Why the “middle miles” are where B2B’s real AI opportunity sits
This episode previewed an upcoming main stage session at the B2B Exchange at Shoptalk Fall in Nashville, titled Where AI Agents Are Working in B2B Commerce and Why. Andy opened by framing AI as an unusual case, a topic that is simultaneously overhyped and underhyped depending on where you look. Most of the public conversation focuses on the first mile, discovery through tools like ChatGPT or Anthropic’s own front-end shopping features, and the last mile, delivery innovations like drones and driverless trucks. Andy’s own book, Bot to Bot, argues the more consequential B2B activity is happening in what he calls the middle miles: pricing, quoting, availability checks, and distributed order management, much of it running behind the scenes without buyers realizing AI is involved at all.
Andy tied the urgency directly to response speed. He described what he called the golden five minutes, the modern successor to what used to be called the golden hour, in which a buyer’s odds of converting with a given vendor fall off by roughly 80 percent if that vendor takes longer than five minutes to respond. Nearly 80 percent of B2B customers, he said, purchase from whichever vendor responds first, regardless of price. That dynamic is a big part of why AI agents fit B2B so well: they can respond instantly and consistently at a scale humans cannot match.
Andy pointed to two live examples of middle-mile AI already in production. Sysco’s Swap and Save feature uses order history, pricing data, and product compatibility models to recommend cuisine-specific product substitutions, such as suggesting a lower-cost domestic beer a restaurant can still sell at a premium price point. Fastenal has built an agent that proactively flags unusual consumption, telling a customer they are using an abrasive disk 20 percent faster than usual and asking whether they would like to increase their safety stock, an industrial version of the Internet of Things idea applied directly to reordering.
Jon Feldman: agents are already doing the practical work, not the magic
Joining the conversation was Jon Feldman, Director of Product Marketing at Salesforce, working within Salesforce Commerce on the B2B side. He described AgentForce as Salesforce’s umbrella name for its full range of AI agent capabilities. Feldman agreed with Andy’s middle-miles framing and split the opportunity into two categories: agents that help operate the site itself, such as merchandising, and agents that help operate the backend supply chain.
On merchandising, Feldman described AI agents that can pull an existing catalog from wherever a company currently manages it and push that data into a new B2B storefront, cutting setup time significantly for companies trying to get a digital B2B channel live quickly. On the supply chain side, he gave the example of an order routing agent that monitors weather data and automatically reroutes orders around a warehouse when it detects an approaching severe weather event that would otherwise disrupt fulfillment from that location.
What B2C companies moving into B2B should prioritize first
Brian asked what advice Feldman would give a B2C company evaluating where to apply AI as it expands into B2B. Feldman said the core difference is that B2B runs on relationships in a way B2C generally does not, and AI should be used to amplify that relationship rather than replace it.
You wouldn’t expect to call into a sales agent and have them be like, well, I got ten percent of the catalog available, I just can’t help you with anything else, go send a fax. B2B experiences are more tolerant of not having a TikTok feed in it, but they’re much less tolerant of what the actual speed of experience and transaction is.
Jon Feldman, Salesforce
In practice, that starts with catalog completeness. B2B buyers, he said, care far more about transaction speed and inventory accuracy than about a flashy front-end experience.
Reorders, subscriptions, and the shift toward agent-to-agent negotiation
Brian raised the post-purchase environment, noting how much of B2B revenue comes from recurring reorders and replenishment rather than first-time purchases. Feldman said Salesforce’s approach is to make reordering as fast and frictionless as possible across whatever channel a buyer prefers, including reordering directly through WhatsApp. He described two levels of agent involvement: today, an agent can complete a reorder transaction directly with a person, freeing up a salesperson’s time for higher-value conversations. Looking further out, he expects buyer-side and seller-side agents to negotiate directly with each other on price and terms for routine reorders, removing friction from transactions that are largely repetitive and time-consuming today.
Feldman offered his own practical example of agentic workflow: he uses Claude Code to populate a Google spreadsheet directly from a keynote script he writes for Salesforce’s Dreamforce conference, handing a production company structured data without retyping it himself. Both hosts agreed the near-term reality of agentic AI is less dramatic than the popular narrative suggests. Andy closed by noting that while it is difficult to make firm predictions about where B2B AI will be in three to five years given how quickly the space is moving, the directional signal is unmistakable: toward AI absorbing repetitive, computation-heavy work so people can focus on higher-value activity.