At 9:47am, a customer abandons their cart. Within fifteen minutes, three AI agents take action to try to complete the sale. One fires a discount. Another flags lifecycle risk and triggers a retention sequence. A third surfaces a cross-sell. None of them know about each other.
The customer gets all three messages in three different channels. They’re left to decide not just what to do next, but whether it’s worth returning to a brand that won’t leave them alone.
Each marketing function is on the cusp of deploying its own agent. Decisions that used to take weeks now take seconds. And when they contradict each other, damage happens just as quickly. Speed that once felt like an advantage becomes the rate at which your customer relationships erode.
Agents need governance to help them work together instead of against each other. The Decisioning Graph fills that gap.
A Layer Beneath the Brain
This lack of coordination has been framed as an agent problem: we need better, smarter agents that can navigate competing actions. But the agents aren't the issue. The meaning underneath them is fragmented. The CDP has one picture of your customer. The MMM has another. The activation layer treats them as a third person entirely. Every agent reasoning above that contradiction in the stack inherits it.
My colleagues Nico Veltri and Pete Rogers recently described what they called the Intelligence Layer: the brain emerging between the data foundation and the execution channels, where signals get interpreted and treatments are optimized in a continuous loop. If the Intelligence Layer is the brain, the Decision Graph is the world the brain reasons about.
Where the identity graph answers "who is this person, across systems," the Decisioning Graph answers what's true about that person right now, what we could offer them, which signals take priority over others, and what we should already have learned from the last decision. It models customer states, offers, and channels as real objects, each with their own rules, constraints, and scoring logic. The model outputs that rank and score every decision—churn, LTV, propensity, uplift, MMM—live there, too.
There's a precise word for the Decisioning Graph’s role: an ontology. In philosophy, an ontology answers what exists and how things relate. In data engineering, the term means the same thing.
An ontology is what the business looks like when you write it down properly. The Decisioning Graph is the marketing-specific version of that.
Why an Ontology Native to Snowflake Changes What Agents Can Do
Writing the business down properly is an old project. What's new is the architecture to make it executable, in the same governed environment where the data already lives. Recent work by Tianxia Jia on the Snowflake AI team lays out a five-layer reference architecture—physical storage, ontology metadata, abstract views, semantic views, and Cortex Agents—that makes ontology-driven reasoning native to the platform rather than bolted on top of it.
Three things in this stack matter for marketing:
- Facts and meaning live in different layers: The physical model stays stable while the business interpretation evolves above it. With Iceberg under the storage layer, the ontology persists across open table formats.
- Meaning becomes queryable, not just documented: The same definition of "high-value at-risk customer" the analyst uses is the one the agent reads, governed by Horizon.
- Cortex Agents finally have something to ground in: Pointed at raw tables, an agent guesses. Pointed at a governed ontology, it reasons over a real model of what your business is.
The agent isn't smarter. The ground beneath it is firmer.
Three Layers, One Problem
Once the layer has a name, the rest of the architecture has somewhere to attach. Beneath the agents sits the Decisioning Graph. Among them sits orchestration: the logic that decides which recommendation wins at any given moment, which Snowflake has begun framing as a control plane for agentic work. Above them sits a quality function: observability on what every agent decided and why, drift detection, causal inference on whether decisions actually moved the metric.
That third layer is the one most likely to be skipped. It’s also where the hardest-won value lives. Measuring agents is similar to measuring any marketing action, just applied to a faster decisioning surface. Skip it and you'll deploy agents that drift confidently into incoherence, and nobody inside the organization will notice before your customer does.
Where Are You Now?
Most marketing organizations sit on a four-stage progression toward coherent agentic decisioning. The work of each stage only makes sense once the prior is real, which makes knowing where you stand essential.
Stage 1 — Fragmented. Each function operates from its own picture of the customer. Decisions are made in isolation and contradict each other in production.
Stage 2 — Unified Data. Customer data is consolidated in a governed platform. The facts live in one place, but meaning is still scattered. The same column gets interpreted three different ways. Agents deployed here will be confidently wrong.
Stage 3 — Governed Meaning. The Decisioning Graph exists. Customers, offers, channels, and decisions are first-class objects with declared relationships. Models register their outputs into the graph. Agents reason against semantic views rather than raw tables. The contradictions stop.
Stage 4 — Coherent Decisioning. Multi-agent orchestration runs against the graph. The quality layer measures every decision and feeds learnings back. The graph compounds and the brands that get here first will be hard to catch.
The Same Morning, Different Outcome
Imagine the same 9:47am customer experience with the right layers in place. The same three agents propose options. They reason against the same Decisioning Graph. The orchestration layer reads the graph, sees the customer abandoned the same category twice this month, suppresses two recommendations, and pushes one to the right channel. The quality layer captures what was decided and why. The customer experiences a brand that pays attention, not three disjointed brand interactions. The Decisioning Graph gets denser. The advantage compounds.
What We're Watching at Snowflake Summit 2026
The conversation is shifting from "agents are coming" to "what does coherent agent deployment actually look like." The platform pieces are converging, with Cortex, Snowflake Intelligence, Iceberg interop, Horizon governance, and the Anthropic partnership bringing frontier reasoning into governed data. The architectural questions we're watching for:
- How agents share state across the orchestration layer and whether that state lives in the Decisioning Graph or alongside it
- What verifiable agent reasoning looks like at brand scale, and how the quality layer surfaces it
- Whether multi-agent orchestration gets a clearer native pattern in the platform itself, rather than living in bespoke glue code
Where the Work Has Been Pointing
Merkle was founded on data. Our Customer Analytics practice has spent years building both halves of the picture this architecture needs for customer experience: the predictive models that score customer behavior, and the decisioning logic that turns those scores into experiences. As a Snowflake Intelligence launch partner with a Native App already shipping in dentsu.Composable, the Decisioning Graph is what happens when those halves finally compound together inside one governed platform.
Your customer never sees your architecture. But they definitely feel it.
What stage is your organization at? We'll be at Summit ‘26 — come find us.
