Most demos of "AI for procurement" show you the happy path. A purchase order goes out, a clean confirmation comes back, the numbers match, everyone smiles. It looks effortless because the example was chosen to be effortless.
Real purchasing departments don't live on the happy path. They live in the exceptions — the confirmation that comes back €4 light, the supplier who ships from a warehouse in a different country than the one on file, the item that's out of stock the moment the order is placed. That's where buyers actually spend their day, and it's where automation either proves itself or quietly falls apart.
We spent the last week deep inside exactly these problems while building Lisa AI, our procurement agent. Here's what the work surfaced — not as theory, but as the kind of detail that decides whether automation survives contact with a real supply chain.
A confirmation that "almost" matches is the whole problem
When a supplier confirms a purchase order, the interesting question is never "do the numbers match?" They rarely match exactly. The real question is: does this difference matter, and what caused it?
That sounds simple until you try to write it down. A €2 difference on a €5,000 line is noise. The same €2 on a €10 line is a different conversation. So tolerance can't be a flat number — it has to flex with the value at stake. And once a deviation clears the tolerance, the next question isn't whether to flag it but how loudly: a 5% variance and a 60% variance are not the same alert, and treating them the same is how buyers learn to ignore alerts altogether.
The genuinely hard part is the layering. A confirmation can be off at the line level but fine at the order level once surcharges net out. It can carry the right total across the wrong number of line items. It can come back cheaper, which is "good" for the invoice and still a signal that something in the order was misunderstood. Useful automation has to reason across all of these at once and come back not just with "exception" but with a probable cause. The value isn't in catching the mismatch. It's in saving the buyer the twenty minutes of figuring out why.
The data you need is usually the data that's missing
Here's a problem that doesn't show up in any product brochure: to estimate when goods will actually arrive, you need to know where they ship from. And that field is almost always empty.
A purchase order will faithfully record the supplier's registered business address. But a company headquartered in Munich might ship from a warehouse in Poland or a factory in Asia. The "ship-from" that determines transit time typically doesn't materialize until the advance shipping notice arrives — long after you needed the estimate to set expectations with the requester.
The instinct of a brittle system is to give up here, or worse, to quietly use the wrong field and produce a confident, wrong answer. The instinct of a useful one is to reason with proxies: fall back to the supplier's country as a reasonable approximation, pair it with the delivery destination to estimate transit realistically, and — critically — upgrade the estimate the moment better information appears in a confirmation note or origin field. Good procurement automation is not the system that has perfect data. It's the system that degrades gracefully when it doesn't, and gets sharper as reality fills in.
Relationships are the data your systems don't hold
Most companies have invested heavily in product information and master data. Those systems are very good at attributes — what a product is, what it costs, how it's classified. They are surprisingly poor at relationships — which product substitutes for which, which items are typically bought together, which SKUs were recalled alongside others, which private-label line competes with which brand.
This gap becomes expensive at the worst possible moment: a stockout. When a customer's chosen item is out of stock, the classic outcomes are all bad — backorder and risk the cancellation, have a salesperson manually ring around for an alternative, or simply lose the line. The lever that changes the math is a vetted, ranked list of acceptable substitutes available at the moment the order is placed, not assembled afterward. That requires holding "this substitutes for that" as an explicit, reasoned relationship — something attribute-based systems were never built to do.
The honest caveat matters here too: the substitutions that matter commercially are not always the equivalences a regulator recognizes, and in regulated categories a customer may refuse a swap on liability grounds. Automation that pretends those constraints away isn't helpful; automation that encodes them is. The point isn't to replace the category manager's judgment — it's to make that judgment available at machine speed, at the moment it's worth money.
Regulation keeps moving the ground underneath
None of this sits still. In medical device distribution, for example, the EU's EUDAMED database becomes mandatory for its first modules on 28 May 2026, with legacy devices following later in the year. That single shift changes the master-data layer underneath every procurement workflow it touches.
The lesson generalizes well beyond one regulation: the data foundation under procurement is not a fixed thing you automate once. It moves. Systems built as rigid, hard-coded pipelines tend to break on exactly these changes. Systems built to reason — to treat rules as inputs they can adapt rather than logic welded into the plumbing — bend instead of break. When you're choosing how to automate, that flexibility is not a nice-to-have. It's the difference between a tool that ages well and one you're re-buying in two years.
The unglamorous conclusion
The thread running through all of this is the same: the value of AI in procurement is not in the cases that were already easy. It's in the messy, ambiguous, incomplete, regulation-shadowed cases that consume a buyer's actual day — the deviation that needs a cause, the missing field that needs a sensible proxy, the stockout that needs a substitute, the rule that just changed.
That's a less flashy promise than "fully autonomous purchasing." It's also the one that holds up on a Tuesday afternoon when the confirmation comes back €4 light and nobody has twenty minutes to spare. That's the bar we hold Lisa AI to — and, more and more, the bar we think the whole category should be measured against.
Lisa AI is Recall Space's procurement agent for purchasing automation, material planning, and ERP integration. If your team spends its day in the exceptions, we'd like to hear about it.
