Manual vs automated exception resolution: analyst, rules engine, AI agent

Manual vs automated exception resolution names the choice between three ways of clearing an invoice that has failed matching: an analyst investigates it by hand, a workflow tool routes it to an analyst who investigates it by hand, or an AI agent runs the investigation itself. The paths differ in one place, who establishes the facts, and the cleanest comparison takes a single failed invoice down each path in turn.

This wiki defines procure-to-pay terms for operations where AI agents carry the operational load, reading receipts and chasing discrepancies, while people hold on to the judgment calls. Abstract comparisons of manual and automated resolution flatter whatever sits in the middle; one concrete case shows where each kind of automation actually stops.

One shorted delivery is enough to run the comparison

Take the case. A packaging supplier invoices $3,860 against a purchase order in SAP: ten pallets of stretch-wrap film at $386 a pallet. The PO says ten. The goods receipt, the warehouse's posting that material physically arrived, says eight. The three-way match, the line-level comparison of invoice against PO against receipt, fails on quantity: two pallets billed and unreceived, a $772 variance far outside any quantity tolerance, the variance band a company pays without review (see match tolerances and thresholds). The invoice drops out of touchless processing and becomes an invoice exception, specifically a quantity and receipt mismatch, with hold code Q1 and the payment clock still running.

What actually happened is mundane. The carrier split the shipment. Eight pallets arrived on the first truck Monday morning; the remaining two arrived on a second truck Thursday and were receipted then. Every path below starts from that same Monday failure; the paths differ in who assembles the facts, and when.

The manual path runs on queue position and email

Worked manually, resolution starts with triage. The AP analyst who owns this hold type opens the queue Monday with ninety-odd exceptions in it and sorts by dollar value, the rational way to protect discounts and month-end accruals when there is more queue than day. A $3,860 invoice sits below the fold. It gets opened Wednesday.

The analyst then does the lookups: the PO in SAP, the goods receipt history against it, the invoice image. The records show eight of ten received, which confirms the mismatch without explaining it. The explanation lives outside the ERP, so the analyst emails the buyer: did the supplier short-ship us, or is a delivery still in transit? The buyer is in meetings, forwards it to receiving Thursday, and the answer lands Friday morning: second truck arrived Thursday, receipt posted, all ten in. The analyst re-runs the match Friday afternoon and the invoice posts.

Elapsed: five business days. Actual investigation: perhaps twenty-five minutes, across four analyst touches plus one each from the buyer and a receiving clerk. Nobody was slow; the delay was structural, with facts in three places and a person carrying messages between them.

A rules engine moves the exception and leaves the investigation alone

Now run the same invoice through a rules-based workflow tool, the usual first layer of automation bolted onto the ERP. The tool reads the hold code, applies its routing table, and by 9:05 Monday the exception sits in the right specialist's worklist: packaging category, quantity hold, invoice PDF attached, a 48-hour SLA timer running.

That routing is worth something: the two days behind bigger dollar amounts are gone, and the timer keeps the hold from quietly aging past month-end. But look at what arrives in the worklist: the same invoice, the same hold code, the same attachment. The specialist still opens SAP, still finds eight of ten, still emails the buyer, still waits on receiving. Rules route on fields the matching engine already produced; they cannot read a shipping notice or put a question to a warehouse. The investigation, where the elapsed time and cost actually live, is untouched. Elapsed: three to four days instead of five, with an identical set of human touches. The context problem in exception resolution covers why that investigation resists being written as rules.

An agent runs the investigation, then waits for the truck

An AI agent picks the exception up minutes after the match fails and starts where the analyst would have started two days later: PO, receipt history, invoice image. Then it keeps going, into records no routing rule can parse. The supplier's ASN, the advance ship notice sent when goods leave the dock, shows the order left in two consignments, eight pallets and two, the second due three days out. The packing slip on the first goods receipt confirms eight pallets on that truck. Within an hour the picture is complete: nothing is lost, the goods are on the road, and the correct action is to wait for the second receipt.

So the agent waits, with intent. It marks the case as a split shipment pending receipt, holds payment action, and watches the PO. Thursday, receiving posts the second goods receipt. The agent re-runs the match within the hour, all ten reconcile, and the invoice posts with the trail attached: ASN, both packing slips, receipt history, and a note that the variance was transit timing. No email was sent and no one was interrupted. Elapsed: three days, every hour of it a truck on a road.

What manual vs automated exception resolution costs on one invoice

Put the three runs side by side:

1. Manual: five business days elapsed, six human touches across three people, twenty-five minutes of investigation diluted into a week. Multiply by ninety holds in that Monday queue and the arithmetic becomes the AP backlog.
2. Rules-based: three to four days elapsed, the same six touches. The tool bought queue time and visibility and paid nothing toward the investigation.
3. Agent: three days elapsed, zero human touches, minutes of machine work. Elapsed time equals the physical constraint, with nothing stacked on top.

Manual and rules-based resolution add coordination cost, the waiting and forwarding and re-asking, on top of what the physical world requires. Spread across thousands of holds a month, that layer is most of what the cost of invoice exceptions page tallies, from missed early-payment discounts to accruals built on stale holds.

Each path breaks in a different place

The sharper way to compare manual vs automated exception resolution is by failure mode. Manual resolution breaks on volume: one analyst clears a few dozen investigations a day, the queue grows whenever volume rises, and triage by dollar value quietly starves every small hold. Rules break on context. A routing table could never have read the ASN, and a shortcut rule like "auto-clear quantity holds under $1,000" would have paid this supplier on Monday for two pallets that, at the moment the rule fired, were still on a truck.

Agents break on genuine judgment. Change one fact in the case: the ASN shows a single consignment of ten pallets, the packing slip says ten, the dock count says eight, and the supplier insists everything shipped. Now someone must decide whether to short-pay the invoice, file a carrier claim, or accept the supplier's proof of delivery, a call no precedent fully settles. A well-built agent recognizes the boundary and escalates with the workup attached: what it checked, what each record showed, and the single open question. The person decides in minutes instead of reconstructing the file for an hour, which is what exception escalation best practices aim for and a forwarded PDF never delivers.

Fragment builds AI agents that work exceptions the way the third path describes, inside a company's existing SAP or Ariba environment: reading ASNs, receipts, and resolution history, clearing timing cases on their own, and escalating genuine disputes with the investigation already done. See how the workflows run or request a demo.