Procurement After Spreadsheets: A Thesis on the Future of Small Business Operations

Abstract

Procurement is the largest line item in most small and medium businesses and the least-tooled function in the modern SMB stack. The middle of the market — between $250-a-month inventory apps and $50,000-a-year enterprise ERPs — has been structurally empty for two decades. Five forces are now collapsing that gap: language models that can read unstructured supplier communication, POS systems that finally expose real-time consumption, statistical forecasting that handles intermittent demand, the cost of working capital under sustained high rates, and the rise of bidirectional buyer–supplier platforms. This essay sets out the first principles of what procurement software be, the four wrong models we reject, the system we are building, and the macro case for why this is a hundred-billion-dollar replatforming. We close with the open problems we are working on and the operating-system thesis for SMBs that follows.

I. The forgotten function

For a typical small business — a 12-seat café, a 3-location grocery, an e-commerce dropshipper, a cabinet shop — the largest line item on the income statement is not rent, not payroll, not marketing. It is cost of goods. In food service, ingredients run 28–35% of revenue. In retail, COGS routinely exceeds 60%. In light manufacturing, components consume 50–70% of every dollar in. The single biggest determinant of whether a small business survives the year is whether it bought the right things, in the right quantities, at the right prices, and at the right times.

And yet procurement — the discipline of buying — is the most under-instrumented function in the SMB technology stack. Accounting has QuickBooks. Payments have Stripe. Marketing has Klaviyo. Customer support has Intercom. Payroll has Gusto. Procurement has a Google Sheet, an inbox, and a phone call.

The gap is not for lack of attempted solutions. NetSuite, SAP Business One, Acumatica, Microsoft Dynamics, Sage Intacct, Epicor, Infor, Odoo, ECI, Brightpearl, Cin7, Fishbowl — all of these claim to handle procurement. None of them are deployed in any meaningful number of businesses under 50 employees. The reason is structural. Enterprise ERPs are sold by the implementation, not the seat. They require six-month onboarding cycles, a six-figure budget, and a dedicated systems analyst. The SMB owner-operator who is making purchasing decisions at 11pm after closing is not the customer. They never were.

What the SMB owner-operator has instead is what we will call the artisanal procurement stack: a recurring spreadsheet of items and quantities, an email folder of supplier replies, a calendar reminder for Tuesday's order, a set of memorized phone numbers, a WhatsApp thread per vendor, a stack of paper receipts, and a quiet anxiety that something has been forgotten. This stack works. It works for tens of millions of businesses. It is also the single largest source of operational drag in the SMB economy. Every hour the operator spends reconciling a confirmation email is an hour they are not spending on the customer.

The thesis of this essay is that the artisanal procurement stack is about to be replaced — not by the descendants of ERPs, but by a fundamentally different category of system whose primitives are workflows, language models, and bidirectional supplier graphs rather than configurable modules and journal entries. The replacement will happen during the second half of this decade. By 2030, an SMB owner who is still managing reorders by spreadsheet will look the way a 2010 founder looked who was still doing payroll by hand.

II. Why the middle was empty

The economic geography of business software has, for a generation, been bimodal. At the top, enterprise procurement is well served by Coupa, SAP Ariba, GEP, Ivalua, Jaggaer, and a long tail of source-to-pay specialists. At the bottom, Shopify apps and bookkeeping plugins offer narrow slices — Stocky for inventory, Inflow for retail, Sortly for asset tracking, Order Desk for routing, Zapier for everything else. Between the two is a chasm.

The chasm exists for three reasons.

First, distribution economics. Enterprise sales motions cannot be applied to a $600-LTV customer. The math does not work: a $50,000 implementation budget cannot be amortized across a $50-a-month subscription. Conversely, a self-serve product cannot carry the configuration burden of a procure-to-pay suite designed for Fortune 1000 supply chains. Anyone who has tried to "scale down" an ERP has discovered that the work of removing complexity is harder than the work of adding it.

Second, the supplier side. Every procurement system that has ever shipped has been built buyer-first — a tool the buyer uses to send orders. The supplier is downstream and unmanaged. They reply by email, by phone, sometimes by fax, sometimes through a portal the buyer must learn. This means the buyer-side system can never close the loop. It has no way to know whether the order was confirmed, modified, partially shipped, substituted, or invoiced — unless the buyer manually re-enters that information. So they don't. So the system rots into a glorified PDF generator within ninety days, and the operator returns to email.

Third, the data substrate. Procurement is downstream of two data sources that, until recently, were not standardized for SMBs: real-time consumption (point-of-sale data) and supplier catalogs (price lists, MOQs, pack sizes, lead times). Without consumption, you cannot calculate how fast you are using something. Without standardized supplier data, you cannot calculate what the order should be. SMBs had neither in machine-readable form. Square shipped its API in 2014. Shopify did in 2009 but only became dominant after 2017. Toast went public in 2021. The data plane is now there. The systems built on top of it are not yet.

