Segments covered: merchant wholesalers that take title to goods, with examples from industrial supply and food-service distribution; manufacturers' sales branches and agents or brokers are outside this guide's lens. Linked workflows: purchase-order intake, inventory forecasting, catalog enrichment. Linked roles: operations leaders, finance leaders, sales operations leaders.
Distribution connects information to physical commitments. A quantity becomes inventory allocation; a product code becomes a warehouse pick; a forecast can influence a purchase. The value of AI depends on whether those transitions become more reliable and economical, and in a business with a thin spread between gross margin and operating cost, small errors carry through to profit quickly.
This guide combines three kinds of material. Sector figures come from US Census Bureau and Bureau of Labor Statistics publications and are cited with their survey, period, and limits. Recommendations about sequencing, ownership, and controls are proposed practice: editorial reasoning, not findings. The distributor in the sequencing section and the $200 million firm in the economics section are invented. No distributor was interviewed and no deployment was observed for this chapter.
How the business makes money
A merchant wholesaler buys goods, holds them, and resells them to businesses, earning the difference between sales and the cost of goods sold. What remains after operating expenses (warehousing, delivery, selling, order handling, and administration) is small relative to sales, and inventory and receivables tie up working capital. Those three quantities, gross margin, operating expense, and working capital, are where an AI effect has to show up.
The Census Bureau's final Annual Wholesale Trade Survey covers 2022; the survey has since moved into the Annual Integrated Economic Survey. Its 2022 tables were revised to benchmark against the 2022 Economic Census and re-released on August 31, 2026. For merchant wholesalers excluding manufacturers' sales branches and offices, those revised tables report:
| Measure, 2022 | All merchant wholesale (NAICS 42) | Durable goods (423) | Nondurable goods (424) |
|---|---|---|---|
| Sales | $7.89 trillion | $3.74 trillion | $4.16 trillion |
| Gross margins as a percent of sales | 20.3% | 27.2% | 14.1% |
| Operating expenses as a percent of sales | 13.3% | 17.2% | 9.8% |
| Year-end inventories | $917 billion | $546 billion | $370 billion |
Source: Census Bureau, 2022 Annual Wholesale Trade Report, benchmarked tables 1, 3, 4, and 5. The population is US employer firms primarily engaged in merchant wholesale trade; estimates are not adjusted for price changes, and the Census Bureau suppresses estimates that fail its quality standards. AWTS methodology
Two ratios follow by our own arithmetic, not as Census-published figures, and the Census Bureau notes that figures derived from its tables carry the same limitations and should not be attributed to it. The gap between the published gross-margin and operating-expense percentages is about 7 points of sales for the aggregate (20.3% − 13.3%). Year-end inventory is about 11.6% of annual sales ($917 billion ÷ $7.89 trillion). Both are sector aggregates. Neither describes a particular firm, and the gap is not a profit measure: it omits interest, taxes, and other items.
Segments differ enough to matter. In the same tables, grocery and related products wholesalers (4244) show a 16.2% gross margin and 11.7% operating expenses, while hardware, plumbing, and heating wholesalers (4237) show 32.3% and 20.2%, and machinery, equipment, and supplies wholesalers (4238) show 31.5% and 20.8%. A food-service distributor handling many low-value lines and an industrial distributor selling technical parts face different economics, different error costs, and different data problems. This guide uses both as examples and does not treat the sector as uniform.
Margins and expenses differ widely by segment
- Operating expenses
- Gross margin
- Spread between them
Show the data
| Segment | Gross margin, % of sales | Operating expenses, % of sales | Spread (our arithmetic) |
|---|---|---|---|
| All merchant wholesale (42) | 20.3% | 13.3% | 7.0 pts |
| Durable goods (423) | 27.2% | 17.2% | 10.0 pts |
| Nondurable goods (424) | 14.1% | 9.8% | 4.3 pts |
| Grocery (4244) | 16.2% | 11.7% | 4.5 pts |
| Hardware, plumbing, heating (4237) | 32.3% | 20.2% | 12.1 pts |
| Machinery and equipment (4238) | 31.5% | 20.8% | 10.7 pts |
For scale, the Bureau of Labor Statistics estimates wholesale-trade employment at 6,072,100 in August 2026, seasonally adjusted and preliminary. BLS Current Employment Statistics, series CES4142000001 That count says nothing about how work in the sector is changing.
