Trevor Pope on Building an AI Harness, Swapping Models Without Rebuilding, and Governing AI Agents in B2B Commerce

What an AI Harness Is, and Why the Question Matters

Hein noted that B2B buyers walking a show floor are starting to tune out the phrase “we’re doing AI.” Pope’s answer was to look at how the organization connects its AI tools to its work. Marketing and commerce teams use a commerce platform, a content management system, and experimentation and personalization tools. Each of those needs to know who is signed in, what each person is allowed to do, and what every person and every agent has done, with the actions recorded and open to audit.

The same layer holds a single set of brand guidelines for AI to follow, so each employee does not write their own. It also controls how file repositories, customer relationship management (CRM) systems, and marketing automation tools connect to models such as Claude, ChatGPT, Gemini, or Microsoft Copilot.

The harness is what allows you to actually bring everything together.
— Trevor Pope, Optimizely

Underneath sits one model or several, connected to the applications so they can take action on the company’s behalf. Pope also pointed to the Model Context Protocol (MCP), which Anthropic introduced in November 2024 as an open standard for connecting AI assistants to the systems where data lives. Hein offered an analogy that stuck with both of them.

If your digital business is the cart and the LLM is the horse, the harness is the thing that you strap the horse into that you use to pull the cart.
— Jason Hein, B2B eCommerce Association

Keep the Horse Swappable

Hein raised a provocative point. Every system a company connects to a single model creates a dependency, and the longer that connection stays, the greater the chance the company has picked the wrong model. Model providers benefit when a business’s systems tie tightly to their platform, which makes flexibility more valuable.

Pope described how Optimizely handled that choice. The company treated AI as a product, not an add-on, and chose not to build its own large language models (LLMs), since specialized labs already do that. Its strength is the layer where commerce and marketing work gets done. Optimizely started its agent platform on Google’s Gemini, then added Claude as a second model when new capabilities and fixes mattered.

You don’t always want to use the same horse.
— Trevor Pope

Because of the component design, the company added the second model without changing anything for users, agents, or context. Hein pointed out why that matters in B2B, where integrations are the step that most often derails a replatforming project.

Imagine how our lives would be different in this ecosystem right now if we could hot swap our PIMs, our ERPs, our e-commerce platforms like a harness lets us hot swap LLMs.
— Jason Hein

Hein asked whether a harness replaces integration platform as a service (iPaaS) tools, which move and translate data between systems. Pope said the two complement each other. An iPaaS excels at moving large volumes of data quickly, while a harness works with people, sets up workflows, and takes action. B2BEA’s guide to the systems that power B2B commerce describes integration layers as the connective tissue of the whole ecosystem, and a harness adds an AI-specific layer on top.

Execution Is Where the Platform Earns Its Place

Hein framed the thinking behind Optimizely’s approach. AI models are getting better at making choices, and a platform vendor does not need to win that race. It can make execution as easy as possible, so a team acts quickly on a decision, and it can make it simple to swap in a better decision-maker when one appears.

Pope agreed and tied it to what he sees at manufacturers and distributors.

Their marketing department isn’t big enough.
— Trevor Pope

Expectations for personalized experiences and campaigns keep climbing, and the knowledge of how a company works has to live somewhere that people and agents can both reach. The difference he highlighted is between AI that reads something and gives a report, and AI that drafts the new page or product detail change and then asks whether the team wants to test it.

Pope has also tested model choice for quality. Optimizely released quality guidance that gives an agent examples of strong output, so a team can change the underlying model and see how quality shifts. The company is finding it can reach better quality at lower cost, and it lets users choose a model for each agent or let the system choose automatically in chat.

Examples From Distributors and Manufacturers

Pope shared several patterns from customers.

• A regional wire distributor replatformed its commerce site, then used agents for experimentation: suggesting changes, setting up variations, interpreting results, and helping with development. The goal was to run more and better tests.

• Another distributor uses agents to find products that slipped through its product information management (PIM) system, such as active items missing a category or data, and to analyze logs and build ad hoc reports on category completeness and top sellers.

• An electrical manufacturer and a high-speed networking manufacturer use agents for content creation and translation, so small teams can run more campaigns with better content for lead and quote generation.

