How to make a distributor AI-native: a playbook

An ordinary morning, not a reported case.

At 7:40 a mechanical contractor emails a PDF purchase order: 38 lines, half in the contractor's own part numbers, two for fittings you stock in boxes of 25 and the PO lists as eaches. Page three adds a clause charging you for late delivery. An inside rep keys it, looks up the cross-references, checks the contract price, and misses the clause because the branch has six more POs waiting. At 8:15 a supplier emails a 6% increase on copper fittings, effective in 30 days. Nobody loads it until the first invoice arrives at the new cost, after three quotes went out at the old one.

Neither event is unusual. The first will cost a credit memo for the wrong pack size and maybe a late penalty nobody agreed to on purpose. The second costs the margin on three quotes.

The first figure is the whole budget. About 20 cents of gross margin per sales dollar pays for branches, trucks, inside sales, buyers and credit, and payroll and benefits are close to half of the 13 cents it costs to run the business[1,2]. The second is the squeeze: supplier costs are moving faster than sell prices. On top of that, tariffs went on in 2025, were struck down by the Supreme Court on 20 February 2026, and are being refunded to importers[27,30]. Every one of those changes lands on order lines that people still key from PDFs.

The third figure counts any AI use at all. At the other end, Sysco has put a target of about $100 million of fiscal 2027 savings on "an artificial intelligence (AI)-powered transformation of business processes and customer engagement"[16]. Most mid-size distributors sit between the two.

A distributor is AI-native when agents do the reading, keying, matching and chasing on every order line and supplier document, and people spend their time on pricing judgment, customers, credit and exceptions, with one set of rules and one record across all of it. The test is simple to run on any line: who keyed it, who set the price, and can you show it six months later when the customer disputes the surcharge.

Few distributors will replace Prophet 21, Eclipse, Infor, NetSuite or SAP to become AI-native. Agents get added around the ERP: from the ERP vendor, from order-capture and pricing tools, or set up by the firm. Each comes with its own settings. The ERP governs its own agents; the pricing tool governs its own. None of them governs an agent that reads an email, checks a contract price in the ERP, looks up a certificate in the tax engine and replies to a customer in one task.

OrchKernel sits in that gap. Agents ask it before they act; it checks the rules, holds what needs a person, and records the result. It does not replace the ERP, the pricing engine, the tax engine or the WMS, which stay the systems of record. Details are in the OrchKernel blueprint.

An order line passes eight desks between the request and the supplier claim that may follow it. At each, an agent can read and draft today, and something should wait for a person.

Categories, not recommendations. Every performance figure below is the vendor's own claim, not an industry benchmark.

The large distributors put it in their filings. Sysco names inventory forecasting, "improved coding efficiency", routing and back-office automation as the sources of its target[16]. Wesco "established an AI Governance Council in 2025"[13]. Fastenal uses AI "to analyze customer usage data to recommend optimized parts and quantity" for its vending devices[11].

The buyer side is automating too. Wesco warns that customers' "algorithmic or AI-enabled pricing, procurement or sales tools" could "pressure margins"[13].

Adoption is low and rising. 19.2% of wholesale businesses used AI in the two weeks to 20 September 2026, up from 12.6% in November 2025, and 24.7% expect to within six months. In the same survey, 49.0% of information firms and 43.9% of professional services firms used it[6]. A firm where one person used a chatbot once counts, so this says nothing about depth.

Where the money went. Into tools that sit on the ERP, and into food: GrubMarket, which owns food distributors and builds their software, raised $50 million at a valuation above $3.5 billion in 2025[53]. We found no AI-native industrial distributor at scale. The incumbents outspend most startup rounds: Wesco reported $62.1 million of digital transformation costs over the twelve months to Q2 2026[14].

The spread differs a lot by line of trade. In 2022, gross margin was 32.0% of sales for hardware, plumbing and heating wholesalers, 31.6% for machinery, equipment and supplies (MRO and fluid power), 27.6% for electrical and electronic goods and 16.2% for grocery[1]. Across all merchant wholesalers, operating expenses took 13.2% of sales[2].

People are the largest cost you control. Payroll and fringe benefits were 48.9% of wholesaler operating expenses in 2022, about 6.4 cents of every sales dollar and about a third of gross margin (computed). Data processing and purchased computer services were 0.6% of operating expenses[2]. Fastenal puts employee-related expenses at 70% to 75% of its SG&A[11].

At public distributors, overhead eats most of the gross profit. SG&A took between 55% and 88% of gross profit across nine of them in fiscal 2025: Fastenal 55%, Core & Main 56%, Grainger and Applied 64%, Watsco 66%, Wesco 71%, DXP 72%, Global Industrial 80% and Distribution Solutions Group 88% (computed)[8].

Percentages of revenue[8]. Wesco sells large electrical and datacom projects at thin margins; Fastenal sells fasteners and vending programs at high ones. Compare yourself with your own line of trade.

