Finding the Intelligence in AI: Where Agents Belong in Manufacturing and Distribution

Richard's problem was simple to state. White River's phones shut off at 5:00 PM Central but its biggest market, California, was going to voicemail after 3:01 PM Pacific. He wanted customers to be able to reach the company 24 hours a day, 365 days a year, and over the past year he built six AI agents to make that happen.

Here's our view on where this is going. The manufacturers and distributors who win with agents will be the ones who decide, step by step, where intelligence belongs. In practice that comes down to three things:

1. Give agents the routine work around a sale.

2. Keep people on the judgment.

3. Match every step to the lightest tool that does the job.

Richard's build runs through this piece, alongside independent research and examples from companies building tools for this market.

AI is everywhere. Value is harder to find.

Most B2B companies are already doing something with AI. Far fewer are getting much out of it.

MIT's NANDA initiative put a number on it. Its 2025 study, reported by Fortune, found that 95% of generative AI pilots produced no measurable impact on the P&L. More than half of AI budgets went to sales and marketing tools, while the biggest returns showed up in back-office automation.

We read that as a placement problem. Companies are putting AI in places where it looks impressive, and skipping the unglamorous work where it would pay off. The fix starts with deciding where it belongs.

1. Give agents the routine work around a sale

Here's what that looks like in practice. After hours, an agent named June answers every call and works out what the caller needs. The agents check live stock in NetSuite in about three seconds, a speed it took Richard ten rounds of tuning to reach. They can find the right product when a caller only has a description, set up trade accounts, send catalogs, and summarize every conversation for the right sales rep.

Richard also told me he had no idea how many calls they had been missing until the agents went live. People were calling at 8 p.m., midnight, and 2 a.m. As he put it, the company needed to be reachable when customers wanted to reach it.

The same pattern shows up in the tools being built for this market. Oro's AI SmartOrder reads purchase orders that arrive as emailed PDFs or images and turns them into draft orders, and its SmartAgent answers product and order questions around the clock. Oro says SmartOrder can cut manual entry by up to 80%, though that's the company's own figure.

There's a customer experience case here as well as an efficiency one. McKinsey's 2026 B2B Pulse found that inconsistent information across a supplier's teams is now the top reason buyers switch suppliers, ahead of not being able to reach a knowledgeable rep. Consistency is something agents do well. Watching Richard's agents, one detail caught our attention: they mention the trade account discount every single time. People forget to do that. 

2. Keep people on the judgment 

Richard was deliberate about this from day one. He designed the agents to work alongside his sales reps, taking on the tasks they didn't enjoy and setting them up for the next conversation. During business hours, an agent only picks up if a caller has been waiting more than 60 seconds. When a customer asks for a person, the agent transfers the call, and if the rep is busy it sends them a summary to follow up on. Every after-hours conversation lands in the right rep's inbox by morning. His reps went from asking whether this would take their jobs to saying they get it.

He's also clear about what the agents can't do yet. They can't cancel or change an order, which matters when a customer has just put $5,000 on a card for the wrong product. People still handle that. As the system has grown, he added a supervisor agent that reads every conversation and sends him a weekly report flagging the ones that need a human's attention.

Oro builds the same rule into SmartOrder: the AI produces a draft, and a person checks and amends it before it becomes an order. Forgepoint takes the same approach with quotes, keeping pricing and engineering decisions with its customers' teams.

Gartner expects that by 2028, organizations using multiagent AI for 80% of customer-facing processes will outperform competitors, with AI handling routine interactions and people stepping in for complex or sensitive cases. That's the split Richard designed for.

3. Match every step to the lightest tool that does the job

We shortened artificial intelligence to "AI," and perhaps we sometimes forget what the second word really means. Intelligence is only needed at the points where a decision gets made. As we've built agents inside the association, model choice has become one of the biggest decisions we make. Quick math is faster in a Python script than in any model. Some tasks just need to run on a schedule. Some need a small, fast model, and a few need the heaviest one available.

Richard learned the same lesson the hard way. His first agent took more than 100 hours, and at one point Claude told him that several steps in his workflows didn't need a language model at all, because a simple automation would get the same result faster.

Model choice mattered in front of customers too. On his video agent, a heavier model spent so long reasoning that the pauses made the agent look broken. He switched to a lighter model, pulled fewer documents into each answer, and the conversations got smoother. He also caps calls at around ten minutes, because past that point the agents start losing track of what was said earlier.

Anthropic gives agent builders the same advice in its guide to building effective agents: find the simplest solution possible and add complexity only when it's needed. Fixed workflows suit well-defined tasks, and agents earn their place where a task needs flexibility and judgment from the model.

Gartner predicts that by 2027 companies will use small, task-specific models at least three times as much as general-purpose ones, because they respond faster and cost less to run. It recommends combining multiple models and workflow steps where one model falls short, which is a fair description of what Richard built.

The data underneath all three

None of this works without product and system data the agents can trust. Richard's agents can explain the difference between two similar panels because he gave them a library of product documents to draw from. They can describe stock levels because he built a direct connection to NetSuite. Before that connection existed, the agents couldn't check inventory at all.

McKinsey calls data the harder problem for most established companies, because agents can't make trusted recommendations when customer, product, and pricing data are fragmented. For many distributors, that product data still lives in PDFs and spec sheets.

That's the problem Kaavio is working on with industrial and science distributors, pulling descriptions, images, and structured data out of supplier documents so it can be used online. It's unglamorous work, and it decides how useful an agent can be.

Where this is heading

So far, we've talked about sellers using agents. Buyers are moving faster. Deloitte's 2026 research found that nearly 40% of B2B buyers already use agentic AI in purchasing, to evaluate products, configure orders, and benchmark prices. Only 24% of suppliers use agents in their sales process, though 67% plan to.

More requests for quotes, stock checks, and order changes will arrive at odd hours, some of them sent by a buyer's own agent. The suppliers who answer quickly and accurately, and who know exactly when to bring in a person, will win that business.

There's a compounding effect too. If you've called a supplier after hours and nobody picked up, you probably stopped trying. Once you get a good answer at 9 p.m., you'll call at 9 p.m. again. Richard says he's already seeing that. His next step is letting agents change and cancel orders, which he calls the missing link.

Where to start

If you're looking at your own quoting, order, or service process, map it before you build anything.

1. Write down every step, from the moment a request comes in to the moment it's closed.

2. Mark the steps that need a person's judgment: pricing calls, engineering review, exceptions, relationships.

3. Mark the steps that need a model to read, interpret, or write something.

4. Mark the steps that only need a rule, a script, or a schedule.

5. Automate the rule-and-script steps first, add models where they earn their place, and keep people on the judgment calls.

In our experience, more of the work lands in the rule-and-script group than people expect. That makes the first wins faster and cheaper to get.

Guest: Richard Enriquez
Richard Enriquez is Director of Digital Commerce and AI Systems at White River Hardwoods, a family-owned manufacturer of hardwood moldings. He built a team of six AI agents that answer customer calls and website chats around the clock, all without a background as a programmer.

Host: Justin King
Justin King is the Chief AI Officer at the B2B eCommerce Association. He hosts the Open Laptop Sessions and leads the B2B NEXT AI Community. His work centers on designing AI-native systems where coordinated teams of AI agents run real business operations. He’s known for bridging executive vision with hands-on technical execution, building systems that move beyond demos into everyday use.