AI for Invoice Processing in Consumer Goods: Proving the Invoice After the Deduction

A consumer goods business has two invoice problems. The first is familiar: supplier invoices waiting to be read, coded, matched and approved. The second costs more. It is the invoice you raised against your customer, which your customer rewrote before paying it. Payables gets treated briefly here, and the deduction gets the rest.

On the sales side the customer holds the money

On the payables side you hold the money, so you control the clock. On the sales side the customer holds the money and has already spent it. The remittance arrives net of shortages, price differences, promotional claims, fines and adjustments, with a reason written by somebody in another company. Your job is not to approve an invoice. It is to prove, after the fact, that the invoice you sent was right.

A grocery retailer's own invoice processing is the engine that generates these deductions. The same mechanics, read from the other end, are yours to disprove. One side of the transaction is automated. The other is a spreadsheet and one experienced person who is permanently behind.

Payables first, because it is the smaller half

Ingredients, packaging, freight, warehousing, media, agency fees and the service invoices nobody raised an order for. An agent reads each one, codes it and routes it against whatever commercial paper exists. Three parts are specific to consumer goods. A co-packer invoice carries a conversion charge, a materials element covering what you supplied free issue, a yield line and sometimes a minimum run fee, so an agent rebuilds it from the production declaration and the bill of materials and tests declared yield against the tolling agreement. Freight invoices are tested against the rate card. And the invoice for display units built for a cancelled promotion stays hard: somebody has to decide whether the spend was committed.

Turning a remittance into a list of things to prove

Cash arrives as one amount covering many invoices, less many deductions. It may come as an EDI 820 remittance, an EDI 812 adjustment, a self billing document that is the customer's own calculation of what it owes, or a PDF in a shared mailbox. The reference is usually the customer's claim number, absent from your ledger, and the narrative is free text along the lines of "shortage ref 7 claim".

An agent reads all of those formats into one structure: cash received, invoices covered, and each deduction with its code, narrative, amount and date. It applies cash to open items, isolates the residual, then searches your records for whatever each narrative points at. The output is a worklist where every line carries a candidate invoice and a candidate cause.

Deduction codes that mean something different at every customer

Every customer has its own code set, and the same word means different things across them. A code labelled as a shortage at one account covers damage in transit at another. Codes change without notice when a customer upgrades a system, and the most expensive deductions often arrive with no code at all, only a sentence. An agent maintains a mapping from each customer's codes and narrative patterns to your internal reason taxonomy, and keeps an explicit unmapped bucket rather than guessing. A deduction filed under the wrong reason disappears into a category nobody reviews.

Proving what left the yard when the receipt says otherwise

Quantity deductions are the highest volume and the most winnable, because the evidence either exists or it does not. An agent assembles the chain for each claimed shortage, from order and despatch note through loading record, carrier paperwork and proof of delivery, then flags whether the gap between claimed receipt and confirmed despatch is supported.

When the unit of measure changes between the order and the receipt

Your invoice is in cases. The customer counts in consumer units, outer packs, layers or pallets, using a conversion factor held in its own item file. If that factor is stale after a pack change, every delivery of the line generates a shortage that is really an arithmetic error. An agent tests the claim against both conversions, which turns a recurring claim into one data fix at the customer.

When the customer weighs it again

Where goods are sold by weight, your invoice reflects despatch weight and the customer pays on its own weighing. An agent checks the difference against the agreed tolerance and whether the drift is one directional, because a consistent drift at one depot is a conversation, not a credit.

Which price applied, and on which date

Price deductions turn on effective dates. The customer compares your invoice to the price it believes applied on the receipt date, using the file it loaded and the calendar it holds, so an order placed before a change and received after it fails that test even when your invoice is right. An agent reconstructs the price history, identifies which price applied at order and at receipt, and names the document that supports it: the confirmed price file, the promotional agreement or the standing trade terms. Narratives read like "price diff inv 4417" and resolve three ways. Your error, their stale file, or a disagreement about which date governs.

Testing a promotional claim against what was signed

A promotional claim is a deduction for money you agreed to fund, arriving separately from the sale it funded and often months later. To test it an agent matches four things: the agreement, the mechanic, the period, and the volume that actually sold or shipped under it. It then recalculates and reports the gap, which usually has one of four causes. The promotion ran outside the agreed window, the mechanic claimed is richer than the one signed, the volume includes stores never in scope, or the same support was claimed twice, once as a deduction and once as an invoice.

Fines, penalties and the appeal window that closes quietly

Service level charges each carry their own evidence requirement and their own deadline. A late delivery or missed booking slot needs the booking record, the gate time and the carrier timeline. A labelling or packaging compliance charge needs the artwork approval and the specification in force on the production date. A fill rate penalty needs the order and the agreed allocation, which is where an allocation you were instructed to make resurfaces as a penalty. An agent pulls the evidence per charge type, checks the measurement period and the measure against the trade terms, and watches the appeal window. Deadline tracking matters more than accuracy here: an arguable deduction never appealed in time is money gone.

Claims that arrive without an agreement behind them

Returns and damage claims are often aggregated, poorly referenced and raised long after delivery, so an agent groups them by line, depot and period and tests whether the volume claimed is plausible against what shipped. The one to watch is the goodwill deduction, taken because a store marked stock down, with nothing agreed behind it. Distributors claim differently again: on a price decrease they claim protection on stock held, resting on a declaration you cannot verify, so an agent reconciles the declared holding against your shipment history.

Deciding what to dispute, and what the evidence pack must contain

Not every deduction is worth challenging. A small claim needing two hours of evidence gathering and a depot phone call is a loss even when you win. An agent scores each open item on amount, evidence completeness, historical win rate for that reason and account, and days left to the deadline. Adopt AI Partners builds the packs to the customer's own format, because a dispute rejected on format is indistinguishable from one that was wrong. Each carries the invoice, the proof, the agreement and a short written case.

The agent reads, matches, classifies, evidences, drafts and tracks. It does not decide. It never writes off a deduction, never accepts a penalty, never releases a payment and never agrees a settlement.

Where AI fails here

Some deduction narratives contain no usable information. A line reading "adjustment" with an amount and nothing else cannot be classified by any system, and pretending otherwise moves the guess upstream. Those need a call, and the goal is a better narrative at source.

Proof of delivery often proves less than people assume. A signature against a pallet count says nothing about the cases inside, and an agent treating it as conclusive builds confident disputes that collapse on first challenge.

Agreements that live in an email thread cannot be matched. Where a promotion was agreed verbally, amended in a message and never written to the trade terms, there is nothing to reconcile against. The agent can report that the evidence does not exist. It cannot manufacture it.

Which disputes to pursue is not an optimisation problem. A tense negotiation, a buyer about to review a range, a joint business plan up for renewal: all change the answer and none appear in the data.

The worst failures are the confident ones. A deduction wrongly classified as valid is written off unreviewed, and a wrongly recalculated claim becomes a dispute you lose in front of a buyer. Value thresholds, review bands and sampling of accepted items are the difference between a usable system and an expensive one.

Start with the deductions nobody disputes

Pick the customer with the highest deduction volume and pull a full year of codes. There will be a cluster of small, repetitive, same reason claims nobody ever challenged because nobody had the hours. Have an agent classify that population, match it to evidence and produce the packs, with a person approving every submission. Measure recovery rate and time per claim against your baseline, then widen by reason code rather than by customer. The first win is rarely the hard disputes. It is the ones quietly conceded. To talk through where your own deduction population sits, email hello@adoptaipartners.com.