
If you ask an AI tool why your brand lost share last quarter, you’ll probably get a fast answer. It will be well-phrased, it will carry a tone of authority, and it might even look right at first glance. But look closer for meaning and impact beyond the string of words: Can you actually defend this?
There is a crucial distinction between a plausible answer and a defensible one. A plausible answer looks or sounds correct. A defensible answer holds up when someone asks where the number came from. It aligns your organization around the real risks and opportunities in pricing, promotion, distribution, and more.
AI is a remarkable partner and asset, but not all AI is equal because not all data is equal. What the model is built on determines the kind of answer you get, what you do next, and how that works out for your business.
Is Data Still the Biggest Barrier to AI in Retail and CPG?
The real barrier in today’s operating environment is trust, not the volume of data. After all, most manufacturers and retailers have plenty of data.
That said, even the most talented analysts can't defend a number built on data nobody can verify. Fortunately, there are ways to gauge trusted data.
Organizations can start with four questions:
Can each number be traced back to its source?
Is it granular enough, down to the UPC and the store week, to answer the question you're asking?
Does it reflect how your business defines its categories and competitors?
Is it recent enough to act on?
If you hesitate on any of these, your AI will give you plausible answers, and maybe even good ones. They may not be defensible.
Where Does an AI Analytics Platform Get Its Numbers?
An AI analytics platform gets numbers from whatever data it's built on. But that foundation decides everything; if the data is verified, granular, and traceable, the answers can be, too. If it isn't, a model won’t save you.
One shortcut has become common, and it backfires. A company might pull raw data extracts into a data lake and point a generic AI tool at it. It looks efficient, but usually fails because the extracts lack richness. They're missing UPC-level detail and the custom hierarchies on which the business actually runs. The AI also can't correctly calculate non-addable facts, the measures you can't simply sum across stores or products. So it does what models do with gaps — it just fills them.
Data Provenance and Data Lineage
Data provenance is where your data originated: who collected it, how, and from what source. Data lineage is the path it took from that source to the answer on your screen, including every adjustment and calculation along the way.
You need both. When a number shows up in a brand review, someone should be able to walk it back to the store, the week, and the item. If you can't, the answer isn't defensible, however good it sounds. A trustworthy AI answer must be traceable to its source and verifiable by a person.
Verified POS Data and Decision-Grade Data
Verified information at Circana begins with census-level point-of-sale (POS) data. That means store-week, UPC-level data that is recent and highly granular. Our POS data is calibrated with more than 1,100 U.S. retail partners. We bring it together with our panel data through Multi-Outlet Bias Detection (MBD) adjustment, so the two views of the market stay consistent with each other.
Verified data is only the starting line, however. The data also has to be decision-grade, meaning it reflects how you see your market: your competitive set, your hierarchies, your flavor and pack profiles. For example, a generic system likely treats all cereal the same. But your brand team doesn't, and neither should your AI.
How Should AI and Licensed Market Data Work Together?
AI and licensed data can be used in tandem, and work best inside clear AI data governance, with usage rights spelled out and a person checking the output. The data brings accuracy, while governance keeps its use appropriate and human review keeps the output honest.
AI data governance is the set of rules for what data flows into AI tools, who can access it, and how answers get checked before anyone acts on them. With licensed market data, usage rights are part of that picture. Your license terms shape how the data can be used, including with AI. Good governance turns those terms into everyday habits, so teams can move quickly without wondering whether they're allowed to.
Then there's review. My advice is to treat AI like an intern: It's fast, it's enthusiastic, and it's often right. But you wouldn't send an intern's deck to the CEO without an expert looking at it first. AI deserves the same check, especially when it lacks the domain context specific to your business. That's usually where hallucinations creep in.
Despite conjecture swirling around the future of AI, we don’t foresee a time when the human is completely removed from the loop. AI can certainly take on the lower-value data preparation, giving people the chance to spend their time on judgment. Done well, that shift can make human decision-making 10 times more productive.
How Do Circana's Liquid Data and Liquid AI Work Together?
Circana's Liquid Data® is the verified foundation, and Circana's Liquid AI™ sits on top of it. Think of Liquid AI as a new front end to Liquid Data.
Liquid AI has three parts:
Emiri® Query pulls accurate numbers directly, without anyone needing to learn power-user interfaces. You ask a question, and it goes to the verified data for the answer.
Brainwave™ codifies your company's own domain expertise. It runs automated, repeatable analyses, like brand reviews, the way your team would run them.
Synapse™ is a curated, conversational research assistant, similar in feel to a general-purpose AI chatbot. It brings Emiri Query, Brainwave, and internet access together in one place.
Here's how that plays out in one scenario. A brand team is heading into a quarterly review. Circana’s Emiri Query pulls share, price, and distribution for the category. Brainwave applies the team's hierarchies and runs the review exactly as they've defined it. Synapse helps them dig into the why: which competitor moved, which pack lost ground. An analyst checks the work, and the team decides whether to shift promotion spend next quarter.
The time savings are real. A deck that would typically take four to six hours to build can come out of the machine in 15 minutes, with another 15 minutes spent verifying that it all looks right.
That verification step matters most. Insight only matters when it becomes a repeatable decision, and that's what these tools have been built to do.
Explore Circana's connected solutions across the decision chain — from AI-enabled measurement to price, placement, media, and marketing — and put a true competitive edge to work.

FAQs
Who may use licensed market data with AI?
The terms of your data license decide that. Before you connect licensed market data to any AI tool, check what your agreement allows, which uses are covered, and who on your team has approved access. Strong AI data governance turns those terms into simple rules people can follow day to day.
Where should a human review AI answers?
Before any output informs a decision or reaches leadership. Treat AI like an intern whose work gets checked by an expert. Look hardest where the model lacks client-specific context, where the data has gaps or suppression, and anywhere pricing or profitability is involved. That's where plausible answers most often slip through. Circana’s Liqid AI solution significantly lowers hallucinations/errors, reducing the need for human review.
What happens when an AI number is wrong?
It should get caught, traced, and fixed. With clear data provenance and data lineage, you can follow the figure back to its source, find where it went off track, and correct it. That traceability is what separates a defensible answer from a plausible one, and it's why a person stays in the loop.
How often should the data underneath an AI tool refresh?
Often enough to match the decisions you make with it. Store-week POS data gives you a recent, granular view that suits most pricing, promotion, and distribution calls. There's always some latency, though, so in fast-moving situations pair fresh data with human judgment rather than assuming the tool sees today's shelf.





























