Salesforce wired its AI agent directly into the website that generates pipeline. Not a sandbox. Not a proof-of-concept in a low-traffic corner. The actual lead-generation engine, the one producing 200 signals per visitor that feed attribution models and determine the next experience in the buyer journey. Forrester principal analyst Joe Cicman calls this the "no plan B" move: removing the organizational fallback so the team has no choice but to make the agent work.
For marketing ops, the interesting part isn't the bravery. It's what broke, what had to be rebuilt, and what the rest of us should instrument before we try the same thing.
Outcome Measurement Isn't Activity Logging
Salesforce discovered that knowing what agents did (which responses they gave, which topics they handled) wasn't the same as knowing whether those interactions produced good outcomes. Traditional analytics told them the agent fired. They couldn't tell whether the conversation moved a visitor toward qualified pipeline or just generated a polite dead end.
The fix: LLM-based evaluation techniques to score agent behavior against outcome criteria. That's a different instrumentation layer than most teams have today. If you're running agents in any part of the buyer journey, the question is whether your observability stack measures downstream impact (pipeline created, meetings booked, customer effort reduced) or just upstream activity (conversations started, topics matched).
The trade-off if you skip this: dashboards full of agent activity and zero ability to tell whether the agent helped or hurt conversion. Directional attribution is fine here, but you need at least one outcome metric tied to a stage your GTM team already trusts.
Knowledge Ops Is the Hidden Workload
Salesforce had to create agent-specific knowledge assets, content stored in their CMS but never published to the public website. Content management started becoming knowledge management, in Cicman's framing.
This matters because agent-ready knowledge is different from customer-facing content. It includes internal context, edge-case handling instructions, qualification criteria the agent needs but a human visitor shouldn't see. If your content ops team isn't already building and governing these unpublished assets, you're going to hit reliability problems the moment you put an agent somewhere consequential. Freshness, provenance, access controls: someone has to own the review cadence, and that someone probably sits in marketing ops or RevOps.
Agents Reshape the Paths You Measure
Forrester's companion "Customer Zero" case study makes a point that should keep RevOps leads up at night: agent success depends on an organization's ability to observe, measure, and govern the changes agents introduce, because agents can reshape qualification paths, demand flows, attribution, and operating models.
The agent doesn't just sit inside your existing funnel. It can reroute the funnel. If your attribution model assumes a specific sequence of touchpoints (web visit, form fill, SDR outreach, meeting), and the agent creates a shortcut or a detour, your pipeline reporting drifts. You won't notice until the numbers stop reconciling, and by then you've been making budget decisions on bad data for weeks.
The diagnostic before you deploy: map the qualification path the agent will touch. Identify every downstream report that depends on that path. Decide how you'll detect drift (a weekly reconciliation between agent-influenced pipeline and your existing attribution baseline is a reasonable starting point).
Escalation Design Is a Conversion Lever
A Five9-commissioned survey reported by CMSWire found that 92% of consumers give an AI agent fewer than 10 minutes before escalating, and 46% cited repeated misunderstanding as a reason to seek a human. Thirty-five percent said they often or always have to repeat themselves after transfer.
That last number should drive your escalation design. The handoff from agent to human isn't an edge case; if context doesn't transfer cleanly, you're burning trust at the exact moment the buyer is already frustrated. Salesforce's guardrails documentation recommends layered controls (input validation, output validation, content filtering, action authorization) rather than relying on a single safeguard. A single prompt-level guardrail is not a safety net; it's a suggestion.
What This Means for Your Next Agent Deployment
The hypothesis worth testing: if we instrument outcome metrics (not just activity metrics) before deploying an agent into a revenue-impacting journey, we'll detect failure modes within the first two weeks instead of discovering attribution drift a quarter later.
Setup: define one primary outcome metric (qualified meetings influenced, pipeline created, or customer-effort score). Add one guardrail metric (escalation rate, context-transfer completion). Set a stop-loss: if the outcome metric drops more than 15% from baseline over two weeks, pull the agent back to a lower-stakes touchpoint and diagnose.
The uncomfortable truth in Cicman's "no plan B" framing is that organizational commitment works precisely because it makes the cost of failure visible. But visible failure only teaches you something if you've built the measurement to see it. Most teams haven't. The boats are already burning; the question is whether you brought a compass.
