As senior vice president of product at Optimizely, Kevin Li has spent the past year building the case that the constraint on AI-generated content isn’t the model — it’s whether the organization has codified what it wants to say.

In this conversation with MarTech, Li explains why general-purpose AI tools stall inside a marketing organization, where enterprise content time actually goes, why review has to move from outputs to inputs, and what a buyer should demand before accepting that a CMS is agentic. Along the way, he describes an internal campaign that produced 800 individualized landing pages from a 17-page brief one marketer wrote in three days.

The interview has been edited for length and clarity.

Q: Vibe coding started as a way to build software quickly with a general-purpose model. Where does it break down inside an enterprise marketing organization?

If you think about what vibe coding is, it’s an articulation of one individual’s opinion about what good software looks like. So it’s good for prototyping, it’s good for sharing designs, it’s good for brainstorming, it’s good for storytelling.

But when it comes to a brand actually showing up for the customer, that needs to be an approved thing, where, as an organization, you should actually agree on what it is that you want to say. Organizations have a brand identity, a set of values that they stand for, and it shouldn’t be up to the individual to decide that on the fly. 

The key difference is the codification and the standardization of what the message is. 

Q: So, if you want to take advantage of AI to create code or create content, what does the agent need to know about the brand? Where do organizational mandates to use AI go wrong?

We have a phrase: “Process before magic.” People think agents do magic, but actually the process is the important part. 

A human designer would not make a new graphic for a brand without brand guidelines. They’re not opening up Photoshop being like, “I feel like using the color yellow today.” You start from your brand’s color palette. When you build a deck, you don’t just open up PowerPoint and go, “new slide.”

The agent needs whatever the human needs. I think people weirdly have this expectation that, because it’s an agent, it’s somehow magical and knows the things. But you can’t skip steps. It’s the exact same thing you would do with training a human. It’s like baking a cake. You can’t do that without ingredients just because a robot is doing it. 

Another thing that organizations don’t spend enough time thinking about is whether agents are now powerful enough to do autonomous work. And what does adoption actually mean? Does adoption mean every human employee using agents as an assistant, or does adoption mean there’s a whole lot of mundane work that’s super boring, no one’s doing it, and now we can have an autonomous agent do it?

Q: What important work is currently going undone?

There are many value-added things that are important, but the nature of human employees is that you’re limited by the working hours of humans. There could be, let’s say, 100 units of important work, but if you only have enough people to take care of 20 units, then only the first 20 units get done. The other 80 units are still important, but they’re not getting done. 

Here are some examples. Every brand should check every single piece of content and refresh it to make sure it has the right SEO keywords. No brand actually does it. Every brand should make sure and scan every single page for accessibility. No brand ever does that. Everybody should take down content that’s actually outdated. No brand does that. 

Another example: How many images in the digital asset management repository are off brand. How many of them don’t have tags? How many of them don’t have the right attribution? 

So is it valuable to do these things? Would the company be better if all of this was done? Yes. Do humans do it today? No. Because they have something more important to do.

Q: You’ve described enterprise content production as weeks of handoff. Where does the time actually go, and what does fixing it look like?

Today, where it goes is the slippage in between handoffs. It’s almost like a manufacturing process. If a process takes 10 steps and, on the first handoff, the second person isn’t ready to do the next step immediately, what happens to that piece of work? It waits. 

The weeks go by when every single handoff sits waiting for something specific. 

My colleague, Olivia, sends me stuff to review, then a day later she has to chase me. I apologize for that. The actual time it would take to do something is maybe two hours, but the calendar time it takes to do that thing is two weeks.

The key thing in speeding things up is to look at what parts of the process agents can do. Because an agent has infinite capacity to immediately address something. 

If there’s a 30-minute wait for humans to do something in a 10-step process, let’s say you remove five of those steps and an agent takes over, there’s zero minutes wait. You’ve just saved two and a half hours. 

Ask yourself: does this require an actual human to have eyes on it? Here we have a concept of virtual teammates, and I’m starting to build one that essentially acts as a second brain. I guarantee you Olivia would love it if she wrote something and didn’t need to wait for my availability for approval. She could ask a virtual Kevin for its opinions and get a first draft without being blocked by my calendar. Because an agent is infinitely available at any time.

Q: So, brands that adopt hyper-personalization need to be comfortable with putting content out there without human review – because it’s impossible to review thousands of personalized iterations. If you can’t have a human review every output, what replaces review?

The brand guidelines are baked in as input and brands need to review the input, although for decades brands have reviewed the output. 

Brands need to move towards reviewing inputs and not outputs because the outputs are constructed on the fly. As in, it is not possible to review all of the outputs. So review is still important, but where review happens is different.

To provide the agent with the right kind of input going in, you need to start by documenting the right process. I’ll give you an internal example. When we wanted to create account-specific landing pages for 800 potential clients, we started by asking Laura, an account-based marketer at Optimizely. We asked her “if you could have infinite time to make this the best personalized page, what would that look like?” She took three days and wrote 17 pages about how she would build it. 

We took that and put it in the agent. So the agent isn’t vibing what it thinks a good page is. The agent is fed with instructions for what good looks like. And we’ve built the production chain to produce those 800 pages.

Q: That process has to live in something. What should a buyer check to distinguish a real agent platform from a CMS with AI features bolted on?

I think buyers are getting smarter about this to be able to tell the difference. The more technical term for a standalone agent platform is a harness.

The harness is what controls everything. When evaluating them you want to know if you can build your own agents. Does it have evaluations? Does it have guardrails? Does it come with different LLMs? Does it do routing to different LLMs? It’s a product category in and of itself, and is very different from what a CMS is.

With the vendors that are pretending to have agentic platforms but are really just offering souped-up CMSs, you mostly hear about features of a CMS that they’ve turned agentic. That’s very different from having an agent platform.

The reason why we want to build an agent orchestration platform for marketing is because the world of marketing as a domain is much bigger than CMS as a piece of technology. We see the opportunity to influence marketing and not just the website.

Q: And why does marketing need its own harness, rather than just using Claude or ChatGPT?

We describe our platform as a harness that’s specifically for marketing work. I’ll give you a couple of good examples. Claude Code is a harness. ChatGPT is a harness. OpenClaw is a harness. These are general purpose harnesses, or, really, they’re general purpose with a slant of being good for software engineering because that’s where they got started. 

If you’re a developer, that’s a harness for you. But if you’re a marketer, our harness is actually the best for you to do your work. I’ll give you some examples of what ours is better at. Ours is better at brand packs; ours is better at PowerPoint skills, in particular. Ours is better at on-the-fly image creation with greater control. 

If you want ChatGPT to do a new image, you need to ask it to regenerate. But a lot of times, that’s not what brands need. They just want to fix one little thing, or they want to say, “here’s a base template, and let me make modifications.” 

So thinking about the workflows, it’s sort of like: what’s the difference between Microsoft Word and a CMS? They both store text. But they are purpose-built to be different. One is a word processor; one is an enterprise piece of software. Marketing workflows are marketing-specific. For example, look at the way we do embedded review. If you ever try to have two people using a general-purpose harness review the same cloud artifact, it’s a pain in the butt. You copy and paste a weird link that’s shared via the Internet. There’s no collaboration. So how do multiple people approve something? 

Harvey is a harness for legal, built around data-safe rooms, so that two parties negotiating an M&A deal each see only what they should see. We don’t need that; we want everything shared for collaboration. Marketing isn’t unique in this. Every domain needs its own harness.