AI extraction confidence thresholds: when to require review
Set AI document extraction confidence thresholds using local tests, field-level risk and review capacity, with a reusable matrix and clear stop conditions.
Stanic & Co.
Perspectives for business leaders
Set AI document extraction confidence thresholds by field and consequence, then test them against documents your business actually receives. Use a tiered approach: lower scores require human verification; middle scores remain under review while you audit performance; higher scores may bypass extraction review only after testing. Do not treat a score above 0.95 as proof of correctness or permission to release a payment.
The practical question is not simply where to put the cut-off. It is how much incorrect data you are prepared to let through, how you will detect it, and whether your team can handle the resulting review queue. The process and numbers below are illustrative recommendations, not thresholds prescribed by government guidance or evidence about any particular product.
Set AI extraction confidence thresholds around the decision
Start by defining what crossing the threshold permits. Saving an extracted invoice date into a draft record is different from posting an invoice to the ledger, changing supplier bank details or releasing payment. For this article, “auto-approve” means accepting an extracted field without routine extraction review. It does not mean authorising the downstream transaction.
The Australian Government’s practical AI use guidance describes flagging low-confidence extractions for staff to check and correct. It also recommends starting with one document type, defining the fields and validation rules, and testing accuracy before rollout. It does not supply a universal numerical threshold. Your cut-off therefore needs local evidence rather than a borrowed industry benchmark.
Before testing, ask the provider what its score represents: a field, a word, a whole document, or something else. Ask whether it has evidence connecting score bands with correct extraction for your intended documents. Unless that relationship is demonstrated, do not interpret 0.95 as a 95% probability that the field is correct. If the system provides only a document-level score, do not assume it supports field-level release decisions.
Set separate consequences for separate fields. A reference description may be eligible for limited automatic acceptance, while a bank-account change remains subject to independent verification regardless of score. This complements the broader question of which extracted fields need a human check: here, the decision is how to test a numerical routing rule after those risk boundaries are established.
Test candidate cut-offs against known answers
Use an evaluation batch before allowing any score band to bypass review. As an illustrative starting point, select 300 historical documents covering the suppliers, layouts, scan quality and exceptions you expect to process. This is a practical exercise size, not a statistically sufficient sample for every risk. Keep a separate group of documents untouched until you have selected a candidate threshold, so the final check is not merely a retest of examples used to tune it.
Have a trained reviewer establish the correct field values from the source documents. Resolve ambiguous originals through your normal business process rather than forcing a supposedly correct answer. Record the document type, layout, field, extracted value, score, verified value and whether the difference would affect a business action. Include missing fields: a test that records only returned values can overlook an important failure category.
For each candidate cut-off, calculate two quantities: bypass volume = fields accepted without routine extraction review ÷ all evaluated fields; observed error rate in the bypass group = incorrect accepted fields ÷ accepted fields checked. Record the counts as well as the percentages. Ten correct results out of ten do not provide the same evidence as a much larger checked group, and neither establishes a guarantee about future documents.
Define pass conditions before reading the results. An illustrative pilot could require no observed material amount errors in the proposed bypass group, all validation failures successfully held, and a review queue that can be cleared before the processing deadline. These are screening conditions, not proof of safety. If material errors appear, retain review for that field, narrow the eligible document group or improve the input before testing again.
Reusable confidence-to-review matrix
Copy this record for each document-type and field combination. It is an illustrative routing template, not a source-prescribed rule. Inputs: [Document type] [Layout or supplier group] [Extracted field] [AI confidence score, 0.0–1.0] [Consequence of error] [Validation results] [Test evidence] [Named owner]. If the product uses another score scale, establish its meaning rather than mechanically converting it to this one.
Band A — score below 0.85: manual review required before accepting the field. Band B — score from 0.85 through 0.95 inclusive: sample-audit candidate; during initial testing, review every field before release. Permit unsampled fields to bypass review only after a separately approved evaluation. Band C — score above 0.95: auto-approval candidate, subject to the evidence and overrides below. These boundaries deliberately include scores exactly equal to 0.85 and 0.95.
Example entries: [Supplier invoice] [Established layout] [Invoice date] [0.97] [Draft-record error] [Date checks pass] [Approved test evidence] → eligible for limited automatic acceptance. [Supplier invoice] [Established layout] [Subtotal] [0.91] [Incorrect posting] [Arithmetic checks pass] [No approved bypass evidence] → review before release; record for the middle-band audit. [Supplier invoice] [Any layout] [Changed bank details] [0.99] [Payment diversion] → independent verification, overriding the score.
Pass only when the field and layout are approved, the score band has supporting test evidence, required checks pass and the review owner is available. Stop automatic acceptance for missing scores, unreadable originals, changed bank details, failed arithmetic, duplicate indicators or an untested layout. Keep affected records on hold; never make an exception silently pass because the queue is busy. Record the override reason so it can be distinguished from an extraction error.
Calculate review capacity before choosing the faster setting
A cut-off that sends nearly everything to a person may remove much of the operational benefit. A lower cut-off may clear the queue while admitting errors you have not measured. Compare candidate settings using observed routing volumes and review time, rather than choosing the percentage of automation you would like to report. Measure document-level effort too: checking several fields together may take a different amount of time from checking each separately.