Where AI could touch the value chain
The table maps major distribution processes to the capability involved, the chapter that covers it, and where a benefit would land. "Not yet covered" means the Field Guide has no chapter for it; the row is there so readers see the whole business rather than only the parts written so far.
| Process | Capability | Canonical chapter | Benefit mechanism | Benefit location |
|---|---|---|---|---|
| Customer order capture from emailed documents | Extract, classify | Order intake | Added capacity; fewer entry errors, credits, and reshipments | Operating expense; risk |
| Orders already arriving by EDI or portal | None needed | Ordinary integration | Already structured; exception handling may still need work | Operating expense |
| Product master and catalog attributes | Extract, generate | Catalog enrichment | Faster item setup; fewer clarification calls; possible search and order effects (unmeasured) | Operating expense; revenue hypothesis |
| Replenishment and purchasing | Predict | Forecasting | Lower average inventory or less expediting at the same service level | Working capital; operating expense |
| Supplier invoices and receiving documents | Extract | Document extraction; finance leaders | Added AP capacity | Operating expense |
| Customer order-status questions | Retrieve | Customer portal and ordinary software first | Fewer calls if status data is reliable | Operating expense |
| Pricing and quoting | Rules, predict | Not yet covered | Margin effects; contractual and authority questions | Revenue; gross margin |
| Warehouse slotting, picking, and delivery routing | Optimize | Not yet covered | Labor and freight efficiency | Operating expense |
| Receiving inspection and damage checks from images | Classify | Computer vision | Fewer missed damage claims; review effort follows the false-alarm share | Operating expense; risk |
| Credit and collections | Classify, predict | Not yet covered | Fewer write-offs; faster cash | Working capital; risk |
Three observations follow from the map, as editorial reasoning.
First, the three written workflows share a dependency: the item master. Order intake maps customer part numbers to SKUs; catalog enrichment maintains the attributes and pack quantities that mapping relies on; forecasting needs stable item identity across substitutions and supersessions. A weak item master degrades all three at once, which is why the sequencing section starts there.
Second, the order-capture row splits in two. Orders that already arrive as structured EDI or portal transactions have nothing to extract. The work there is still real: customer part numbers must be cross-referenced to SKUs, units of measure mapped, contract prices reconciled, ship-to locations validated, and order acknowledgements returned. Those are the same item-master and mapping problems that degrade every workflow in this guide, and ordinary integration and rules address them better than a model reading text. A distributor whose largest customers order by EDI may find that the addressable volume for document intake is the long tail of smaller accounts.
Third, several of the highest-value decisions in distribution, pricing, credit, and warehouse operations, sit outside the current guide. Their omission is a coverage gap, not a judgment that AI has no role there.
Follow the order through the business
An incoming request needs a recognized customer, an unambiguous item, a quantity and unit, commercial terms, and a delivery destination. A system can extract those fields correctly yet still create a bad order if the catalog mapping is wrong. A forecasting system can improve a statistical score yet worsen stock availability if the replenishment policy ignores supplier constraints.
Evaluate the combined process. Separate the interpretation of information from authority to commit inventory, extend credit, or purchase goods. That separation makes a pilot easier to measure and mistakes easier to contain, and it is the common thread in the three workflow chapters.
The three workflows should not share one success metric. Intake measures complete, correct orders and critical fields, including duplicates prevented. Forecasting measures replenishment outcomes weeks later, with service level held fixed. Catalog work measures approved, supported attribute values and correct holds where the supplier did not state a value. A single accuracy figure across them would hide the differences that matter.
Constraints that shape every project
Units and packaging levels. Distribution is full of eaches, inner packs, cases, and pallets, and a unit error moves physical goods. GS1's GTIN management rules assign a distinct GTIN at each packaging level above the base unit, and a change in case quantity requires a new GTIN. GS1 GTIN Management Standard, rule 8 That is the standard's rule; whether a distributor's suppliers and customers follow it has to be checked in local data. The order-intake chapter's worked case turns on exactly this kind of case-versus-each ambiguity.
Master data ownership. Customer part-number cross-references, units of measure, pack quantities, and supersessions need named owners and a change process. A correction should produce feedback without automatically becoming a new global rule: one customer's shorthand may be invalid for another. Track the scope and approver of mapping changes.
