Optimizely's Agentic AI Pilot Found the Bottleneck Wasn't the Agents

Optimizely’s internal deployment of agentic AI in its own marketing operations turned up a finding that cuts against the standard pilot-to-scale narrative: workflow design, not agent capability, was the binding constraint. Forrester Principal Analyst Joe Cicman, drawing on Optimizely’s “customer zero” rollout, found that preparation, coordination, and handoff work consumed most of the organizational capacity before a campaign ever reached a customer — so adding capable agents on top of that didn’t translate into more throughput. Cicman’s framing: “An agent can execute a task. A workflow determines whether that task contributes to a functioning system.” Optimizely’s response was to systematize marketing work — making it explicit, structured, and measurable — rather than layering AI onto ad hoc processes. The piece stresses that automation reveals bottlenecks rather than eliminating them, meaning the constraint shifts rather than disappears, and argues technology and marketing leaders have to co-design operating models together rather than treating agentic rollout as a purely technical project.

The finding matches a governance failure mode Salesforce documented in its own internal systems around the same time. That piece describes an internal Slackbot that confidently returned two different, both-wrong figures — 50,000 and 75,000 — for a marketing-qualified-leads query, when the certified dashboard showed 100,000. The root cause was the agent querying a raw data object directly instead of the organization’s certified metric definition; Salesforce’s fix pairs structured “AI Skills” with certified datasets so agents have a single source of truth to query against, rather than reasoning over whatever raw field happens to be nearest.

Read together, both cases make the same point from different failure modes: an agent that’s technically capable can still produce organizational failure — wasted throughput in Optimizely’s case, confidently wrong numbers in Salesforce’s — when the workflow and data-definition layer underneath it isn’t designed first. For marketing-ops and RevOps leaders piloting agentic AI, both pieces argue the gap between a working demo and real operational adoption closes through governance and workflow redesign, not a more capable model.