According to McKinsey, 65% of organizations now regularly use GenAI in at least one business function—yet only 11% have deployed it at scale. The gap between experimentation and business value keeps widening, and the solution, argues Onil Gunawardana, isn't better models.

"Enterprise AI delivers business impact when the AI Product Manager drives the closed loop from idea to impact, not just the product solution," says Gunawardana, whose product leadership at leading Silicon Valley companies has shaped the enterprise AI landscape.

Most organizations are stuck in what Gunawardana calls "pilot purgatory"—where AI outputs never translate into operational action. The problem isn't the model. It's the missing loop.

Impact requires a closed loop: Input flows to Decision, Decision triggers Action, Action produces Outcome, and Outcome feeds back to improve the system. Without this architecture, AI remains a sophisticated demo. With it, AI becomes infrastructure.

Start with Outcomes, Not Technology

Gunawardana recommends starting with a single workflow that already has a scoreboard—contact center operations, sales pipeline management, or IT service desk. The key is selecting a domain where success is already measured, so AI's contribution can be isolated and quantified.

"If you can't point to a metric that will move, you're not ready for AI—you're ready for a demo," Gunawardana notes.

Before building anything, write the one-line business case: current baseline, target improvement, and the economic unit that matters. A contact center might frame it as: "Reduce average handle time from 8 minutes to 6 minutes, saving $2.40 per call." This clarity forces alignment between technical teams and business stakeholders from day one.

Define constraints early. Acceptable latency, accuracy tolerance, and compliance boundaries should be established before prototyping begins. These aren't obstacles—they're design requirements that prevent costly pivots later.

Prototype for Proof, Not Applause

With the business case established, prototyping can begin—but the goal isn't to impress stakeholders. It's to demonstrate task success on real cases.

Gunawardana advocates building a small "golden set" of 20 to 50 representative examples with a simple evaluation rubric. This creates an objective standard for measuring whether the AI actually works.

The velocity accelerator, according to Gunawardana, is writing the product requirements document (PRD) and QA test plan together in Claude Code, Cursor, or your favorite coding environment—and keeping both in code repositories as markdown so AI assistants can access and reason over specs directly. "Write the PRD and the test plan together," he advises. "Let AI iterate against real data. That's where velocity comes from." He urges AI Product Managers to use the latest AI tools aggressively—to be power users of what they're building for others.

This approach creates its own closed loop: build, test, learn, improve. The test plan defines success criteria upfront. AI systems iterate against the test plan using production data. Feedback cycles that once took weeks compress to minutes.

Every prototype should include safe behavior defaults: ask clarifying questions when input is ambiguous, abstain or escalate when confidence is low, and cite sources when relevant. These behaviors build trust and reduce friction in enterprise adoption.

Enterprise Readiness as Competitive Advantage

Once a prototype proves task success, enterprise readiness becomes the critical path. These gates aren't bureaucratic hurdles—they're the adoption path.

Gunawardana identifies three categories that determine whether AI ships or stalls in security review:

• Identity and access: SSO integration, SCIM provisioning, and role-based access control
• Data boundaries: Workspace and tenant separation that enterprises require for compliance
• Auditability: Comprehensive logging of who did what, when, and why

"These aren't platform chores—they're the adoption path," Gunawardana says. "Skip them and you'll be stuck in security review while competitors ship."

Teams that build readiness into their roadmap from the start deploy faster than those who bolt it on after proof-of-concept.

Controlled Rollout Patterns That Limit Risk

With enterprise readiness addressed, the deployment strategy determines how quickly value reaches the organization. Gunawardana outlines three patterns that shorten time-to-value while limiting risk.

Shadow mode runs AI in parallel with existing workflows, measuring performance without taking action. This builds confidence through data before any operational change occurs.

Human-in-the-loop lets AI draft outputs while humans approve final actions. This captures efficiency gains while maintaining oversight during early adoption.

Role-based canary starts with a single team, expands based on success metrics, and limits blast radius if problems emerge.

Gunawardana advises starting in shadow mode and measuring before acting. Rollback triggers should be business metrics, not gut feelings—severe error rates, escalation spikes, or customer satisfaction drops. When triggers are objective, decisions to expand or retreat become straightforward.

Measuring What Matters

Gunawardana recommends a minimum set of metrics for any production AI deployment:

• Task success rate: Whether the AI accomplishes its intended goal
• Escalation rate: How often humans must rescue failed AI attempts
• Cycle-time reduction: Direct tie to the baseline established at the outset
• Cost per successful task: Predictability for capacity planning

"If you can't connect your AI investment to a business metric within six months, you've likely misframed the problem," Gunawardana notes.

Cadence matters as much as metrics. Weekly reviews with the product owner, operations lead, and security representative create accountability and surface issues before they compound.

Morgan Stanley's Eval-Led Path to Measurable Impact

Morgan Stanley and OpenAI have publicly described a robust evaluation framework that ensures reliability and consistency before scaling advisor-facing GenAI. Their approach exemplifies the closed-loop pattern Gunawardana advocates.

The "AI @ Morgan Stanley Debrief" tool, deployed with client consent, generates meeting notes and action items, drafts follow-up emails for advisor review, and saves approved notes directly into Salesforce. CEO Ted Pick has stated the tool could save advisors 10 to 15 hours per week—and 98% of advisor teams have adopted it.

The pattern is clear: conversation flows to summary, summary receives human review, and reviewed action updates the system of record.

Gunawardana points to this as an exemplary implementation. "That's the blueprint. The AI doesn't just answer—it acts, and the action is tracked."

The AI PM as Time-to-Impact Owner

The framework connects: pick a workflow with an existing scoreboard, prove task success with a testable prototype, build production readiness into the product, deploy with controlled rollout patterns, and measure outcomes against a baseline.

This progression mirrors the philosophy behind Gunawardana's 5Ps of Product—a framework for taking products from concept to scaled revenue. The same principle applies to AI: automate the repeatable, focus humans on judgment, and build the platform that lets both work together.

As AI capabilities expand into UCaaS, CCaaS, and operational workflows, competitive advantage will belong to teams that operationalize feedback loops and governance alongside raw capability. The technology is commoditized. The differentiation is execution.

"The AI Product Manager's job isn't to ship models—it's to ship outcomes," Gunawardana says. "Own the loop, and you own the impact."