AI integration
Buying AI document intake software for mixed forms
Test AI document intake with scanned forms, attachments and corrections. Give reviewers field-level evidence and keep duplicate submissions out of the queue.

An emailed application can contain a typed form, a handwritten amendment and a photograph of the same form. For a New Zealand permits office or membership organisation, the job starts with deciding what arrived and which details a person can trust. Define the proposed record, the evidence behind each field and the exception queue before choosing a document model.
Which fields are worth extracting?
Start at the desk where someone keys the application. Watch which fields they copy, which they derive and which they deliberately leave blank. A printed job reference is different from a category someone assigns after reading the description. Keep those jobs separate in the specification so reviewers can recognise an inference.
Write an example using an application date, site address and reference. A handwritten amendment changes the site address but repeats the original reference. Specify who decides whether the amendment supersedes the typed value. Show the reviewer both values beside their page locations. A blank address needs a visible missing-field state and a route back to the applicant.
Where does a confidence score help?
Use the score to prioritise review after measuring performance on your own material. Microsoft’s document evaluation guidance describes scores between zero and one for extracted words and key-value mappings. The same score can carry very different consequences for a misspelt street description and a single incorrect reference digit that sends the file to another case.
Build a test folder with clean originals, rotated scans, duplicate pages and handwriting over printed fields. Include local place names, Māori macrons and addresses with unit numbers. Have two staff resolve ambiguous entries before the demonstration. Score correct, incorrect, absent and referred fields separately. Ask how long it takes to correct a referred field with the source image on screen.
Look at the shape of the queue as well as field accuracy. In a sample batch of ten applications with four suggested fields each, six wrong fields could be concentrated in two difficult forms or scattered across six otherwise usable forms. The number of people interrupted and the time to clear the batch will differ.
Can a retry create a second case?
Stage a connection failure after the destination records the application but before intake receives confirmation. Resend the same item. Intake should reconcile the existing reference, preserve the receipt and show staff a single case. Then submit a genuine correction. Staff must be able to link that version without quietly overwriting the original evidence.
Agree which originals, extracted fields and temporary copies the supplier retains. Trace one synthetic example through the inbox, processing service, review queue, destination and diagnostic logs. The Office of the Privacy Commissioner recommends assessing privacy impacts before using generative AI. At handover, get a field dictionary, sample correction history and a replayable test folder. A buyer can use those materials to decide whether a new form version still belongs in automatic intake.