In a 28,400-lead panel studied by Visionary Marketing, median first response time was 1 hour 42 minutes. Only 14.6% of leads got a response within five minutes, and 4.8% received no response at all. That's not a chatbot problem. That's a routing and assignment problem, and it predates any AI agent by a decade.
The temptation right now is to buy an "agentic" platform, watch a polished demo with clean sample data, and assume the routing mess will fix itself. It won't. According to LeanData's 2026 B2B State of Martech and Revenue Operations Report, 40% of respondents reported data-quality issues that prevented accurate routing or assignment. Same report: AI routing and assignment ranked among the least-adopted agent use cases. The gap between vendor pitch and operational reality is wide enough to lose pipeline in.
What a Pipeline Agent Actually Does
An AI marketing agent is a scoped software worker. It retrieves information, uses tools, completes an assignment, and records what happened. Specify its trigger, permitted systems, output format, exceptions, and the human who owns it. If you can't describe the agent's job in a single sentence with a measurable deliverable, you don't have an agent. You have a prototype.
Meagen Eisenberg, CMO at Samsara, described on the Clarity podcast using agents to investigate lead-routing and SLA questions that previously consumed staff time. She's using agents to surface operational problems (who got this lead, when, and why), not to replace the humans accountable for pipeline. That's the right order.
Keep deterministic steps deterministic. Territory lookup, duplicate detection, lead-to-account matching: these are rule-based operations. They need clean data, governed rules, and an exception queue for edge cases. Use a model to interpret inconsistent information (a company name spelled three ways, a contact with two accounts), then validate its output against written territory and ownership logic.
The Job Worth Starting With
Inbound form submission hits your CRM. The agent enriches the record from allowed systems, preserving the source, timestamp, and confidence score for every field it adds. It checks existing account ownership, applies written fit and territory rules, and prepares the handoff to a named rep inside the hour. Conflicting ownership enters an exception queue for a human. The assignment ends at routing. Pricing negotiation, discovery calls, outreach sequencing: all outside this agent's scope.
Measure three things: submission-to-assignment time, assignment corrections (how often did the agent route to the wrong rep?), and unresolved exceptions. Track the rep's first contact and qualified opportunities separately. A fast assignment doesn't compensate for a seller who never follows up.
Lead-to-account matching is the prerequisite most teams skip. Orphaned leads and broken matching create downstream assignment and SLA failures that agents will amplify, not fix. If your identity resolution is weak, agents will misassign faster.
Buy, Build, or Hire (the Decision Tree)
Buy a platform when the job is standard, integrations exist, and your operator can inspect the work. Demand a test on your records, not the vendor's clean sample data. Build internally when proprietary rules justify custom control and engineering can maintain integrations. Budget for credentials, monitoring, recovery, model updates, and the person clearing exceptions. Hire an agency (like Verto Digital) when domain expertise and implementation capacity are both missing. Scope the engagement around a working job, an evaluation set, and transferable configuration and logs. A governance-first blueprint, where agents operate inside deterministic routing and territory rules with full audit trails, is the deliverable.
The Evaluation That Replaces the Demo
Give every candidate (vendor, internal build, agency) the same historical records and permitted tools. Establish your current process's baseline first. Then test three scenarios:
- Routine work: clean inbound submission, clear territory match, known account.
- Ambiguous inputs: company name misspelled, contact associated with two accounts, territory overlap.
- Broken dependencies: disconnect an enrichment source. Does the agent retry with bounded limits, alert the operator, and maintain recoverable state? Or does it silently assign based on incomplete data?
Score routing accuracy against written criteria. Submit the same event twice and require no duplicate assignments. Report cost per accepted job, correction minutes, staff hours saved, and unresolved queue age. Export logs, configuration, and work products before signing anything long-term.
The hypothesis (make it falsifiable): if we deploy a scoped routing agent with governed territory rules and provenance-tracked enrichment, then submission-to-assignment time will drop below 30 minutes and assignment corrections will stay under 5%, because the bottleneck is manual lookup and CRM hygiene, not decision complexity.
Success = median assignment time under 30 minutes; guardrails = correction rate under 5%; stop-loss = if correction rate exceeds 10% in the first two weeks, revoke write access and revert to manual routing while you diagnose.
Permissions Earned, Not Granted
The right deployment order: read access first, then recommendations, then action with approval. Only after the operator trusts its judgment does the agent get write permissions. Name the actual person authorized to release each class of work. Agents don't independently send outreach, publish content, or authorize spend. More completed tasks don't justify broader access if the correction queue grows alongside them.
The 28,400-lead panel that opened this piece didn't have an agent problem. It had an accountability problem dressed up as a speed problem. An agent that routes a lead in four seconds to a rep who doesn't respond for two days hasn't fixed anything. Start with the job card. Define the trigger, the deliverable, the exception path, and the human whose name goes on the outcome. Everything else is a chatbot demo.