Physical irreversibility. An order entered twice can ship twice. A purchase order sent to a supplier commits cash. Writes into the order and purchasing systems need duplicate prevention and safe retries, as the order-intake chapter shows. Whether a particular ERP honors such identifiers is a local test, not an assumption.
Commercial authority. Customer-specific prices, credit limits, allocations during shortages, and substitutions are commercial decisions with owners. An extraction or forecasting system can propose; it should not grant credit, change a contract price, or substitute a product without an authorized person. Customer price agreements may also carry confidentiality terms that constrain what enters a model's context; that is a question for counsel and the contract owner, not something this guide can resolve.
Regulated and safety-relevant products. Food, chemicals, electrical equipment, and other categories carry labeling, handling, and certification requirements that vary by product and jurisdiction. The catalog chapter treats safety and certification attributes as owner-approved fields that are never inferred. Which requirements apply to a given category is a question for qualified advisers; this guide states none.
Legacy systems and peaks. Distribution ERPs may be old, heavily customized, and hard to integrate. Seasonal peaks change order volume and exception rates. A pilot run in a quiet month may understate exception load; the operations leaders guide shows how to size an exception queue using Little's Law before a trial.
Data readiness is concrete. A usable inventory history needs stable item identity, order and fulfillment dates, stock availability, returns, and lead-time information. Recorded sales are not demand when items were out of stock; the forecasting chapter illustrates why the two can differ. Product information needs units, provenance, and explicit handling of missing attributes. Before selecting models, have an experienced employee walk through several difficult orders: split shipments, substitutions, customer-specific product names, amended documents, and duplicate attachments. These reveal requirements a clean demonstration hides.
The sector's economics lens
For your finance lead
The decisive question in a thin-spread business is where an effect lands: gross margin, operating expense, or working capital. These are different quantities and should not be added together.
The following illustration uses an invented distributor with $200 million in annual sales and the 2022 aggregate ratios from the Census tables above. The ratios are sector averages applied to a fictional firm; a real firm's ratios should come from its own accounts.
| Quantity | Calculation | Illustrative result |
|---|---|---|
| Gross margin | $200M × 20.3% | $40.6M |
| Operating expenses | $200M × 13.3% | $26.6M |
| Spread before other items | $40.6M − $26.6M | $14.0M (7.0% of sales) |
| Year-end inventory | $200M × 11.6% | $23.2M |
Where $100 of wholesale sales goes
- Cost of goods sold
- Operating expenses
- Spread before other items
Fictional firm, per year
$100,000 operating saving against a $14.0M spread
5% less inventory
$1,160,000 released once
$232,000 carrying value a year at 20%
Different quantities; never added together.
Show the data
| Quantity | Value |
|---|---|
| Cost of goods sold, per $100 of sales | $79.70 |
| Operating expenses, per $100 of sales | $13.30 |
| Spread before other items, per $100 of sales | $7.00 |
| Fictional firm: spread | $14.0M |
| Fictional firm: operating saving | $100,000 |
| Saving as share of spread | 0.71% |
| Working capital released once | $1,160,000 |
| Annual carrying value | $232,000 |
Operating expense. A $100,000 annual reduction in operating expense equals 0.05% of sales, and 0.71% of the $14.0 million spread ($100,000 ÷ $14.0M). The same result could come from a 0.05-point improvement in gross margin. The comparison helps decide which lever a project really pulls. It also shows why capacity must become cash to matter: the order-intake chapter's base case releases 105 hours a month, valued at $3,675 of capacity, but that becomes an operating-expense reduction only if the distributor avoids overtime, temporary labor, or a hire. The order-intake economics and the economics chapter explain the distinction.
Working capital. Suppose better replenishment lowered average inventory by 5% without reducing service, a hypothesis the forecasting chapter shows how to test. Taking year-end inventory as a stand-in for average inventory in this illustration, that releases about $1.16 million of working capital once ($23.2M × 5%). At the forecasting chapter's assumed 20% annual carrying rate, the recurring effect is about $232,000 a year ($1.16M × 20%). Those are two different quantities: a one-time cash release and an annual carrying value, part of which is opportunity cost rather than cash expense. Counting the $1.16 million as an annual saving would overstate the result about fivefold. The forecasting economics develop this with a sensitivity table.
Gross margin and revenue. Fewer order errors may reduce credits, returns, and reshipments; better catalog data may affect search, order conversion, or returns. Those effects touch gross margin and revenue, but none has attributable measurement in this guide. The catalog and order-intake chapters record them as hypotheses and assign them no value. Revenue preserved during a stockout is not the same as profit; delivery and procurement costs still apply.
