Optimizely’s Kevin Li says agentic marketing depends on documented processes

src-a7608277

Marketing teams seeking to use AI agents for brand-sensitive work need to define their processes and standards before handing over production, argues Kevin Li, senior vice president of product at Optimizely. In a sponsored interview, Li said the limiting factor for AI-generated marketing content is often not the model itself but whether an organization has clearly specified what it wants to communicate and how work should be done.

His argument challenges the idea that general-purpose AI can reliably produce enterprise-ready marketing simply from a short prompt. Li said rapid, informal “vibe coding” can help with prototypes, brainstorming and design sharing, but brand communication requires organizational agreement on identity, values and messaging. Without that foundation, an individual user is effectively left to decide what the brand should say.

Put the process ahead of the agent

Li summarized the approach as “Process before magic.” He compared an AI agent to a human designer: both need guidance, such as brand standards and approved inputs, to produce work that meets expectations. Organizations should not assume an agent will infer requirements that have never been documented.

That requirement has operational consequences. Li said marketers should distinguish between using agents as assistants for employees and assigning them autonomous work. He pointed to recurring tasks that may go undone because teams have limited capacity: checking all content for relevant SEO terms, reviewing pages for accessibility, removing outdated material, and auditing digital asset management libraries for off-brand images, missing tags or attribution issues.

Those examples frame automation as a way to address work that has value but may not make it onto a human team’s immediate priority list. Li did not suggest that every task should be automated without oversight; rather, he argued that teams need to identify which work requires human judgment and which can be handled according to documented instructions.

Reduce delays between handoffs

Li also described a common source of delay in enterprise content production: work waiting between steps. A task that requires only a few hours of active effort can stretch across days or weeks when reviewers and collaborators are not available at the right time. In his account, the calendar time is often dominated by handoffs rather than the work itself.

Agents could take on selected steps and respond without waiting for a colleague’s schedule, he said. Li described the idea of a “virtual teammate” that could provide an initial review or feedback while a human reviewer is unavailable. That could reduce bottlenecks, though it depends on supplying the agent with the right context and deciding where human approval remains necessary.

For highly personalized campaigns, Li said the review model may need to shift from checking every output to checking the instructions and standards that guide production. He cited an Optimizely internal campaign aimed at creating individualized landing pages for 800 prospective clients. Account-based marketer Laura spent three days writing a 17-page brief describing what a strong personalized page should include. The team then used that documented approach as input for an agent and built a production chain for the pages.

The example illustrates Li’s emphasis on codifying expertise before scaling output. It is an account from Optimizely, not independent evidence that every organization can achieve the same volume or results. The underlying operational question is how a marketing team can translate its standards into usable instructions, controls and review practices.

What buyers should examine

Li advised buyers evaluating claims about agentic platforms to ask whether they can build their own agents, whether the product includes evaluation and guardrail capabilities, and whether it supports multiple large language models and routing among them. He called the underlying control layer a “harness” and described it as distinct from a content management system with AI features added.

In Li’s view, a CMS is one part of a broader marketing workflow, while an agent platform should support work across the marketing function. He argued that purpose-built tools can address needs such as brand packs, presentation workflows, controlled image changes and collaborative review. General-purpose systems, he said, may not provide the same domain-specific workflow support.

These are Optimizely’s product and category perspectives, presented by one of its senior product executives. For marketing leaders, the practical takeaway is to assess the operating process alongside the technology: what knowledge is documented, how agents are evaluated, which safeguards apply, and where people retain decision-making authority. Without those elements, adding an agent may accelerate production without ensuring that its work reflects the organization’s brand or requirements.