So the middle was empty because the unit economics didn't close, the supplier side was unowned, and the data didn't exist. All three constraints have now relaxed at the same time. That is the opening.

III. The five forces, 2025–2030

1. Language models that read reality

Procurement runs on unstructured communication. A confirmation email from a produce supplier reads "got your order, blueberries are out till Friday, subbing strawberries 1:1, price up $0.50/lb, rest ships Wed." A purchase order PDF arrives as a scan with three columns and a handwritten total. A WhatsApp message says "can do 8 cases instead of 10, cool?" Until 2023, parsing this required either humans or fragile regex. In 2024, frontier language models crossed the threshold of being able to extract structured order updates from these messages with accuracy rates exceeding 95%, including from images and PDFs.

This is not a feature. It is a category-defining capability. The buyer-side system can now, for the first time, close the loop without forcing the supplier to change behavior. The system reads the email, updates the order status, adjusts the line items, recalculates the inventory ETA, and presents a digest to the buyer. The supplier did nothing. The buyer did nothing. The data is correct.

The implication is that the comparative advantage of procurement software flips. For thirty years, the way to add value was to force standardization onto suppliers (EDI, portals, punch-out catalogs). The new way to add value is to absorb the variance — meet every supplier in their preferred channel and translate it on the fly. Standardization moves from a precondition to an output.

2. POS as the consumption oracle

You cannot order intelligently if you do not know how fast you are selling. For two decades, this was the binding constraint on small-business inventory management. The point-of-sale system was a closed register that printed receipts. The procurement system was a separate spreadsheet. The two never met.

The platform shift came quietly. Square, Shopify, Toast, Clover, and Lightspeed now collectively cover the majority of SMB transactions in North America. All five expose item-level sales through APIs. This means that for the first time in the history of small business, every sold unit can be deducted from inventory in real time, and consumption rates can be computed at the SKU level on a daily basis with machine precision.

Real-time consumption is not just a better data point. It changes the math of replenishment. With weekly counts, you needed thick safety buffers because the data was stale by definition. With daily consumption, you can drive inventory closer to the demand curve without stocking out. The economic effect is large: a typical food-service operator running 2-week safety stock can compress to 5-day safety stock with daily POS sync, freeing 60% of the working capital tied up in raw materials.

3. Statistical forecasting hits the middle market

Most SMB inventory tools use a primitive heuristic called the min/max method — set a minimum, set a maximum, reorder when below minimum. This is the procurement equivalent of using a thermometer to predict next week's weather. It does not handle intermittent demand (items sold infrequently), erratic demand (items sold in clumps), lumpy demand (both), or seasonality. It does not distinguish between the cup of soup that sells thirty times a day and the specialty bitter that sells four times a month. It treats both with the same bluntness.

The forecasting literature has solved this. The Syntetos–Boylan Approximation, published in 2005, is a bias-corrected variant of Croston's method that handles intermittent demand with statistical rigor. The SBC framework (Syntetos, Boylan, Croston) classifies demand into four regimes — smooth, intermittent, erratic, lumpy — using two parameters: average demand interval (ADI) and squared coefficient of variation (CV²). For each regime, a different forecast and safety-stock model applies. This is undergraduate-level operations research. It has been deployed in Walmart's distribution centers for thirty years. It has never, until now, been deployed in a sandwich shop.

The reason is compute and consumption data, not algorithms. The math is trivial; the inputs were not. With both now available at the SMB tier, statistical forecasting becomes a default, not a luxury. A safety-stock calculation that correctly accounts for demand volatility, lead-time variability, and service-level targets can replace the operator's gut feel without any operator effort. The cost of being wrong drops by an order of magnitude.

4. The cost of being wrong

From 2009 to 2022, the cost of capital was effectively zero. Carrying excess inventory was nearly free. Operators could buffer their way out of forecasting uncertainty. Since 2022, prime rates above 7% have made every dollar tied in working capital expensive again. A $50,000 inventory float that used to cost $500 a year now costs $4,000+. For a business with 8% net margins, that's the difference between a good month and a missed one.

This is structurally important because it changes which features of a procurement system matter. In a zero-rate world, "ease of ordering" is the dominant feature — just send the PO. In a 7% world, "right amount" becomes the dominant feature. The system has to compute the order quantity that minimizes the joint cost of stockouts and carrying, not just generate a piece of paper. The math has to be right.

5. Bidirectional supplier graphs

Every previous wave of procurement software treated the supplier as an externality — an email address to send to. The supplier did the work of replying. The buyer's system absorbed the cost of integration.