• Account-based pages let a manufacturer create and retire landing pages for its top 100 or 500 customers automatically.

Governance: What CIOs Ask First

Hein asked the question a chief information officer would ask: with many models in the barn, how does a company keep control? Pope said CIOs want to know what actions were taken and by whom, whether a person or an agent. Single sign-on and multifactor authentication should cover agents as well, so permission changes apply right away.

More CIOs are concerned about model training than anything else.
— Trevor Pope

They want to know whether company data could ever resurface for another user of the same model, and how transparent a model provider would be about incidents. Pope suggested a department pick one harness, since accounting and marketing may not share one, and set clear rules on who can create agents, start them, or change brand context.

He has seen the posture change quickly. At the start of the year, many organizations ruled out outside AI tools because a security review could take six months. Now many are willing to go through a rigorous process, and Hein said the pace of change feels like the Wild West. Careful governance fits Gartner’s finding that over 40 percent of agentic AI projects will be canceled by the end of 2027, citing escalating costs, unclear business value, or inadequate risk controls.

Customer-Facing AI Is the Next Step

Hein noted that at a B2BEA event in Melbourne, the most common answer to how companies use AI was internal processes, with no customer contact. Pope offered two reasons for the caution. Letting customers use agents on a site opens usage and costs that are hard to predict. And agents may soon visit a manufacturer’s site in greater numbers than people, which calls for a separate experience. He sees an easy first step in giving signed-in customers more agent capability, such as asking about order history.

In his experience, B2B buyers adopt AI about twice as fast as consumers because the technology serves efficiency. That fits Forrester’s research on zero-click B2B buying, which found that 94 percent of business buyers use AI in their buying process. B2BEA’s conversation with 3M’s Jen Watters on AI and digital strategy in industrial B2B shows what an assistant for customers can look like. It also covers B2BEA’s conversation with Søren Handlos on bringing full eCommerce into procurement systems, which looks at how a seller’s experience connects to the systems buyers already use.

On agentic purchasing, where an AI agent buys for the customer, Pope was candid. Agent traffic converts at a low rate today, and many predict it will be huge within five years.

I don’t know how much I believe, but also we’ve seen a lot of change.
— Trevor Pope

Start With the Customer, Then Equip Your Team

Asked what a manufacturer or distributor should do tomorrow, Pope named two paths. The first is to stay close to the customer, especially the largest ones, who tend to innovate first.

What my customer needs and how I’m servicing them is my absolute North Star.
— Trevor Pope

The second is to give employees better tools, because tolerance for weak experiences keeps shrinking, for customers and for the people building them.

There’s so much less tolerance for a bad digital experience.
— Trevor Pope

That ties back to the harness. AI keeps changing faster than almost anything these teams have seen, and Pope said the safest preparation is flexibility, a view Gartner also supports: it predicts that 40 percent of enterprise applications will feature task-specific AI agents by 2026, up from less than 5 percent in 2025.

I more and more value flexibility.
— Trevor Pope

About the Guest

Trevor Pope, Senior Director of Solution Architecture, Optimizely

Trevor Pope is Senior Director of Solution Architecture at Optimizely, where he works with manufacturers and distributors that are increasing service levels and connecting with customers across their digital and physical capabilities. He has been with Optimizely for more than twelve years, focusing on the manufacturing and distribution vertical, and his work includes global enablement of solution architects who focus on distributors and manufacturers. He helps these organizations get value from commerce, experimentation, personalization, and content management applications. Trevor is passionate about helping them meet their business challenges, and he continually seeks to understand each situation and adapt to find the best solution. Connect with him on LinkedIn.

About the Host

Jason Hein is Vice President of B2BEA Academy and Global Director of Programming for B2B eCommerce World at the B2B eCommerce Association, where he builds and scales the association’s practitioner education offerings and leads the development of conference programming across B2B eCommerce World events in APAC, the Americas, and Europe. He also hosts the B2B eCommerce Show, where he talks with practitioners and technology leaders shaping how manufacturers and distributors sell online. With nearly 30 years of experience in B2B eCommerce, he also runs his own consulting practice, Acumental B2B, helping companies fix underperforming B2B sites.