If you’ve written off AI as overhyped, you’re not wrong, at least about most of it. MIT recently found that 95% of companies investing in GenAI have seen zero measurable bottom-line impact.

What that research also found, though, is that the 5% who got it right didn’t have better AI. They had better foundations underneath it. That distinction is worth understanding, because the gap between the two groups is widening fast.

UPS has cut over 48,000 jobs since 2025, with 30,000 more planned this year, under a restructuring built entirely around automation and AI-driven logistics. At their flagship facility in Louisville, robots now outnumber workers 15 to 1.

As one of the most operationally disciplined logistics businesses in the world, when UPS says AI is changing the economics of running a warehouse, that should mean something to anyone running a distribution business.

Why Most AI Projects Fail

The AI your vendors have sold you over the past few years is, for the most part, purely analytical. It watches your data and surfaces recommendations: follow up with this account, review this order, chase this invoice. Helpful, sure, but it’s only adding to your workload since a human still has to act on every single recommendation.

The problem is that most of your team is already at capacity before the AI surfaces a single thing. They have emails to answer, orders to key in, quotes to turn around, and customers who needed a callback an hour ago. That’s just a normal Tuesday. Analytical AI doesn’t reduce any of that load. It piles on top of it, handing people a growing to-do list they have no time to actually act on.

Agentic AI is different. Agents can jump across systems to execute workflows end-to-end. It can read a customer email, pull up the account, cross-check inventory, apply your pricing logic, flag the exception, and close the loop without a human telling it what to do at each step.

The AI models 18 months ago weren’t reliable enough to handle tasks like this, but today they are. So much so that Gartner projects that by the end of 2026, 40% of enterprise applications will include AI agents, up from less than 5% in 2025.

Where Distribution Is Most Exposed

Distribution is one of the highest-opportunity industries for agentic AI, precisely because so much of its operational cost is people doing predictable, rules-based work across disconnected systems.

• Sellers spend the majority of their week on tasks like order entry, data cleanup, pricing exceptions, and figuring out which accounts to call.
• Pricing and quoting teams manually pull price sheets, check contract tiers, apply exceptions, and confirm availability. A single complex quote can take an hour.
• Purchasing and replenishment buyers monitor inventory levels, review reorder reports, and generate purchase orders. This is work that is largely pattern recognition, the exact task category where AI consistently outperforms humans on speed and accuracy.
• Collections teams review aging reports, draft follow-up emails, and log calls, executing rule-based decisions about who to contact, when, and how.

All of these tasks center on someone moving information between systems, then acting on it. That is precisely what agentic AI is made for. The goal is to automate these busywork tasks so that the people doing them can focus on work that actually moves the needle.

A seller who is not spending half their day keying in orders is calling the accounts that have gone quiet or have open quotes, pitching new products, and expanding wallet share. A buyer who is not manually generating routine POs is negotiating better terms with suppliers, building out backup vendor relationships, and getting ahead of supply disruptions instead of reacting to them. That is where the real margin is. Not in doing the same tasks faster, but in freeing up the people who know your business to do the work that compounds over time.

The Real Reason Your Last Pilot Failed

The problem I see with most distribution AI projects is the underlying data infrastructure.

An agent that needs to process a purchase order has to read an email, check a product catalog, look up live inventory, apply pricing rules, and flag substitutions. Moving across five or six systems in a single workflow. If those systems don’t share data in real time, the agent stalls at the first seam between them.

That’s exactly what MIT found in their study. The 5% who successfully scaled AI had deeply integrated it into specific workflows. The 95% who didn’t were running generic tools that didn’t understand their business or target a specific problem.

The good news: this is fixable. Your tools need to be connected on a shared data layer, so an agent can move between your ERP, CRM, WMS, and quoting tool without interruption. Not loosely integrated with periodic sync jobs.

What It Looks Like When It Works

When the foundation is in place, the outcomes are straightforward.

A customer’s PDF purchase order arrives, gets parsed by an agent, matched against your catalog, checked against live inventory, and lands in front of your rep as a confirmed draft in minutes, not an hour. Your product data team’s two-year enrichment backlog clears in days because an agent is continuously pulling from manufacturer sources, writing descriptions, and populating fields across your catalog. Your AP team stops processing routine invoices and handles only the exceptions, because an agent is matching POs to received quantities and escalating only what it can’t reconcile.

The Compounding Problem

Here is what should concern a skeptic more than the technology itself.

Agentic AI is available to everyone. A unified data foundation is not; it takes time and investment to build. The distributor who builds it first doesn’t just run better AI today. They run AI that compounds: getting smarter on their customer data, pricing patterns, and order history every month, while competitors still try to get their systems to talk to each other.

You don’t need to restructure your whole company to get there. You just need to get your data connected and deploy agents on top of it. The distributors doing that today are going to be difficult to catch in two years, not because the technology they’re using is proprietary, but because a six-month head start on compounding is very hard to close.

The clock is running. The foundation is the bet worth making.