A confidence threshold is the score cutoff that decides how an extracted field is handled: values at or above the line are accepted automatically, and values below it route to a human. In commercial real estate document extraction, the threshold turns per-field confidence scores into an accept-or-review decision on every value read from a lease or rent roll.
How Does a Confidence Threshold Work?
A confidence threshold works by comparing each field's confidence score against a fixed cutoff and branching on the result. Fields at or above the line pass through automatically. Fields below it are flagged for human review or reprocessing. LlamaIndex describes this pattern as confidence-based routing: the score acts as a quality gate that separates trusted extractions from uncertain ones.
The threshold is set per field type, not once for the whole document. A base rent figure and a co-tenancy trigger carry different error costs, so they warrant different cutoffs. Invoice-processing guidance from the field commonly sets auto-accept in the 0.85 to 0.90 range for ordinary fields and 0.95 or above for high-stakes financial values. Production pipelines that chain classification, extraction, and validation with these thresholds report reaching 99.5% or better effective accuracy while sending only 5 to 15% of fields to a human.
Band | Condition | Action |
Auto-accept | Score at or above the threshold, validation passes | Straight-through, no review |
Review | Score below the threshold, value still usable | Route to a human to verify |
Escalate | Score far below the threshold or validation fails | Reprocess or re-extract |
Why Confidence Thresholds Matter
A confidence threshold matters because it is the single dial that trades speed against safety in an extraction workflow. Set it too low and errors pass through unchecked. Set it too high and reviewers drown in fields that were already correct. The threshold decides how much of the document a human actually touches, so it governs both accuracy and cost.
The threshold sits directly on the precision-recall tradeoff. Raising the cutoff sends more fields to review, catching more errors but flagging many correct values along the way, higher precision at the cost of throughput. Lowering it lets more fields auto-accept, which speeds processing but lets more mistakes slip through. There is no cutoff that optimizes both at once. The right number is the one where the cost of a missed error and the cost of an unnecessary review balance for that field.
Example
A confidence threshold is easiest to see when the same fields are run at two different cutoffs. An analyst extracts 100 fields from a batch of leases. The model returns a confidence score per field, and 92 of the 100 are actually correct. The table shows what happens at two thresholds.
Threshold | Fields auto-accepted | Correct among accepted | Fields routed to review | Errors caught in review |
0.80 | 90 | 84 | 10 | 2 |
0.95 | 62 | 61 | 38 | 7 |
At 0.80, the workflow auto-accepts 90 fields but 6 of the 8 total errors pass through, leaving 6 wrong values in the output. Raising the threshold to 0.95 cuts auto-accepts to 62, catches 7 of the 8 errors in review, and leaves just 1 error live, but a reviewer now checks 38 fields instead of 10. The threshold move traded 28 extra reviews for 5 fewer errors. Whether that trade is worth it depends on what a wrong base rent costs versus a reviewer's time.
Variations and Edge Cases
A confidence threshold behaves differently depending on how many bands it defines and whether it adapts. The variants below change how sharply the cutoff sorts the work.
Variant | Behavior |
Single threshold | One cutoff, two outcomes: accept or review. Simplest, but blunt on high-stakes fields |
Dual threshold | Two lines create an accept, review, and escalate band for finer routing |
Field-specific | Each field type gets its own cutoff tied to its error cost |
Adaptive | The cutoff shifts as measured accuracy or reviewer capacity changes |
Uncalibrated input | A threshold on overconfident scores routes too few fields and misses errors |
Confidence Threshold vs Confidence Score
A confidence threshold is often confused with a confidence score, but they are different objects. A confidence score is the model's per-field estimate of correctness, a number the model outputs. A confidence threshold is the human-set cutoff that decides what to do with that number. The score is measured; the threshold is chosen.
Put simply, the score is the reading and the threshold is the rule applied to it. A field scoring 0.88 means nothing until a threshold says whether 0.88 clears the bar. One model produces one score per field, but a workflow can apply many thresholds to those same scores depending on how costly each field's errors are.
Frequently Asked Questions
What is a good confidence threshold for document extraction?There is no universal number. Common practice sets ordinary fields at 0.85 to 0.90 for auto-accept and high-stakes financial fields at 0.95 or above. The right threshold depends on how costly an error is and how well the scores are calibrated, not on the number alone.
Does raising the confidence threshold improve accuracy?Raising the threshold routes more fields to review, so it catches more errors but also flags many correct values, trading throughput for precision. It improves the accuracy of the final output only at the cost of more human review. There is no cutoff that maximizes both accuracy and speed at once.
How is a confidence threshold different from a confidence score?A confidence score is the model's per-field estimate of correctness. A confidence threshold is the cutoff a workflow applies to that score to decide whether the field auto-accepts or routes to a human. The score is measured; the threshold is chosen.
Related Terms
Confidence Score
Human-in-the-Loop
Exception Handling
Straight-Through Processing
Precision and Recall