What is a change order in procurement, when agents do the work?

A change order in procurement is the formal amendment of a purchase order after it has been committed to a supplier: a revision to quantity, price, delivery date, or specification, or a cancellation of the line entirely, posted against the PO so that every downstream record validates against the current agreement. The formality exists for what comes after. The goods receipt, the supplier's invoice, and the match between them all check themselves against whatever the PO says, and the change order is how the PO keeps saying something true.

This wiki defines procurement and AP terms for companies where AI agents carry the operational load and people keep the decisions that need judgment. A conventional glossary treats the change order as a transaction type a buyer keys in on request; this page treats it as the mechanism that keeps the source-to-pay chain – everything from sourcing a supplier through paying their invoice – aligned with what was actually agreed, and asks what happens when an agent maintains that mechanism.

The change event and the change order are separate populations

Purchases change constantly. A buyer accepts a two-week delay on a phone call. A supplier's order confirmation – the acknowledgment document stating what will actually ship, at what price, and when – quietly substitutes a superseded part number. A quantity gets trimmed in a planning meeting and everyone nods. Each of those is a change event, and in most companies only a fraction of change events ever become posted change orders. The remainder live on in email threads, call notes, and individual memory while the PO of record goes on describing a purchase that no longer exists.

This is the central fact about change orders that glossaries skip: the recorded amendments and the actual changes are two different populations, and the second is always larger. A company that measures change order volume is counting the changes someone bothered to key in, the smaller and better-behaved set. The unrecorded remainder gets paid for later, indirectly, as blocked invoices, receiving confusion, and month-end accrual noise, and almost nobody tags those costs back to the change that caused them. How the gap opens line by line is the subject of how PO amendments drift from reality; the recorded half alone is enough work that change orders overwhelm shared services teams before the unrecorded half is even acknowledged.

One unrecorded change, traced through everything it touches

Take a concrete case. A buyer has PO 4500071812 open in SAP: 600 molded housings at $18.00 each, a $10,800 line, due October 6. In mid-September the supplier calls: tooling maintenance means they can commit 450 units by October 20, and they would prefer to cancel the balance over promising units they cannot build. The buyer checks the production schedule, agrees, thanks them, and jots it in a notebook. The correct next step is a change order trimming the line to 450 and moving the date; under deadline pressure, it never gets posted.

Everything downstream now validates against fiction. The MRP run – the planning calculation that schedules production around expected receipts – still sees 600 units arriving October 6 and builds a schedule on top of them. When nothing arrives that week, a shortage alert fires, an expediter spends an afternoon chasing the supplier, and the supplier patiently explains that the buyer approved the new schedule a month ago. The production plan gets reworked in a scramble that a posted amendment would have made a calm five-minute exercise back in September.

The damage continues after the goods show up. Receiving books 450 units as a partial receipt against a 600-unit line, so the PO stays open for 150 phantom units. The supplier's invoice for $8,100 matches the receipt and pays cleanly, which makes the remaining distortion invisible: $2,700 of open commitment sits on the books for goods that will never ship, overstating accruals and committed spend until a year-end open-PO review chases the ghost. Had that phone call changed the price instead of the quantity, the same silence would have surfaced sooner and louder, as a price-variance hold – one of the standard invoice exceptions – blocking payment while AP reconstructs an agreement the buyer could have recited from memory.

Each change order type touches a different downstream record

The types are worth keeping distinct, because each one corrupts a different record when it goes unposted:

• Delivery date changes feed planning. MRP schedules production and safety stock around the PO date, so a stale date produces phantom availability, false shortages, and expedites.
• Quantity changes set receiving expectations and open commitments. An untrimmed line leaves phantom open quantity that inflates accruals; an unraised one guarantees an over-receipt someone has to explain.
• Price changes are the ones AP feels. The invoice match compares invoice price to PO price within tolerance – the small variance band a company pays without review – and an unposted price change fails that match every time the supplier bills the new number.
• Specification changes and part substitutions hit receiving and quality. A dock team cannot receive a part number the PO has never heard of, and inspection criteria attached to the old spec test the wrong thing.
• Cancellations end the agreement entirely. Left unposted, the line remains an open commitment that budget owners still see, and a supplier holding an uncancelled PO can still ship against it.

Agents close the gap by reading where changes announce themselves

The reason the gap exists is mundane. Posting an amendment is transcription work for a person who already knows the outcome: the buyer got what they needed from the phone call, and the record keeping serves everyone downstream except the one person expected to do it. Deadlines win, the posting waits, and after enough waiting it quietly never happens.

An AI agent inverts the effort. Changes announce themselves in readable places: an order confirmation whose date or price disagrees with the PO, a buyer's email thread where a delay is accepted in plain language, a receiving pattern of repeated partials that implies a rescheduled balance. An agent watching those signals drafts the change order while the change is fresh – new quantity, new date, source document attached – and routes it to the owning buyer for confirmation. The buyer's job collapses from transcription to a yes or a correction, which is a job that fits inside a deadline-driven day. The full pipeline, from signal detection to posted amendment, is laid out in automating change orders.

What agents leave alone matters just as much. Whether a two-week delay is acceptable, whether a surcharge is worth paying, whether to cancel and resource elsewhere: those are the buyer's calls, and an agent that guessed at them would be making commitments on the company's behalf. The agent's claim is narrower – whatever gets decided should reach the record the same day it gets decided – and that narrow claim covers most of the downstream damage this page has walked through.

Fragment builds AI agents that do this work inside a company's existing SAP or Ariba environment: reading confirmations, mail, and receiving activity, drafting the change orders that keep POs matching reality, and routing each one to the owning buyer for a yes or a correction, with no rip and replace. See the change order workflows Fragment runs or book a demo.