The next wave inverts this. The procurement platform is, simultaneously, a buyer-side system and a supplier-side system. When a small distributor signs up to manage their inbound orders, they get a free supplier portal. The buyers of that distributor see real-time pricing, availability, and order status without any email parsing required. The same database powers both sides.

This sounds like a small architectural choice. It is, in fact, a network-effect engine. Every supplier added increases the value of the platform to every buyer. Every buyer added increases the value to every supplier. The unit-economic flywheel is identical to the one that built Stripe, Shopify, and Twilio. It compounds.

IV. First principles

If you are designing the procurement system for the next two decades, you are not building a smaller ERP. You are building something whose architectural primitives are different. Six principles follow.

Principle 1: Procurement is a workflow, not a database. The mental model of an ERP is a database with screens on top. The mental model of a procurement system should be a workflow with data underneath. The distinction matters at every layer of design. In a database-first system, the user is asked to maintain master data — suppliers, items, prices, MOQs, pack sizes — before they can do anything useful. In a workflow-first system, the user makes orders, and the system absorbs the master data as a byproduct of orders happening. The first model has 11-month onboarding times. The second has 11-minute ones.

Principle 2: AI at the edges, not the center. Language models are extraordinary parsers and analysts. They are mediocre decision-makers. The right architecture uses AI at the edges — reading emails, parsing PDFs, extracting line items from a photo of a delivery slip, answering ad-hoc questions in natural language — while the central decision (what to order, when, from whom, at what price) remains a deterministic, auditable, human-controlled act. We do not want AI agents that auto-purchase. We want AI translators that absorb the world's unstructured procurement reality and present a clean, reviewed picture to a human who clicks send.

Principle 3: Meet suppliers where they are. Every procurement system in history has tried to drag suppliers onto its platform. EDI, supplier portals, punch-out catalogs, vendor onboarding flows. This always fails for SMB suppliers, who have margins too thin to invest in platform integration and who are ordering from a buyer who represents 0.4% of their book. The system has to support every channel a supplier already uses: email, phone, WhatsApp, web portal, EDI, fax. The buyer should not know or care which one the supplier prefers.

Principle 4: SMBs configure nothing. The defining mistake of mid-market procurement software has been its insistence on configurable workflows. Every SMB workflow looks bespoke from the inside and identical from the outside. There are exactly four things SMBs do with procurement: replenish their stock, dropship their orders, build POs from a recipe, and reorder a list. Every one of them can be inferred from the data. None of them require a settings page.

Principle 5: Inventory math is statistical. Min/max thinking is a relic of paper-based inventory. The right inventory math is statistical, with explicit demand-pattern detection, service-level targets, and confidence intervals. Compute CV² and ADI for every item daily. Classify each item as smooth, intermittent, erratic, or lumpy. Apply Croston/SBA for non-smooth. Compute safety stock as z × σ × √(lead time). Calculate days of stock with explicit decay handling: I(t) = I₀ × (1−d)^t. None of this is novel science. All of it is novel for SMBs. The arbitrage is in deployment, not invention.

Principle 6: Money and movement reconcile in one place. A purchase order, an invoice, a payment, a receipt of goods, and a journal entry are five views of the same event. In an ERP, they live in five tables across three modules with four reconciliation steps. In the right system, they are one object whose state evolves. The PO becomes the invoice becomes the payment becomes the receipt becomes the COGS entry, with every step audited and visible. The accountant does not reconcile a separate ledger because there is only one ledger.

V. The four wrong models we reject

The lighter ERP. The failure mode of every ex-NetSuite product manager who tries to "do NetSuite for SMBs." Produces a system with the conceptual surface area of a real ERP but a fraction of the depth, which is the worst of both worlds. The right system is not a smaller ERP; it is a different shape.

The inventory-only tool. Stocky, Sortly, Inflow, Cin7. Inventory databases with reporting. They tell you what you have. They do not act. The operator still has to open a separate window to send the email to the supplier. Without procurement, inventory tools are spreadsheets with charts.

The procurement-only tool. Procurify, Tradogram, Precoro. Build POs nicely but have no consumption signal. The user has to know what to order before opening the tool. Half a system.

The marketplace. Faire, Mable, Pod Foods, RangeMe. Catalogs that lock the buyer into a curated supplier set. Works for new SKU discovery; doesn't work for the 90% of orders an SMB places to existing suppliers. The marketplace model also flips the economic relationship: the buyer becomes a lead, the supplier becomes the customer. We do not believe procurement software should be a marketplace.

VI. The macro thesis

How big is this opportunity? Three approaches converge on the same range.