Consider an explicitly illustrative workload of 500 invoices a day, with one extraction-review task per flagged invoice. At a candidate cut-off of 0.95, suppose 150 invoices need review. At two minutes each, that requires 300 minutes, or five staff hours daily. At a candidate cut-off of 0.85, suppose only 60 need review: 120 minutes, or two hours. The apparent difference is three staff hours per day, before audits, escalations and corrections.
Now include audit capacity. If the lower setting bypasses review for 440 invoices and you inspect a rounded-up 10% sample at two minutes each, that adds 88 minutes. Total daily oversight becomes 208 minutes, about three hours and 28 minutes. The saving against the first setting is smaller than the headline routing difference. This calculation says nothing about whether either setting has an acceptable error rate; that requires the known-answer test.
Make the operations or finance process owner accountable for the risk-versus-capacity decision. The technical owner implements the routing, logs and holds; the review lead estimates staffing and escalation capacity. If the queue exceeds capacity, restrict intake, add authorised review capacity or return the affected documents to the manual process. Do not lower the threshold merely to make overdue work disappear.
Hypothetical failure: a decimal error passes the threshold
Imagine a business sets its invoice confidence threshold at 0.60 to reduce review work. An invoice for $1,250.00 is extracted as $12,500.00 with a score of 0.78. The workflow accepts the amount and, because extraction acceptance has also been treated as payment approval, the invoice is paid. Nobody identifies the error until the next quarterly audit. This is a hypothetical failure, not a client result or a claim about a specific system.
The overpayment is $11,250: $12,500 minus $1,250. A higher cut-off would have routed this particular score for review, but that does not establish a safe cut-off for other errors. The more important design failure is that a single extraction score controlled a financial action without separate checks. A numerically plausible but incorrect result still needs controls appropriate to its consequence.
In this scenario, proposed detection controls include comparing the total with subtotal and tax, matching against an approved purchase order where available, flagging an unusual amount for that supplier and keeping payment authorisation separate. Do not assume these checks always catch the problem: an arithmetic check, for example, may pass if several related amounts were extracted incorrectly. Treat each check as additional evidence, not a substitute for the full control design.
On discovery, pause automatic acceptance for the affected amount field and document group. Identify records processed since the last satisfactory check, compare them with originals, and reconcile any postings and payments. Preserve the incorrect and corrected values with their record identifiers. Refer payment recovery to the finance owner, then retest before reinstating the route. Write these steps into the operating procedure before the first live bypass.
Make review measurable without treating it as a guarantee
Specify what a reviewer must do: open the source, locate the field, compare the value, check the applicable business rule and record any correction. A confirmation button alone is not a useful definition of verification. Design the queue so the original and extracted value can be compared, and give reviewers a way to mark the source as ambiguous rather than compelling them to accept or guess.
Human review is another control to evaluate, not an assurance of correctness. In your pilot, include a second check of selected reviewed records to see whether reviewers are missing errors or disagreeing about the correct interpretation. Measure correction time and escalation time, not only the number of approvals. If reviewers cannot reliably resolve a field, more review of the same kind is not necessarily the answer; the source or workflow may need changing.
For organisations subject to the Australian Privacy Principles, the OAIC’s commercial AI guidance explains the APP 10 obligation to take reasonable steps to ensure personal information collected, used and disclosed is accurate. It also recommends ongoing performance reviews, staff training and monitoring. Ask a qualified privacy adviser how those obligations apply to your information and proposed uses; this matrix is not a compliance determination.
Assign the review lead to check a defined sample of bypassed records and report errors by field, layout and score band. As an illustrative initial rule, inspect 10% of eligible bypassed documents daily, with at least one check from each active layout where practical. That is a workload proposal, not a statistically validated assurance level. A material error should trigger a hold and investigation rather than waiting for the monthly average to deteriorate.
Choose a simpler input when thresholding is the wrong control
Do not use confidence-based bypass if the provider cannot explain the score sufficiently for your decision, you lack reliable known answers, or an incorrect accepted value would have consequences your controls cannot contain. It may also be unsuitable where document layouts change frequently or volumes are too small to maintain useful evaluation evidence. In those circumstances, AI may still prepare drafts, but the score should not determine release.
Compare extraction with a standard supplier template, structured digital form or an existing structured data exchange. A form can require an amount, currency and reference number in separate fields, avoiding the need to extract those values from a free-form page. It does not ensure that the submitter supplies correct information. Validation, identity checks and business approval remain necessary; the benefit is removing one source of interpretation, not eliminating error.
Allow for the alternative’s costs: supplier onboarding, form maintenance, exceptions and staff support. For a small set of cooperative suppliers, those costs may be preferable to maintaining score thresholds, audit sampling and an exception queue. For varied incoming documents that you cannot standardise, a restricted extraction workflow may be more practical. Compare both approaches using the same processing deadline, correction effort and consequences of error.
Finally, make the approval temporary and specific. Record the eligible fields and layouts, score bands, audit plan, accountable owner, stop conditions and rollback route. Reassess after changes to the model, extraction configuration or document population. The NCSC’s secure AI development guidelines treat security as a lifecycle requirement and include logging, monitoring and update management during operation. For this workflow, retain enough version and routing information to identify what needs rechecking. The decision to approve is a decision to operate those controls—not simply to select 0.95.
Sources
General information only; verify sources before making business decisions.