By 2023, 64% of marketing leaders had implemented or were implementing AI marketing automation, according to benchmarks compiled by WorldMetrics and Gitnux. A year later, AI integration into content workflows jumped from 43% to 69%. The adoption curve isn't the question anymore. The question is what happens when the person building your AI workflow has no engineering training, no deployment checklist, and credentials to your CRM.
That's the tension BlueRock CMO David Greenberg laid out in a recent DemandGenReport.com Q&A. Marketers have become "citizen developers," constructing agents and workflows that connect systems, make decisions, and take actions across production environments. Most marketing leaders haven't updated their operating model to account for it.
The Workflow That Looks Like a Shortcut Until It Isn't
Greenberg's example is specific. A demand gen manager builds an AI workflow for event follow-up: ingest the attendee list, research each company, cross-reference CRM history and engagement data, score prospects, draft personalized outreach, route to the right rep. What used to take days now runs in minutes.
"Marketers no longer have to accept the limitations of the software they buy," Greenberg told DemandGenReport.com. "They can increasingly build the applications, agents and workflows they need themselves."
The failure mode is quieter. That workflow holds credentials, touches customer data, and has permission to act across multiple systems. Nobody filed a security review. Nobody documented the triggers or thresholds. The experiment became operational software without anyone treating it that way. One unexpected change can hit data security, brand consistency, and campaign performance simultaneously, because AI systems that make decisions don't silo their failures the way traditional martech does.
Governance as Operating Model, Not Approval Queue
Greenberg draws a line that matters for anyone running marketing ops: "The mistake is treating governance as an approval process rather than an operating model."
If every new workflow requires a security ticket and three meetings, builders either stop building or route around the process. Neither outcome produces pipeline. The alternative: define boundaries around what systems and data AI can access, what actions it can take, then give builders freedom inside those boundaries. Low-risk experimentation moves fast. Higher-impact actions (customer-facing comms, deal routing on high-value accounts) get additional controls.
The cautionary view from practitioners like Stephan Bisser is worth noting: citizen development works for simple, self-serve use cases, but once you're dealing with integrations, complex logic, or enterprise governance, you need professional developers or at minimum guided development with IT involvement. Both views are right, depending on the workflow. The operating model should account for that spectrum.
The Three Moves Greenberg Recommends (and the One Most Teams Skip)
Greenberg's 30-day playbook for moving from casual prompting to repeatable AI workflows:
- Identify a handful of real workflows worth building. Look for repetitive work, manual handoffs, or processes where the team already understands the problem deeply. Don't start with the most complex workflow. Start with the one where the owner can articulate exactly what "good" looks like.
- Give teams a secure environment designed for agentic building. Boundaries, visibility, controls. This is the step most organizations skip, Greenberg says.
- Teach people how to build well. Hands-on learning around AI development practices, not just better prompting.
The skipped step is the second one. Teams jump from identifying the workflow to shipping it, with nothing in between to enforce permissions, log actions, or flag behavioral drift. Agents adapt as they execute, so what worked Tuesday may behave differently Thursday. Without runtime visibility, the team won't know until the outcome shows up in pipeline metrics or a customer complaint.
Where This Connects to Pipeline
When asked about the stat that 36% of employees don't understand why they're expected to use AI, Greenberg called it the bigger red flag: "Adoption without a clear understanding of the problem you are trying to solve creates a lot of activity without necessarily creating value."
For growth leaders measuring AI investment against pipeline, this is the diagnostic. Are your AI workflows producing more activity (more emails sent, more content generated, more leads scored) or better outcomes at the stages that matter? Reply-to-meeting rate. Meeting-to-opportunity conversion. Win rate on AI-sourced pipeline versus manually sourced. If you can't answer that, you're measuring adoption, not impact.
The hypothesis worth testing: if you pilot one bounded AI workflow (say, event follow-up enrichment and routing) with clear permissions and human review on outbound, then qualified pipeline from that event cohort should improve versus a control group that gets the old manual process. Measure at the opportunity stage, not at the MQL stage. Set a stop-loss: if outbound reply rates drop below your baseline by more than 15% after two weeks, pause and diagnose.
The Constraint That Changed
Greenberg's core argument: AI removed a long-standing constraint. Marketers who understood the customer and the process couldn't build the systems they needed. Now they can. But the new constraint is governance, ownership, and measurement. The teams that treat their AI workflows like the production software they actually are will compound their advantage. The ones that don't will discover the failure mode when something breaks across three systems at once and nobody documented how it was built.