Top-down TAM. The U.S. has 33 million small businesses; Canada has 1.2 million. Conservatively, 35% have meaningful procurement. That is 12 million businesses. At a blended ACV of $1,000/year — a mix of subscriptions, payment processing, and supplier-side revenue — that is a $12 billion ARR ceiling for the procurement layer alone in North America.

Bottom-up wallet share. A typical SMB spends 1–3% of revenue on operations software. For a $2 million revenue business, that is $20,000–$60,000/year. Procurement is plausibly 15–25% of that, or $3,000–$15,000 per business. Across 12 million businesses: $36–180 billion of potential annual spend.

Comparable replatformings. Shopify reached $7.1B revenue replatforming SMB e-commerce. Toast reached $4B replatforming restaurant POS. Square reached $5B replatforming SMB payments. The pattern is consistent: replatforming a fragmented SMB function produces a $5–15B revenue business at maturity. Procurement is structurally larger than any of these because it sits upstream of payments, inventory, and fulfillment.

The convergence point is a $5–20B ARR opportunity for the eventual category leader, supporting a $50–200B enterprise value. The high end is the supplier-side flywheel: as the platform reaches density, it absorbs adjacent functions — supplier financing, embedded payments, freight booking, predictive cash. Each is a $1–5B opportunity. Stacked, the ceiling is in the trillion-dollar enterprise-value range.

VII. Why incumbents won't pivot

QuickBooks and Xero's primary objects are journal entries. Procurement is a workflow that produces journal entries as a side effect. Inverting that hierarchy cannibalizes the bookkeeper channel that drives their distribution.

Shopify and Toast have shipped thin procurement features (Stocky, xtraCHEF) and abandoned or under-invested in both. Procurement is upstream, low-frequency, and B2B, while their muscle is downstream, high-frequency, and B2C.

NetSuite, SAP, Acumatica are implementation-led. Average sale takes 6–9 months and costs $50,000+. They cannot serve the operator who needs to send a PO tonight without breaking their economic model.

The category is not under-served because it is hard to build. It is under-served because the incumbents are structurally prevented from building it.

VIII. Open problems

Forecasting under regime change. Every statistical forecast assumes the future resembles the past. SMBs experience tariffs, supplier failures, viral demand spikes. The detection layer is hard.

Cross-supplier item identity. When a buyer orders the same product from two suppliers under different names and pack sizes, identifying it as the same product is a non-trivial entity-resolution problem at scale.

Trust UX for AI agents. A model that proposes orders is useful. A model that auto-sends them is dangerous. The interaction design that captures the value while preserving operator agency is unsolved.

Supplier financing. Net-30 is unsecured credit at zero APR. As the platform sees both sides, it can underwrite this credit. The product is obvious; the credit risk model is not.

Real-time landed cost. Freight, duties, fuel surcharges, port fees, insurance can add 8–25% to landed cost. Computing this in real time, especially cross-border, is operationally painful.

IX. The SMB operating system

Step back from procurement. The artisanal SMB stack — the spreadsheet, the email folder, the WhatsApp thread — is not unique to procurement. It exists in scheduling, payroll variance, customer follow-up, inventory reconciliation, supplier vetting. Every function of a small business is run by an operator who is holding too much in their head.

The deeper bet is that the procurement substrate — items, suppliers, lines, orders, recipes, capital — is the load-bearing substrate of the SMB operating system. POS sits on top, accounting sits next to it, payroll runs adjacent. But the central object — the daily question of what is happening in my business — is best modeled as a graph of trade relationships and the events that flow through them. Build that substrate well, and the rest of the stack becomes views into it.

This is a thesis about a category, but it is also a thesis about who small business owners get to be. Right now, the average SMB operator is a part-time procurement analyst, a part-time inventory clerk, a part-time accounts payable specialist, and a part-time supplier relationship manager. They became all of these things by accident, and they are mediocre at all of them, because no one can be excellent at four jobs simultaneously while also running the business that nominally employs them.

The thesis is that procurement, done right, gives the operator back the time to be excellent at the one job that actually matters: serving the customer. That is what we mean when we say the system runs itself. Not literally — the operator still presses the button. But the cognitive load of figuring out which button, and when, and at what quantity, and at what price, and from whom — that load is gone. It is replaced by a system that has been watching consumption all week, parsing supplier replies all night, and presenting one screen on Tuesday morning that says: here is the order. Send?

That is the future. The decade after that is the rest of the small-business operating system. We are starting with procurement because that is where the largest line item lives, and where the artisanal stack hurts the most.

If we are right about this, the prize is the operating layer of the small business economy, which is approximately 50% of GDP in most advanced economies. If we are wrong, we will have built an extremely good procurement system. Both outcomes are acceptable.