Double counting. Labor saved correcting an order error is already inside assisted handling time; do not count it again as avoided error cost. An inventory reduction that came from accepting more stockouts is a policy change, not a forecasting gain. A return avoided may already be reflected in lower handling effort. State the counterfactual for each benefit and count each improvement once.
The financial reporting location of any cost or benefit is company-specific. Finance should classify costs under the applicable reporting standards, transaction facts, and company policy; the finance leaders guide and the economics chapter keep that question separate from the operating analysis.
What is measured about adoption and outcomes
For your analyst
Adoption. The Census Bureau's Business Trends and Outlook Survey (BTOS) asks employer businesses every two weeks: "In the last two weeks, did this business use Artificial Intelligence (AI) in any of its business functions?" In the collection period of September 21 to October 4, 2026, covering September 7 to 20, 2026, the estimate for wholesale trade (NAICS 42) was 21.6% answering yes, with a standard error of 1.46 points, against 23.8% for all sectors. For the same period, 25.6% of wholesale businesses said they expected to use AI in the next six months. BTOS sector data, published October 8, 2026
Read these figures narrowly:
- They are shares of businesses, weighted to represent employer firms, not shares of employment or sales. A small wholesaler and a national distributor count equally. BTOS methodology, version 5
- Multi-unit companies operating in several sectors or states are reported in a separate category, not under NAICS 42, so large diversified distributors may not be in the wholesale estimate.
- The question was reworded in November 2025 from AI use "in producing goods or services" to use "in any business function," so earlier and later estimates are not directly comparable. Census Bureau, May 2026
- "Use in any business function" covers anything from a sales team drafting emails to a production forecasting system. It says nothing about which workflow, at what authority, or with what result.
- BTOS is an experimental data product with its own response-rate and nonresponse limits, described in its methodology.
Share of businesses using AI, by sector
- This guide’s sector
- Other sectors in the guide
- All businesses
Show the data
| Group | Estimate | Standard error | 90% interval |
|---|---|---|---|
| Information | 50.2% | 1.43 pts | 47.8–52.6% |
| Finance and insurance | 40.1% | 1.86 pts | 37.0–43.2% |
| Professional, scientific, and technical services | 42.7% | 1.01 pts | 41.0–44.4% |
| Health care and social assistance | 24.8% | 1.01 pts | 23.1–26.5% |
| Wholesale trade | 21.6% | 1.46 pts | 19.2–24.0% |
| All employer businesses | 23.8% | 0.41 pts | 23.1–24.5% |
Outcomes. We found no attributable, independently documented measurement of AI effects on distributor order accuracy, exception rates, inventory levels, service levels, or margins, and this guide reports none. Vendor case studies exist in this market; the Field Guide does not use vendor-reported results without labeling them as such, and none is cited here. The workflow chapters describe how a distributor could measure these outcomes in its own operation.
Who is involved
| Role | Owns | Guide |
|---|---|---|
| Operations leader | Process performance end to end: order accuracy, exception queues, fulfillment, service levels | Operations leaders |
| Finance leader | Business cases, realization of capacity into cash, working capital, cost classification, payment and posting authority | Finance leaders |
| Sales operations or inside-sales leader | Customer records, quoting data, and what reps may commit | Sales operations leaders |
| Order-entry and customer-service staff | Entry, exception handling, and customer clarification | Covered within the order-intake chapter |
| Product-data stewards | Approved catalog values, safety and compliance attributes | Covered within the catalog chapter |
| Purchasing and planning | Supplier orders, replenishment policy, overrides | Covered within the forecasting chapter |
| Engineering or IT | Interfaces, duplicate prevention, traceability, fallbacks | Engineering leaders |
Order operations owns entry and exception handling. Product-data staff own approved catalog values. Purchasing owns supplier and replenishment decisions. Engineering owns reliable interfaces and traceability. An AI supplier may provide technology, but it cannot resolve unclear authority within the distributor. Purchasing managers, category managers, warehouse managers, and credit managers do not yet have role guides.
