In this guide 6 sections
A custom AI model is ready when it scores well on real samples it never saw in training and its mistakes cost less than the work it removes. Set a confidence cut-off. Below it, a person decides. Judge accuracy only on answers the model releases alone, because those answers reach suppliers, customers and the ledger with no second look.
Test on samples the model has never seen
Ask for a held back set. These are real documents, such as supplier invoices, scanned delivery notes and customer emails, set aside before any tuning began, so the model has never read them. A model scored on the examples it learned from will look excellent and then fail on its first day with live paperwork.
Insist on recent, messy samples. Include the crooked phone photo, the invoice with a handwritten correction, the delivery note stamped over its total and the email mixing a refund with a new order. Clean samples flatter a model. Messy ones tell the truth.
Have your own staff mark the correct answer for each sample, because their reading becomes the standard and they will live with the tool every morning.
- Samples from the last few months
- Every supplier or sender type represented
- Poor scans, photos and odd layouts included
- Correct answers marked by your own team
Score each field, not the whole document
A single accuracy figure hides the trouble. An invoice tool may read the supplier name correctly nearly every time and still misread the tax line, the unit of measure or the purchase order number. Ask for results per field, meaning invoice number, date, total, line quantity and unit of measure, each with its own count of right and wrong answers.
Ask a sorting tool for a table of confusions, which shows which categories get mistaken for which, such as refund requests landing in the delivery queue. Customers notice that long before you do.
Watch the rare cases. A category that appears in only one message out of a hundred can be the costly one, such as a legal notice or a safety complaint.
Set a confidence threshold with a person behind it
Most models attach a confidence score to each answer, so pick a cut-off and write it on the release sheet where nobody can quietly move it. Answers above it pass straight through, and answers below it go to a person with the model’s guess shown beside the source document.
Choose the line by testing. Raise it and fewer answers pass alone. Fewer are wrong, too. Lower it and the reverse holds. Plot your held back samples against several cut-offs, then read the trade-off with the person who will do the reviewing and the person who pays for mistakes.
Every week, pull a small random slice of the passed answers and read them against the source documents, because drift shows up there first. Supplier layout changes are the classic cause.
- Cut-off chosen from test results, not a guess
- Low confidence goes to a named reviewer
- Reviewer sees the source beside the answer
- Weekly spot check of passed items
Weigh a wrong answer against the work removed
Accuracy has no single pass mark. A mis-sorted ticket costs a re-route. A wrong bank account on a supplier payment costs real money and an awkward call, so ask what the worst realistic error would do in each workflow.
Count the hours the tool removes against the hours your staff spend checking, correcting and re-sending whatever it got wrong in the same week. A model that is right most of the time but needs every output rechecked by a clerk has saved nothing, however good the demonstration looked on the day.
Where a mistake is expensive, such as a payment run, a customs entry or a legal reply, keep a person as the final approver of every release. The model only prepares the draft.
Release in a narrow slice and keep the exit open
Start with one supplier group. Run the model beside the old manual process for two weeks, compare the two sets of results every morning and log each disagreement with its document. Disagreements teach you more than agreements do.
Agree the stop rule before launch. If the error rate on passed items rises past a limit you chose in advance, the tool reverts to review-everything mode until someone finds the cause.
Write down who owns the model after release, who retrains it and how new samples get added; without an owner, accuracy slowly decays and nobody notices.
AI release sign-off sheet
Single sheet, four blocks. The first lists the held back sample set and its size, and the second is a table of fields or categories with the correct and wrong counts. The third records the confidence threshold and the named reviewer, while the fourth holds the cost of a wrong answer, the stop rule and the owner after release.
Open a blank worksheet to printQuestions owners ask
What accuracy should I demand before launch?
There is no universal number, because the right figure depends on what a wrong answer costs you in rework and lost goodwill. Demand results per field on your own held back samples, then decide together whether the remaining errors are acceptable once a human check sits behind them.
How many test samples are enough?
Enough to cover every document type, every supplier layout and every rare case, with several examples of each so one lucky or unlucky sample cannot sway the result. A few hundred real items is a common starting point, though your own mix of suppliers and categories sets the true need.
Can accuracy get worse after release?
Yes, it can, and sooner than most owners expect. A new supplier, a redesigned invoice layout or changed wording in customer emails pushes the model outside what it learned, and weekly spot checks catch the slide early. Retraining on fresh samples is the repair.
Do we need a person checking forever?
Low confidence answers always need a person. For high confidence ones, you can thin out spot checks as evidence builds, but keep a standing review wherever a mistake is costly, such as payments or tax lines. Payment checks never fall to zero.