A note on the e-commerce figure. Census counts a sale as e-commerce when the order or its terms come through the internet, an extranet, an EDI network or email[4]. So the 28.6% includes PDFs that someone still keys by hand, and the other 71% came by phone, fax, counter, rep visit and the like. Count your own channels before deciding which inbox to automate.

Wholesalers are squeezed on cost changes. In the September 2026 Census survey, 49.3% of wholesalers said the prices they pay had risen in the past two weeks; 25.6% said the prices they charge had[6]. Each supplier increase has to reach list and matrix prices, contract accounts that allow a pass-through, open quotes and open orders before it reaches an invoice. When that takes weeks, the difference comes out of margin.

Then the tariffs reversed. Tariffs imposed under the International Emergency Economic Powers Act were passed into prices through 2025. On 20 February 2026 the Supreme Court held that "IEEPA does not authorize the President to impose tariffs"[28]. The same day, the President imposed a 10% import surcharge under Section 122 of the Trade Act for 150 days, from 24 February to 24 July 2026[29]. CBP now refunds IEEPA duties to importers of record through its CAPE process[30]. Grainger booked $43 million of refunds against cost of goods sold in Q2 2026, after growth that came partly from "price inflation as tariff costs are passed" through[10].

That made surcharges an audit question. A customer who paid a 2025 tariff surcharge can ask what happened to it. Answering means knowing which invoices carried it, which tariff it recovered, and who approved the logic.

This may be the first agent worth building. Matching a supplier price file to your items and listing everything it touches is slow, careful and mechanical. The decision to apply it stays with the pricing analyst, and the record of that decision is what you show the customer later.

Three rules make it work. Contract-locked items are excluded unless the contract allows a pass-through. A surcharge is a separate line tagged with its source. A refund runs through the same approval, so the firm decides on purpose whether to pass it back.

Six stages that follow the order line: data you can trust, then drafts that people release, then the supplier side, then touchless lines under rules, then pricing, inventory and credit decisions with guardrails. Stages 1 and 2 can run side by side. Stage 4 waits until the rules and data are proven, because that is where the legal exposure sits.

Across all stages, the agent that came with your ERP acts with whatever permissions it was given. In roll-ups running several ERPs, the same customer may have three part-number lists, and an agent matching against one of them will miss lines the others would catch. Merge them in Stage 0.

Stage 0 for one branch or desk, then Stage 1 in draft mode on one inbox, with the first supplier price file run through approval near the end. We do this with design partners.

Agents can prepare every one of these: the summary, the comparison, the draft. A named person makes the call, and the log shows who.

The recurring requirements: know when your agent has committed the firm, be able to explain every price difference between customers, keep surcharges traceable to their source, and never let an agent blind the firm to a red flag or a missing certificate.

Real cases first. Most are not AI failures. The worst recent distributor losses were outages, and agents connected to every system raise the stakes of credentials and stop switches.

Scenarios, not reported cases. Each maps to the control points that stop it.

Inferred from the work, not from a survey; we found no data on how distributors are changing roles. Wholesale employment was 6.08 million in September 2026, up 0.4% on a year earlier[7]; no public data yet shows AI reducing headcount at sector level. In each role below the agent assembles the order, the comparison or the claim, and the person keeps the decision.

Record a baseline before Stage 1 and track the same numbers at each stage. Census and SEC filings give sector margins, the inventory-to-sales ratio and public-company overhead. For every workflow measure below, no public benchmark exists, so compare each stage with your own Stage 0 numbers.

OrchKernel sits between AI agents and the distributor's systems, as drawn in the missing layer. Agents ask it before they act; it checks the rules, holds what needs a person, and records what happened.

What it is not. OrchKernel is not an ERP, a pricing engine, a tax engine, a WMS or a rebate system, and those stay your systems of record. It cannot clean your item master, verify a bank change by phone, screen against denied-party lists by itself, calculate tax, or keep the ERP running through a ransomware attack. It can stop agents acting outside the rules, and show who approved what.

Where a distributor needs a control whatever tools it uses, who owns it, how OrchKernel enforces it, and what stays in another system.

Limits. An agent inside a vendor's product that does not go through OrchKernel follows that vendor's controls, so list those. OrchKernel is source-available under the Business Source License and runs on your own servers, so you can read the code that enforces these controls.

• Workflow benchmarks for distributors: line fill rate, order lines per inside sales hour, quote win rate, override rate, rebate capture, DSO. NAW, MDM and buying groups may hold them for members; nothing public could be opened.
• How orders split by channel at a mid-size distributor. Census counts emailed orders as e-commerce, so it cannot show how many lines are keyed by hand.
• How much of distributor profit comes from rebates and supplier programs, and how much is left unclaimed. Only vendor claims exist.
• An independent survey of distributor AI use by workflow. The Census survey asks whether a firm used AI at all, not where or how deeply.
• What the ERP vendors' own agents do with permissions, logging and approvals. Their pages could not be opened for this review.
• How customers' procurement agents will negotiate with distributors' agents, and what contract terms about AI will look like. Plausible, not yet evidenced.

Last reviewed October 2026. Sources were read in October 2026; dates are publication or data dates.