Sponsored Content from Orocommerce

Eighty percent of business-to-business companies say they have deployed artificial intelligence (AI). Only eight percent say they have integrated it across their operations. That gap, documented in the B2B AI Benchmark Report, defines the central problem facing wholesale distributors: the distance between switching on an AI tool and getting durable value out of it. 

The gap was the starting point for a DSG Tech Talk conversation between Ian Heller, chief strategy officer for Distribution Strategy Group, and Aaron Sheehan, vice president of strategy at OroCommerce. Sheehan, who works with distributors deploying AI across commerce and back-office systems, said the failure to move from pilot to production usually traces to two technical barriers: data and connectivity. 

“Nothing quite surfaces the gaps in your data like just pointing an AI tool at it,” Sheehan said. Decades of information accumulated in ERP systems and aging SQL databases carry inconsistent schemas and null fields, and those holes become visible the moment a model is asked to produce insight from them. The second barrier is interoperability. A generic AI tool wired into an ERP, a product information management (PIM) system, a CRM platform and an ecommerce catalog inherits the data-quality problems of every system it touches. 

Where AI Actually Lands on the Shop Floor 

The board-deck version of AI adoption and the shop-floor reality are two different things. What Sheehan sees most often is Microsoft Copilot spreading through the office suite, writing emails, analyzing spreadsheets and building presentation decks, then extending into SharePoint and document systems. Excel, he said, is the primary vector for generative AI moving through a distributor, because any large body of human-readable data is a natural place for a language model to sit and start working. 

Adoption beyond that is inconsistent. Point solutions have added AI features inside transportation management systems, PIM tools and ERP modules, but the depth of use varies widely from one distributor to the next. 

The lowest-risk, highest-return entry point, Sheehan said, is order automation. OroCommerce built a tool that converts a purchase order from a PDF into a digital order because customers asked for it. The task it replaces is familiar: a sales rep opens Outlook on one monitor, logs into the ERP on another and keys in the order by hand. Automating that step returns time to a sales team, Sheehan said, that a distributor is paying to build relationships and close business, not to perform data entry. 

The benchmark data tracks with that maturity level. Forty-eight percent of distribution leaders describe their AI results as positive but not transformational. Sheehan called that “exactly what a mature first inning should look like if you sold it as a first inning internally.” The discontent, he said, appears where AI was sold as magic and executives expected late-inning performance from an early-inning capability. 

The Interface Shifts, the Infrastructure Gets More Valuable 

One question distribution technology leaders are wrestling with is whether AI becomes the interface while the ERP, CRM, PIM and commerce platform recede into the plumbing underneath. Sheehan said that is the framing OroCommerce hears most often, and he expects it to prove out over time, but not yet, because the data is not clean enough to support it. 

He drew a distinction between systems of record and systems of action. An ERP should be a system of record, with logging, auditability and security. It does not have to be the system of action. Decoupling the two puts pressure on the seat-based licensing model much of the software industry is built on, since value shifts from counting user seats to outcomes, tokens or flat fees. 

The benchmark found that 60 percent of companies are adding AI to systems they already own rather than replacing them. Sheehan said that fear is rational and the strategy is sound. Embedded AI built by the vendor already carries the business logic and context of the system it runs inside, so it understands what a state change or a workflow means. A third-party tool bolted on top has to be taught all of that from scratch. 

That view shapes his skepticism toward marketing language. When a vendor claims to be “AI native,” Sheehan said, he assumes it means part of the stack was rebuilt, not all of it, and that the term will not carry a stable meaning for another couple of years. He is equally wary of the argument that distributors should vibe-code their way out of their software contracts. The cost of that approach, he said, tends to surface in about six months, when something breaks or gets hacked and the contractor who built it has moved on, leaving undocumented code no one can maintain. A distributor’s data and business rules are its moat, he said, and those are precisely the assets that belong in secure, governed systems. 

The DiversiTech Sequence 

The buying sequence that follows from that logic shows up in one OroCommerce customer. DiversiTech ran 12 ERP systems and chose to put a commerce layer on top rather than consolidate all those systems first. 

Sheehan called the sequence wise. An ecommerce platform holds the most context about what customers are buying, because it has to surface order history, catalog entitlements, tier pricing, quotes and purchase orders at the line-item level. Establishing that layer as a common operating system for both the sales team and customers lets a distributor make downstream ERP changes, including acquiring or divesting businesses, without retraining everyone. It also flattens the 12 different ways a dozen ERPs describe a product or a customer into one coherent database. That consolidated, standardized data, he said, is the foundation AI needs before it can automate anything reliably. 

Governance Separates the Value Generators 

The benchmark data on governance is stark. Only 4 percent of companies have comprehensive governance in place, yet at least 95 percent of the distributors generating high value from AI have at least basic governance. Sheehan acknowledged the correlation-versus-causation question, whether governance drives success or simply marks a well-run company, but said the pattern is not a coincidence. 

He compared AI governance to PCI compliance for payment data: a program that has to keep evolving, not a certificate filed in a drawer. Governance that was valid in January, he said, may be obsolete by August given how fast models and products change.  

That matters because pointing high-volume automation at a business surfaces decisions no one anticipated: an order type the company has never seen, a margin it is unsure about, a customer request it has never approved or denied. Someone has to own each of those calls. IDC’s guidance, cited in the conversation, is to name agent owners in every function. Sheehan said the natural owners already exist. A pricing manager owns the pricing agent. A credit manager owns the credit agent.  

AI should not be treated as an IT-owned realm that imposes rules on the rest of the business. 

Autonomy Stops at Reversibility 

Data quality remains the barrier few executives want to fund. Two-thirds of distributors say their data is not standardized across systems, a problem that predates the current AI cycle by 20 years. Sheehan’s advice is to start narrow rather than boil the ocean: pick one use case, document the data it touches and fix that slice first. He pointed to Waymo, which mapped and launched one city at a time, as the model of disciplined, sequenced investment that eventually reaches scale. 

On autonomy, Sheehan said distributors should treat an AI agent like a new hire that needs supervision, not an infallible oracle. At a recent OroCommerce customer advisory board, every participant said they would not allow AI to write data to their ERP under any circumstances. The line Sheehan drew is reversibility. Anything irreversible, such as ending a customer relationship or refusing an order, requires a human signature. A single credit-score dip cutting off a long-term customer is the kind of unexplainable call that agents must be built to escalate, not execute. 

That constraint extends to the technology’s next phase. Agentic commerce, in which AI buys, pays and triggers fulfillment on its own, remains largely slideware for distributors, Sheehan said. Gartner’s most recent digital commerce hype cycle places agentic buying at the far bottom left, five to 10 years from mainstream adoption. Machine-to-machine commerce already exists in distribution as electronic data interchange (EDI), he noted, which raises the question of what agentic commerce would do that EDI does not. Where he does see movement is discovery, with AI agents beginning to identify suppliers on behalf of buyers based on brands, certifications and shipping coverage. 

The competitive stakes sit in that discovery layer. Clean, standardized data lets a distributor add channels, attract buyers and appear when an agent or a crawler goes looking for a supplier. Incomplete data leaves a distributor invisible to the systems that increasingly decide who gets found. The distributors closing the gap between deployment and integration, the benchmark shows, are not the ones with the best model. They are the ones that fixed their data and brought their people along.