The four layers of enterprise marketing agents.

Why a stack, not a feature list

The marketing AI category has spent eighteen months shipping "agent" features without agreeing on what an agent is. The result is a market of tools that all claim agents and rarely deliver them — a chatbot here, a subject-line generator there, a flow builder with a sentence input on top. Buyers can't tell what they're evaluating.

A stack is a forcing function. It says: an enterprise agent is the thing that happens when these four layers exist together. Take any layer out and the work doesn't hold. Look at a vendor's deck and ask which of the four they ship; the gaps are usually obvious.

Foundation is the layer enterprise procurement reads first. It's also the one most vendors gloss past. Multi-model isn't marketing copy — it means a buyer can pick OpenAI, Anthropic, or Google for a workspace and switch without rewriting the prompts. Multi-cloud isn't a deployment toggle — it's whether your agents can sit inside your AWS, GCP, or Azure environment with the data plane you've already approved.

Integration is where Foundation gets practical. Agents have to read from your CDP, write to your warehouse, fire webhooks your internal tools listen for, and authenticate against your SSO provider. The layer is invisible when it works and blocks every layer above it when it doesn't.

What we ship at this layer: encryption at rest and in transit, SSO via SAML 2.0 with SCIM, scoped bearer tokens, signed webhooks with retries and replay, OpenAPI 3.1 spec, multi-region data residency, and BYO keys for sub-providers. Pre-signed DPA for any customer who asks. SOC 2 Type II audit in progress; report expected Q2 2026.

A foundation model alone is a generic writer with internet-scale opinions. An agent grounded in your context is something else: it knows your brand voice, your product catalog, your audience definitions, your suppression rules, your opt-in regimes. The grounding layer is what makes the difference between "sounds like AI" and "sounds like our brand."

In our model, Context is delivered through Reusable Skills — modular capabilities composed once and inherited by every agent that needs them. A discount-code generator built for one cart-recovery agent becomes available to every loyalty, win-back, and upgrade agent in the workspace. A brand-voice reviewer tuned during onboarding gets enforced at generation time across every channel. An inventory-aware product picker queries your catalog directly instead of guessing.

Skills are versioned. They're testable. They roll back. The cart-recovery agent and the post-purchase agent share the same suppression-list checker by design — fix it in one place; both update.

Orchestration is the layer where agents stop talking and start doing. The value of marketing AI was never in generating fluent text — that was solved two years ago. The value is in the per-recipient decisions that turn fluent text into a working program: which channel reaches this customer, which sequence belongs to this cohort, which send time matches this audience's rhythm.

Six channels run through the same orchestrator: email, SMS, WhatsApp, push, RCS, and in-app. The orchestrator decides per recipient at send time. Cart-recovery might fire SMS for the recently engaged and email for the cooling. The win-back might wait three days for users with high Composite Engagement Score and three hours for users about to churn. The launch announcement might use WhatsApp in LATAM and email in North America — same agent, different decisions, no parallel campaigns to maintain.

The orchestrator's job is decisions, not authorship. The Skills that ground the work, the Foundation that powers it, and the Observability that watches it are all in service of decisions made well, per recipient, every minute.

Drift compounds. Models drift. Audiences drift. Brand voice drifts. The launch you shipped in February sounds different by August because the data the agent is grounded in shifted three percent every week without anyone noticing. Observability is the layer that catches drift before it becomes deliverability collapse.

Three tools live here. Evaluations test the agent against real past sends before it ever touches a live audience — a synthetic test set generated from your historical data, scored automatically, gated in CI. Live monitoring watches every send for bounce-rate spikes, complaint-rate anomalies, and engagement deltas with context — not just "something is up" but "bounce rate up 3%, concentrated in Gmail recipients added in the last 30 days, likely list quality." Composite Engagement Score tracks per-contact engagement across every channel and signal — a metric that survives Apple Mail Privacy because it doesn't depend on opens.

The audit log captures every action: model used, prompt, output, decision rationale, recipient, send time. Compliance reviews stop being archaeology; they're a query.

The stack as a procurement test

The fastest way to evaluate a marketing AI vendor is to walk down the stack and ask what they ship at each layer. Foundation: do they have multi-model? Multi- cloud? SSO? Context: do they support reusable Skills, or is every agent built from scratch? Orchestration: how many channels, and are decisions made per recipient at send time, or per campaign at design time? Observability: evaluations against real data, or just dashboards?

Most vendors have one or two strong layers. The best have four. The worst sell you a fluent demo and leave the integration, the grounding, and the watch as your problem to solve.

We built PolarGX with all four layers from the start because Polaris Innovation spent the last decade learning, the hard way, that any one missing layer makes the rest of the work fail quietly. Foundation alone is infrastructure looking for a customer. Context without orchestration is a search index. Orchestration without observability is a hope. Observability without the layers underneath is a dashboard with nothing to watch.

Stack-shaped, all four layers, shipped together. That's what Enterprise Agents look like.