AI in procurement for businesses — controlling spend you cannot see

Ask a business owner what they spend on purchasing and you get a confident number that is wrong. Not because they are careless — because the real number is scattered across dozens of suppliers, several people with company cards, a WhatsApp group where someone says "just order it", and an approval process that exists in theory.

Procurement in most small businesses is not badly run. It is unrun — nobody owns it, so it happens everywhere at once, invisibly. That invisibility is the problem, and it is exactly the kind of problem software is good at making visible.

The three leaks

Maverick spend. Purchases that happen outside any process — someone needs something, they buy it, it gets reconciled later if at all. Each instance is reasonable. In aggregate it means you have no negotiating leverage (you cannot consolidate spend you cannot see), you miss volume you were entitled to, and you discover commitments after they are made. The money is not stolen; it is just spent without anyone deciding it should be.

No approval discipline. Approvals that are verbal, retrospective, or skipped when busy. The purchase order — if it exists — is raised after the goods arrive, which makes it a formality, not a control. By the time finance sees it, the money is committed and the leverage is gone.

No visibility into what you actually buy. Because purchases scatter across suppliers and people, nobody can answer basic questions: how much do we spend with this supplier, are we buying the same thing from three vendors at three prices, is our spend with anyone large enough to negotiate. The data exists, in fragments, and nobody has assembled it.

None of these is dramatic. Together they are a steady, invisible margin leak, and invisibility is why they persist — you cannot fix what you cannot see.

What controlling it actually looks like

Not bureaucracy. The goal is not to make buying a pen require three signatures — over-controlling procurement just drives people back to maverick spend to get their jobs done, which is worse. The goal is proportionate control: friction that matches consequence.

Approvals before commitment, sized to the amount. A small purchase needs a light touch or none. A large one needs a real approval, before the order, from someone accountable. The principle is the same as approval workflows people actually follow: make the routine frictionless and reserve the scrutiny for where it matters, or people route around you.

A record of what was ordered, from whom, for how much — so the questions become answerable. This is the foundation; you cannot negotiate or consolidate without it.

Three-way matching on the way in. Did what we ordered match what arrived match what we were invoiced? This is the buy-side twin of the AP discipline in AI in accounts payable, and it catches both errors and the slow overcharge.

Where AI helps with purchase orders and supplier spend

Making spend visible. Pulling scattered purchases into one picture — by supplier, by category, by person — is a data-consolidation job, and it is where the first insight comes from. Most businesses have never seen their spend assembled, and the assembly alone surfaces the "we buy this from three suppliers at three prices" problems that are pure recoverable margin.

Routing approvals intelligently. Sending the right purchase to the right approver, with context, and chasing it so it does not sit — the same reliable-follow-up value as in collections.

Catching the anomalies. A price that crept up. A supplier you are suddenly spending much more with. A duplicate order. A purchase that skipped approval. These are the leaks, and surfacing them automatically is high-value, low-glamour work nobody has time to do by hand.

Reading the documents. Quotes, order confirmations, delivery notes, invoices — extracting and matching them is exactly what document AI does, and it is what makes three-way matching feasible without a data-entry team.

Supplier choice and negotiation stay yours

Deciding what to buy and from whom. Supplier selection involves quality, reliability, relationship and risk — judgement the data does not contain. AI can tell you that you spend a lot with a supplier whose prices have risen; whether to switch depends on whether the alternative can actually deliver, which is a call, not a calculation.

Negotiating. AI can arm you with the spend data that gives you leverage. It does not sit across the table. The negotiation is yours.

Fixing a culture of maverick spend. If people buy outside the process because the process is too slow, no tool fixes that — a faster process does. The technology can make approval quick enough that following it is easier than dodging it; it cannot make a slow, annoying process attractive. Get the process right first, as we argue in the new shadow IT: control that gets in the way gets bypassed.

Assemble a year of supplier spend before buying software

Not with a procurement system. With visibility.

1. Assemble your actual spend — by supplier and category — for the last year. This alone will surprise you and will point at the biggest leaks.
2. Find the concentration. Where you spend most is where control and negotiation pay off first. Ignore the long tail of tiny purchases at the start.
3. Put light approval on the large purchases only. Do not bureaucratise the small ones; you will lose the war and the goodwill.
4. Add three-way matching on the way in, so you catch what you are actually charged.
5. Then widen the discipline as it proves its worth.

The mistake is starting with a heavy process that covers everything. You annoy everyone, they route around you, and you have less visibility than before plus a system nobody uses. Start where the money is, keep it light, and expand from evidence.

Common questions

What is maverick spend, and why does it matter if each purchase is reasonable?

Maverick spend is purchasing that happens outside any process — someone needs something, buys it, and it gets reconciled later if at all. Each instance is defensible; the aggregate is not. You have no negotiating leverage, because you cannot consolidate spend you cannot see, you miss volume you were entitled to, and you discover commitments after they are made. The money is not stolen. It is just spent without anyone deciding it should be.

Should every purchase need an approval?

No, and uniform approval tends to create the problem it is meant to solve. Over-controlling procurement drives people back to maverick spend to get their jobs done, which leaves you with less visibility than before. Aim for proportionate control: friction that matches consequence. A small purchase needs a light touch or none; a large one needs a real approval, before the order, from someone accountable.

How do I start getting control of what my business buys?

Start with visibility rather than a procurement system. Assemble your actual spend, by supplier and category, for the last year — this alone will surprise you and will point at the biggest leaks. Then find the concentration, because where you spend most is where control and negotiation pay off first. Put light approval on the large purchases only, add three-way matching on the way in, and widen from there.

What is three-way matching, and can AI handle it?

Three-way matching checks that what you ordered, what arrived and what you were invoiced all agree. It is the buy-side twin of accounts payable discipline, and it catches both outright errors and the slow overcharge. AI is what makes it feasible without a data-entry team: reading quotes, order confirmations, delivery notes and invoices and matching them up is exactly what document AI is for.

Related: AI in accounts payable and approval workflows people actually follow.

Also worth reading: managing your suppliers well.