Expertise specific to this industry. A general AI builder lacks distribution data knowledge: the item master, customer part-number cross-references, units of measure and packaging levels under GS1's GTIN rules, and the EDI mapping that structured orders already depend on, which ordinary integration and rules handle better than a model. It also lacks local knowledge of whether the ERP honors duplicate-prevention identifiers and who holds authority over prices, credit, and substitutions. This guide places that expertise in-house, with product-data stewards, order operations, and purchasing, treats the AI supplier as technology only, and sends price-agreement confidentiality and product-safety requirements to counsel and qualified advisers. See Staffing an AI pilot and Choosing and working with an AI partner.
Sequencing: a first move and what follows
The following is a fictional, proposed sequence. Tarnwick Supply is an invented regional distributor of food-service disposables and janitorial products with two customer segments: a few large institutional accounts that order by EDI, and roughly 1,800 small restaurants, schools, and offices that email orders as PDFs, spreadsheets, and photographed order forms. It carries about 9,000 active SKUs from 140 suppliers.
First: measure and repair the item master. Before choosing any AI project, Tarnwick samples a month of emailed orders and a list of recently added items. The invented findings: most intake corrections trace to customer shorthand that maps to no SKU, and many new items lack case quantities. The first investment is a cross-reference and pack-quantity cleanup with named owners, which no model is needed to perform.
Second: catalog enrichment for new and changed items. With owners and a schema in place, Tarnwick runs the catalog enrichment workflow on new-item setup from supplier spec sheets, with a steward approving every record and safety attributes left to their owner. Tarnwick's small accounts order recently added items heavily, so missing pack and unit data would surface as intake exceptions the pilot could not resolve; that is why enrichment comes before intake here. A distributor whose intake exceptions come mainly from part-number mapping could reverse the order, as the operations leaders guide shows.
Third: document intake for the emailed segment only. Tarnwick pilots order intake for small accounts, staged with human approval, one customer group at a time, with duplicate prevention tested before any write to the order system. The EDI segment is excluded.
Later, if the data supports it: forecasting. Tarnwick first starts recording stockout periods and substitutions, because its sales history understates demand for items that ran out. Only after several months of reliable availability records does it consider a forecasting trial on a bounded item family, compared against its existing reorder rule.
Counterexample: when to wait. Suppose Tarnwick's measurements had shown instead that 85% of order lines arrive by EDI and the emailed remainder is a few dozen orders a day, and that its item master has no owner and conflicting pack quantities across locations. Then document intake would address little volume, and every AI workflow would inherit the master-data errors. The right call would be to fund master-data ownership and ordinary EDI exception handling, and to revisit AI when the remaining unstructured volume and the data justify it.
Questions for operators, investors, and vendors
- What share of order lines arrives by EDI or portal, and what share as unstructured documents? Which customers account for the unstructured volume?
- Who owns the item master and customer cross-references, and how are changes approved and scoped?
- Where does order-entry labor go today: keying, clarification, correcting mappings, or resolving credit and allocation holds?
- Does the claimed benefit reduce operating expense, improve gross margin, or release working capital? Which paid expense or balance changes, and when?
- Are stockouts and substitutions recorded, so that sales history can be distinguished from demand?
- How does the proposed system prevent duplicate orders and purchase orders on retry, and has that been tested against this ERP?
- What authority would the system have: propose, stage for approval, or write? Who can change that?
- How is performance reported: per document, per order, per line, and per critical field, with denominators and including rejected outputs?
- For a vendor claim, what was the customer population, the baseline, and the measurement period, and was the result independently verified?
- How will the process run if the system is unavailable during a seasonal peak?
When to wait
Defer if master data is inconsistent and unowned, transactions cannot be reversed or reconciled, most volume already arrives structured, or the remaining unstructured volume is too small to justify maintenance. Defer forecasting if stock availability and substitutions are not recorded. Structured customer ordering, better order forms, a customer portal, EDI onboarding for large accounts, and master-data cleanup may remove more work with fewer uncertainties. The Field Guide treats those as legitimate outcomes of the analysis, not as failures to adopt AI.
Evidence still needed and related reading
Attributable operator accounts are missing for every claim about how distributors run these processes: the share of unstructured orders by segment, exception rates and their causes, master-data ownership arrangements, and how stockouts are recorded. No independently documented outcome measurement of AI in distribution has been identified. More recent sector financial ratios will come from the Annual Integrated Economic Survey, whose wholesale detail should be checked against the 2022 figures used here before this chapter is updated.
Continue with purchase-order intake, inventory forecasting, catalog enrichment, operations leaders, finance leaders, and the economics chapter.