Skip to content

CPD lab · AP, Purchasing

Purchase three-way match

A fortnight of purchase orders, goods received notes and supplier invoices for a fabricator: match every invoice line, apply the tolerance policy with your AI tool, and decide what is paid now.

All the ways to get this lab

Comes with Raising and approving spend: procure-to-pay for requesters, and with the year. 29 labs open now.

The brief

You work in accounts payable at Aldwick Fabrication Ltd, a metal fabricator. It is Friday 13 November 2026 and the payment run is being prepared. The pack covers purchase orders, goods received notes and supplier invoices for the last fortnight. Every company, person and figure is invented, so you can paste it into any AI tool.

The files are the purchase order lines, the goods received notes (GRNs), the supplier invoice lines (15, all with VAT at 20%) and the register of budget holder sign-offs for the two service orders. Nothing on these orders has been invoiced or paid before this fortnight.

Match each invoice line to its purchase order (PO) line and to the goods received against it. Then decide what to do with the line, under the policy below. Pay nothing yourself: your job is the matching table, for a person to approve.

Policy, applied in this order. (1) Duplicate: if an earlier invoice, by invoice date, has already been matched to the same PO line and the same goods, the later invoice is a duplicate. (2) PO: the PO must be dated on or before the invoice date. (3) Price: the variance is (invoice unit price minus PO unit price) times the quantity invoiced. Only a price above the PO counts. The tolerance is the greater of 2% of the PO value of the quantity invoiced (PO unit price times quantity invoiced) and 25.00. A variance within tolerance is paid at the invoice price. (4) Quantity: for goods, the quantity payable is the lower of the quantity invoiced and the quantity received on all GRNs for that PO line, less any quantity already matched to an earlier invoice. (5) Services have no GRN: the evidence is the budget holder's written sign-off in the register.

Decision, one per invoice line. PAY: the whole line is supported, so it is paid in full. HOLD: the line is not yet fully supported for a reason that will clear without the supplier changing the invoice (goods or sign-off still to come). Pay only the quantity that is supported, which may be nothing, and hold the rest. QUERY: the line is wrong or unapproved (price outside tolerance, or no PO in place on the invoice date). Nothing on the line is paid until Purchasing or the supplier answers. REJECT: the invoice should not be paid at all and goes back to the supplier (a duplicate).

Give one line per invoice line, with the columns in the result table. matched_qty is the quantity payable now (for a service, 1 if paid, 0 if not). amount_payable_now is what you would pay now for the line, including VAT at 20%, to the penny, at the invoice price. It is 0.00 for QUERY, REJECT and for a HOLD with nothing supported. reason is a short sentence for the reviewer.

A goods received note with no invoice is not an invoice line, so it is not in the table. Mention it in your note.

Then write a short review note: what the AI got wrong, what you checked, and what goes to a person.

Objective: To be able to use an AI assistant to match supplier invoice lines to purchase orders and goods received under a stated tolerance policy, deciding what is paid now and what is held, queried or rejected, and checking its work before anything is paid.

What's in the pack

Every name, figure and email in the pack is invented. Use your own AI tool: Claude, ChatGPT, Copilot or Gemini.

  • purchase-orders.csv

    Purchase order lines raised 28 October to 5 November 2026 (15 lines)

  • goods-received.csv

    Goods received notes, 2-11 November 2026 (14 lines)

  • supplier-invoices.csv

    Supplier invoice lines received 2-13 November 2026 (15 lines, all VAT at 20%)

  • service-sign-offs.txt

    Budget holder sign-offs for the two service orders, at 13 November 2026

The traps

The pack has 7 deliberate traps: cases where AI tools often get it wrong and a person must check. We don't list them here. Finding them is the job, and your feedback names each one you missed.

What you submit

  1. Your result, as a table with the columns invoice_ref, line, matched_qty, decision, amount_payable_now, reason.
  2. The prompt you used.
  3. A short review note, 150 to 350 words: what the AI got wrong, what you checked, and what goes to a person.

How it's marked

  • Your result, checked line by line against the answer key60%
  • Your review note, marked on traits40%

The review note's traits:

  • Quantities matched to goods received
  • Tolerance applied as written
  • PO and receipt trail tested
  • Goods received not invoiced noted
  • What goes to a person, and data kept safe

Pass at 80. Under 80, read the feedback and try again. Then write your reflection and copy the entry for your CPD record.

Price

  • Raising and approving spend: procure-to-pay for requesters, £39. The role track and its labs, this one included. No end date.
  • Everything for a year, £199. Every course, role track and lab for 365 days.
  • Monthly, £19.99 a month, cancel any time. Every lesson, plus 8 marked tasks a month, across every course, role track and lab.

Buying for a team, or want a lab built on your own process? CPD for employers.