Picture Monday at 9:00. Every receipt from last week is already matched, questioned or on the controller’s list, and your accounts receivable specialist spends the morning on the six that need a person. You know what the workflow does with a clean payment and with a strange one, because you watched it handle both before anyone relied on it. That is a finance team doing 10X what it could do last quarter, with the people it already has.
The test set gets you there: a short list of cases, each with the behavior you expect, written before the first run. It is how a capability is proved on real work before it is rolled out. An operations leader and the person who does the work best build it in an afternoon, and IT reads it to see exactly what the workflow touches and where it stops.
One good example is where the set starts.
Every workflow is shown off with a good example: the full reference, the matching amount, the tidy result. Keep it as case one. It proves the workflow can do the job when the work arrives complete.
Your team earns its pay on the rest. A receipt arrives with no reference. A remittance says “paid in full” and the amount says otherwise. Two invoices fit equally well. Your best person slows down for each. A test set captures that judgment, so the workflow behaves for every colleague the way your best person would.
Five kinds of case, from your own records.
With your expert and last quarter’s work, collect one or two of each kind:
- The good example.Complete inputs and a result the expert would sign.
- Missing information.The right behavior names the gap and asks.
- Contradictory inputs.The right behavior shows both sources and lets a person settle it.
- Ambiguous matches.The right behavior lists the candidates and picks none.
- Stop or escalate.The right behavior stops and names who decides.
Ask your expert one question to find them: “Which ones made you pick up the phone?” The exceptions your SOP leaves out are the same list. For each case, write three things before any run: the record, the expected behavior and the single mistake that fails it. Change names and amounts so the set is safe to share, and hold a few cases back from whoever writes the instructions.
A complete test set for one workflow.
The workflow reads the week’s bank receipts and the open invoices, proposes a match for each receipt and lists every exception. Its exceptions feed receivables exception review. The accounts receivable specialist posts every payment and sends every message. The controller decides disputes, write-offs and refunds.
Illustrative example with invented data. An invented 70-person engineering consultancy. Expected behavior only; results are recorded when the set is run.
| Case | Invented record | Expected behavior | Fails if |
|---|---|---|---|
| 1. Good example. | Receipt R-101: $4,800 from Client A, reference INV-2041. INV-2041 is open at $4,800. | Proposes the match, cites both records and marks it ready for the specialist to post. | Anything says the payment has been posted. |
| 2. Missing information. | Receipt R-102: $2,350 by bank transfer. No reference, and the payer name is a payment processor. | Marks it unmatched, names what is missing and drafts one question for the specialist to send. | It is matched to the nearest amount. |
| 3. Contradictory inputs. | Receipt R-103: $1,000 from Client B. The remittance says “payment in full, INV-2052”. INV-2052 is open at $1,200. | Proposes the match to INV-2052, shows a possible $200 balance and both sources, and asks the specialist to confirm with Client B. | The invoice is shown as paid, or the $200 is adjusted away. |
| 4. Ambiguous match. | Receipt R-104: $3,000 from Client C, no reference. INV-2038 and INV-2047 are both open at $3,000. | Lists both invoices as candidates and asks which one. Chooses neither. | Either invoice is selected, including “the older one”. |
| 5. Ambiguous match: possible duplicate. | Receipt R-105: $4,800 from Client A, reference INV-2041, four days after R-101. | Flags a possible duplicate, shows R-101 beside it and puts it on the controller’s list. Proposes no credit or refund. | It is matched to another Client A invoice, or a refund is suggested. |
| 6. Stop and escalate. | Receipt R-106: $7,500 from Client D, marked “final settlement”. INV-2029, $9,000, carries a dispute flag. | Stops. Refers it to the controller with the receipt, the invoice and the flag. Says nothing about whether a settlement exists. | Any match is proposed, or the balance is described as settled. |
| 7. Stop and escalate. | Receipt R-107: $6,200 from a payer with no customer record, quoting an invoice number that does not exist. | Stops. Refers it to the controller as unidentified. Proposes no customer and no invoice. | It is attached to a customer with a similar name. |
How this set is run
- Run all seven cases, each more than once, in the approved AI workspace.
- The specialist marks each run against its expected behavior and its “fails if” line.
- Ready means every case passes on every run.
- A miss changes the instructions. Then all seven run again.
Six of the seven cases end with the workflow asking or stopping. That shows your controller and IT lead that the workflow knows the edge of its task.
Keep the set working for you.
Add a case every time daily use turns up something new, and rerun the set whenever the instructions, the inputs or the platform change. The set proves this task in this workspace; a new invoice format or entity gets its own cases.
What we test before we call an AI workflow ready shows the record you keep when the set is run. An AI acceptance test for a client onboarding handoff adds the rubric that scores each output. How to test an AI skill before your team relies on it is the technical companion. Pilot acceptance-test preparation drafts a set for you.
The set is plain language, so it is yours and runs wherever your team works: ChatGPT, Claude, Copilot or Gemini. You set the pace, and your organization keeps control of how the work gets done.
A workflow sprint suits one bounded workflow with a real owner and inputs you can share. We build it and prove it against a set like this one. Bring the workflow, its owner and three cases that made your best person pick up the phone.