive research and self-service from the start. If the answer can be produced faster and without going through the process complexity of conventional respondent-based market research, sooner or later, people will ask why they cannot access it directly.
But for us, that was not the first question. The first question was whether the technology was good enough to be trusted.
Iva Tasovac
When we started developing PackSee.AI, we already had a strong base to work from. The models were built on years of behavioral research and a large body of data, but we did not want to treat that as permission to launch everything at once.
We wanted to put one focused product into the market, use it on real projects and stay close enough to the process to understand how it behaved outside the internal development environment. That meant keeping our researchers involved.
We were still seeing the inputs, checking the outputs, talking clients through the results and watching what happened when those results were used in practice. Part of that was quality and control, but part of it was simply learning.
We wanted to know whether the questions clients brought to the tool matched the questions we had designed it to answer. We wanted to see where the workflow felt natural and where it created some frictions. We overall wanted to hear what clients trusted immediately and what needed explanation.
We were also cautious of the wider mood around AI at the time, because today most research teams have had some exposure to AI tools. Two years ago, many were still deciding how comfortable they were with them. So, there was a learning curve on both sides.
We were learning how the product worked in the real world while clients were learning where predictive AI could fit into their own research process. So, taking a measured approach gave both sides room to build confidence.
Marija Djordjevic
The first months after launching PackSee.AI taught us something quite basic but very important.
A good model does not immediately guarantee a good research product. Internal validation can tell you whether the model performs, but it cannot tell you everything about how people will use it, what will be immediately understood and which parts can lead to confusion or even where the process around the model still depends on expert judgment.