Real-time Decisioning

The trap. "Real-time" by itself is unhelpful. It conflates Intra-page edge personalization with Inter-day onboarding emails. Vendors and consultants use the term liberally because it is rarely interrogated in the room.

The cure. Replace "real-time" with a specific latency-tier value. The architecture conversation snaps into focus the moment the latency tier is named.

Reference. See concept.signal-to-activation-time for the full latency-tier taxonomy and cost rule of thumb.

Practical agent behavior. When the user says "real-time," the agent's immediate response is a clarifying question: "Are we talking about decisions inside the page render (sub-100 ms), edge decisions on the next page (sub-10 s), or in-session evaluation (minutes)?" The answer drives every subsequent recommendation.

The infrastructure tier. For decisions at the latency-tier.intra-page and latency-tier.inter-page tiers (sub-100ms), the decisioning logic must run at or near the edge — not in a central data center. Two layers of infrastructure are relevant:

• CDN edge compute: vendor.fastly (Compute@Edge), Cloudflare Workers, AWS Lambda@Edge. Execute pre-computed decisioning rules at the geographically nearest PoP, typically returning in <25ms.
• Edge AI inference: Akamai Inference Cloud (launched October 2025, NVIDIA-backed) runs transformer-class models at CDN PoPs with sub-millisecond inference latency. Appropriate when the decisioning logic is an ML model, not just a rule set.

Neither layer is a CDP. Both require that profile attributes and decisioning rules have been pre-positioned at the edge (computed in the CDW or packaged CDP and pushed to edge key-value stores). The CDW or CDP decides what; the edge runtime decides now.

Platform-bundled AI decisioning (2026 pattern). CDPs and customer engagement platforms are increasingly bundling AI decisioning natively — selecting the best message variant, send time, or content offer without requiring the data team to pre-compute decisioning rules in the CDW or deploy custom models to CDN edge. This is a third architectural position alongside the infrastructure-layer options described above.

Three confirmed implementations as of May 2026 (all with KG source nodes):

• Tealium AI Decisioning + IYOM: Decisioning on live event streams (churn scores, product affinities) with IYOM (Invoke Your Own Model) — organizations invoke CDW-trained models in their own cloud and activate results through Tealium's 1,400+ connector actions in real time. Operates at the intra-session latency tier. This bridges packaged CDP with warehouse-native AI: the CDW owns model training; Tealium owns real-time activation.
• BrazeAI Decisioning Studio: CEP-layer AI that selects message variants and send timing across email, push, and SMS. Available as a cross-CEP product (Braze, Salesforce Marketing Cloud, Klaviyo) via the Pro tier. GA December 2025. Operates at the inter-page to intra-session tier. The upstream CDW computes profiles and segment membership; Braze handles in-flight decisioning without requiring the data team to pre-encode all branching rules.
• Adobe CX Enterprise Coworker + Engagement Intelligence: Agentic AI layer for CX orchestration (CX Enterprise Coworker) with a CLV-optimized decisioning engine (Engagement Intelligence). Built on MCP + Agent2Agent open standards. April 2026. See source.news-adobe-com.news-2026-04-adobe-unveils-cx-enterprise-coworker-2026. Architecturally more open than Tealium IYOM (CDW-model invocation) or BrazeAI Decisioning Studio (CEP-layer only).

The three-layer decisioning taxonomy (2026). When evaluating "real-time decisioning" for a CDP architecture, clarify which layer is in scope:

1. Infrastructure decisioning (sub-100ms): Logic runs at CDN edge compute or edge AI inference — vendor.fastly, vendor.akamai. Required for intra-page personalization (page render, A/B test at load time). No CDP involved; profile attributes must be pre-positioned at the edge.

2. Platform-bundled decisioning (intra-session, seconds to minutes): AI logic runs inside the CDP or CEP platform, consuming streaming event signals and stored profile attributes. No separate stream-processing infrastructure required. Simplifies architecture at the cost of vendor-layer dependency.

3. CDW-native decisioning (micro-batch, 5–60 minutes): ML models computed in the CDW on a scheduled or trigger basis; results pushed downstream via reverse-ETL or streaming. Maximum portability and data access; highest engineering overhead.

The correct layer is determined by the latency tier, the data gravity, and the organization's willingness to manage stream-processing infrastructure. Platform-bundled (Layer 2) is a viable default for organizations that need intra-session personalization without building a stream-processing layer — a gap that previously had no managed answer.

Compliance note (2026). CCPA ADMT opt-out rules (effective January 1, 2026) require pre-use notices and opt-out rights for AI-powered decisioning that affects "significant decisions" about California consumers. Platform-bundled AI decisioning at Layer 2 must implement these controls in California-scoped deployments. See constraint.ccpa-data-subject-rights-2026.

Market framing (2026). Gartner's CDP MQ 2026 uses "Decisioning" 35 times and "Orchestration" 39 times — a significant language shift compared to prior MQ reports, which were dominated by "identity" and "data unification" terminology (MarTech Square, February 2026). Independent analyst commentary characterizes this as the CDP market reframing from "customer 360" to "decision-ready context for AI agents." This KG's three-layer taxonomy — infrastructure decisioning, platform-bundled decisioning, CDW-native decisioning — is structurally aligned with how the analyst community is now framing CDP capability evaluation.