Indemnification Clause Review 2026: Policy Guide for 90% vs 95% Auto-Accept Threshold

This guide sets the 2026 confidence-threshold policy for indemnification-clause review: keep auto-accept at 90% and route the 90–95% band to human reviewers.

It defines the exact conditions and validation check required before raising the global threshold to 95%.

How the 90% vs 95% cutoff works

In this review, 90% and 95% are auto-accept confidence cutoffs for indemnification-clause extraction, not measured error rates. A 90% cutoff auto-accepts a clause-level score at or above 0.90; a 95% cutoff auto-accepts only a score at or above 0.95. The percentage is an operating threshold chosen for routing, so it should not be read as an observed false-positive or false-negative rate for indemnification language.

The model emits one confidence score for each indemnification span, and the cutoff is applied to that span. Raising the cutoff from 0.90 to 0.95 does not change the extracted text or the model's score; it changes which spans are allowed to pass without review. In practice, the 95% rule is stricter because it demands more confidence before a clause leaves the queue.

The two-tier rule keeps 90% as the auto-accept floor and routes scores in the 0.90–0.95 band to human review. That band is treated as uncertain: it is not silently accepted, and it is not automatically rejected. The intent is to preserve recall by not forcing every borderline indemnification clause through the stricter 95% gate, while still containing false positives inside a reviewed band.

The cutoff is applied per indemnification clause, not per contract. One low-confidence indemnification clause can therefore be routed to review even when every other clause in the same contract has a score above the cutoff. The contract-level document is not the decision object; the clause-level span is.

A useful check is to record, for each clause, the raw score, the cutoff in force, and the resulting route. If a clause at 0.92 is auto-accepted under a 90% rule but would be reviewed under a 95% rule, that single difference is the entire effect of changing the threshold. Comparing routes this way keeps the threshold debate tied to clause-level decisions rather than to contract-level impressions.

What the evidence actually supports

No grounded source in this review supplies a 2026 contract-clause error rate for indemnification at a 90% versus 95% auto-accept level; the evidence gap is defined by three sources: Analytics ToolKit, OpenRouter, and FasterCapital.

Analytics ToolKit defines the confidence threshold as the inverse of the significance threshold, so a 0.05 significance threshold equals a 95% confidence threshold. It also states that this threshold corresponds to a probability of committing a type I error—a false positive—of (1 − the confidence threshold), which is 0.05 at 95%. That is a statement about cutoff semantics and error probability in a test, not a guarantee that 95% of indemnification clauses are extracted correctly; 95% confidence is not 95% clause accuracy.

OpenRouter, published October 1, 2026, places confidence-based escalation between always using the strongest model and fixed routing rules. Its discussion concerns routing economics and escalation behavior; it provides no contract-clause error figures. Therefore, any claim that indemnification extraction is better at 90% than at 95%, or vice versa, must be measured on your own labeled indemnification clauses rather than imported from that source.

FasterCapital describes indemnification in securities lending as a protective layer and financial safeguard that compensates a lender for losses after a default. That is a domain definition of the protection, not a measurement of extraction accuracy, false positives, or false negatives for indemnification clauses.

The supported evidence is therefore semantic and architectural: how a confidence cutoff is defined and how escalation can be routed. The missing evidence is empirical and domain-specific. A stratified holdout of labeled indemnification clauses, scored at both cutoffs, is the measurement needed before treating either 90% or 95% as an auto-accept standard.

Keep, raise, or two-tier: compare

Option A keeps a 90% global auto-accept level. It lowers review volume, but it also leaves higher false-positive exposure; because no grounded 2026 error rate shows that level is safe for high-risk indemnification clauses, it should not be treated as self-justifying. The burden is on evidence, not on the convenience of fewer reviews.

Option B raises the global level to 95%. It auto-accepts fewer clauses and creates more misses in the 0.90–0.95 band; it wins only when false-positive cost dominates and a stratified holdout shows the recall loss is acceptable. Without that holdout, the stricter gate mainly shifts work and risk rather than proving lower total error.

The comparison below isolates auto-accept volume; the two-tier design is the winner because it keeps medium volume while preserving review of the uncertain band. That middle position matters: it avoids the high exposure of accepting everything at 90% and the high miss rate of rejecting everything below 95%.

The practical check is to compare total error under both designs on a labeled holdout: if the 95% level does not reduce total error enough to offset missed clauses, keep 90% and review the band. The default should remain 90% with human review of 90–95% until that comparison is favorable, and the holdout should be stratified so the uncertain band is represented.

If a later holdout shows false-positive cost exceeds false-negative cost by enough to justify recall loss, revisit the global 95% option; until then, the two-tier design remains the better fit for indemnification clauses.

Costs and numbers that change the call

The cost model for this decision is annual cost = (auto-accepted clause count × false-positive rate × cost per missed indemnification) + (review-routed count × minutes per clause × reviewer hourly cost). It counts review minutes and false-positive indemnification exposure; it does not count model license price. Use it to compare the current 90% auto-accept level with a 95% level while holding every other input unchanged, so the threshold is the only variable.

