It seems like anywhere you looked at Snowflake Summit 2026, you encountered the same core message: that the biggest barriers to enterprise AI have less to do with models or compute and more to do with underlying data foundations.
Few sessions illustrated this more clearly than “Consistency Before Intelligence: The Hidden Work Behind AI Analytics,” where Canva shared what it takes to build reliable AI-powered analytics at scale.
The takeaway from that success story was surprisingly simple. You can’t build trustworthy AI on top of inconsistent metrics.
The Real Problem Isn’t AI. It’s Agreement.
Let’s start with a snapshot of the enterprise. Canva processes more than 400 TB of data every day, serves 265+ million monthly active users, and sees 440 new designs created every second. Yet despite that scale, a listening tour with data scientists and product managers uncovered a familiar problem: “Metrics logic is scattered. There is no single source of truth.”
The issue appeared in two ways: the same metric meant different things to different teams, and duplicate definitions existed across dbt models, BI tools, and ad hoc queries. “Retention,” for example, could have multiple definitions depending on who was calculating it. For humans, this creates confusion. For AI, it creates unreliability.
Why Inconsistent Metrics Break AI
There’s a growing assumption that organizations can simply place an LLM on top of their data warehouse and immediately unlock intelligent analytics. Canva’s experience suggests matters aren’t quite so simple.
When AI encounters multiple definitions of the same metric, it can select the wrong one, generate SQL users can’t validate, and deliver answers with false confidence.
This creates what researchers call algorithm aversion: people are often less forgiving of AI mistakes than human ones. Even if an AI assistant is correct most of the time, a few high-profile errors can quickly erode trust.
As Canva put it, reliability is the number one requirement for AI adoption.
The Semantic Layer Is Becoming Strategic
The touchy subject of reliability is where semantic layers entered the conversation. Historically viewed as BI infrastructure, semantic layers are increasingly becoming foundational AI architecture. Canva described its semantic layer as a centralized place for metric definitions, business descriptions, join and filter logic, example queries, and contextual instructions for AI.
The analogy shared during the session was particularly effective: a semantic layer acts like an onboarding guide for a highly capable intern. Without guidance, even the smartest employee will make assumptions and occasionally get things wrong. AI works the same way.
Designing Data for AI Consumption
Canva’s architecture is deliberately structured around business domains. Each metric belongs to a single domain and is exposed through curated data services with defined ownership and service levels.
Foundation services provide canonical datasets, conformed services model shared business concepts, metric services expose dimensional models, and the semantic layer captures business meaning.
The result is a single, governed interface that serves both traditional analytics and AI experiences—a pattern that is becoming increasingly common among organizations pursuing agentic AI.
The Results Were Significant
The impact of adding semantic context was substantial. Without semantic views, error rates ranged between 10% and 40%. With semantic views in place, accuracy exceeded 90% for low- and medium-complexity questions, response times dropped to under a minute, and query costs were 57 times lower than using a standalone LLM approach.
The session also highlighted an important limitation: high-complexity questions such as “Why is this metric moving?” remain difficult for AI to answer reliably. Encouragingly, systems like Cortex Analyst increasingly know when not to answer, refusing questions that fall outside their confidence boundaries rather than generating misleading responses.
AI Readiness Is an Organizational Challenge Too
Technology alone wasn’t enough. Canva also created an Engineering Council to establish data architecture standards, metric definitions, and organization-wide alignment.
Their reasoning was straightforward: if AI works exceptionally well in one part of the company and poorly in another, users won’t distinguish between implementations. They’ll simply conclude that AI doesn’t work here.
Consistency has to become an organizational discipline, not just a technical one.
The Bigger Lesson from Snowflake Summit
Across Summit sessions, one pattern continued to emerge: the organizations making the fastest progress with AI are not necessarily those with the largest models or the most sophisticated architectures. They’re the organizations that have done the foundational work first—standardizing definitions, establishing governance, building semantic context, and designing trusted data products.
This makes sense when you begin to understand that AI isn’t there to solve inconsistencies in your data. If anything, it runs the risk of amplifying them.
The future of AI-powered analytics will belong to organizations that invest in consistency first, because intelligence without trust is difficult to scale.
Building an AI-Ready Foundation
As enterprises move toward agentic analytics and natural-language interfaces, the question is no longer whether AI can query your data. It’s whether AI understands what that data means.
At Hakkoda, an IBM Company, we help organizations build trusted, governed, and AI-ready data foundations, from semantic modeling and modern data architecture to enterprise AI enablement. If you’re exploring how to make your analytics platform ready for the next generation of AI, let’s talk today.