Before changing the threshold, measure the false-positive rate separately for the 0.90–0.95 band and the ≥0.95 band on a labeled holdout. Classify each indemnification clause as correct, miss, or false positive. The holdout must be stratified by the clause and counterparty patterns that drive indemnification exposure, and examples must not be reused for tuning and final measurement.

Raising the global auto-accept level to 95% requires that the indemnification false-positive cost exceeds the false-negative cost. Furthermore, a stratified holdout must demonstrate that the 95% threshold results in lower total error compared to the 90% threshold with the 90–95% band routed to human review.

If a cost input is unavailable, keep it as a variable and solve for the break-even false-positive rate or break-even review minutes instead of inserting an estimate.

The decision to change the threshold should be based only on changes measured in the holdout variables, such as false-positive rates or review minutes, and not on changes to model license pricing.

Compare the two policies on the same holdout and choose the one with lower annual cost and lower total error; if the evidence is inconclusive, do not raise the threshold. Re-run the comparison whenever the clause mix, reviewer cost, or indemnification exposure changes.

What the evidence does not establish

The 90/95 cutoff does not transfer across clause types, contract families, or model versions without re-measurement. A threshold validated on one population is a local decision rule, not a portable accuracy claim; applying it unchanged to mutual indemnities, a new model build, or a non-English agreement assumes stability that has not been shown.

Use the edge-case table as a routing check before any global change. If the contract or model falls in a "breaks" cell, do not inherit the 90% auto-accept decision; send that subset to review or re-measure it. If it falls in a "wins" cell, the existing path remains appropriate.

The table is deliberately narrow: it names the three checks that must be recorded in the review log; contract family should be recorded as well when the agreement is not a standard vendor form.

Analytics ToolKit’s equivalence of a 0.05 significance threshold to a 95% confidence threshold is a threshold definition, not a clause-performance guarantee: it says how the cutoff is defined, not how often an extracted indemnification clause is correct. Do not read 95% confidence as 95% accuracy. Report per-class precision and recall for indemnification and for the clauses it is most often confused with, then compare the resulting error mix before any cutoff change.

A passing holdout must be stratified by the breaking dimensions in the table and must show the metric that matters for the chosen policy, not just a single aggregate score. Until that evidence exists, do not apply a global change to the breaking subsets; require re-measurement or review for those subsets.

1,000-clause review

This section alone uses the 1,000-clause worked example with a 90–95 review band and illustrative method inputs, not observed 2026 rates. The example contains 1,000 labeled indemnification clauses: 800 score at least 0.95, 120 score from 0.90 to 0.95, and 80 score below 0.90. Its illustrative method rates are 1% false positives in the at-least-0.95 band, 8% in the 0.90–0.95 band, and 15% misses in the below-0.90 band; these are method inputs, not grounded measurements.

The checkpoints convert the bands into expected counts: auto-accepting the top band leaves 8 expected false positives; reviewing the middle band exposes 9.6 expected false positives before acceptance; reviewing or rejecting the lowest band leaves 12 expected misses. These are expected values from illustrative rates, not observed counts.

The comparison matters for workload. The middle band has 120 clauses against 800 in the top band, but its illustrative false-positive rate is eight times as high, so it contributes 9.6 expected false positives against 8 from the top band. The lowest band contributes 12 expected misses, the largest single error count in the example. Review capacity should therefore be sized for the middle and lowest bands even though they contain only 200 of the 1,000 clauses.

Because the rates are illustrative method inputs, they cannot prove that raising a global auto-accept level would reduce total error. They only show how a 90–95 review band changes where errors are caught in this example. A stratified labeled holdout should supply measured rates and compare false-positive cost with false-negative cost before any global auto-accept level changes.

Decision rules for indemnification QA

Apply these five if/then rules in order when reviewing indemnification-clause extraction for 2026. They implement the default of keeping a 90% auto-accept level with human review of the 90–95% band, and they allow a global move to 95% only when the stated cost and holdout conditions are met, including when false-positive reduction justifies recall loss.

Rule 1 is the only path to a global 95% auto-accept level. It requires both a cost asymmetry favoring false-positive avoidance and holdout evidence that the higher cutoff reduces combined error and review cost. If either element is missing, the default operating point remains unchanged.

Rule 3 addresses capacity without changing the global cutoff. When reviewers cannot clear the 90–95% band, the response is targeted sampling of higher-scoring clauses, not a blanket raise that would also auto-accept lower-confidence indemnification language.

Rule 4 makes model or contract-family changes a hard stop. A 300-clause stratified holdout must be rerun before any threshold change, because a new model or family can shift score calibration and error mix even when headline accuracy looks stable.

Rule 5 prevents an incomplete holdout from authorizing a threshold change. If labels, stratification, or reproducibility are missing, the default remains 90% with human review of the 90–95% band.

What to do next

Frequently Asked Questions

Which confidence band is sent to human reviewers by default?

The band between the lower and higher auto-accept cutoffs is a human-review tier.

What cost condition must hold before considering a global threshold increase?

Indemnification false-positive cost must exceed false-negative cost before any higher-cutoff raise is considered.

What validation check is required for a threshold increase?

The increase must be validated on a stratified labeled holdout that shows lower total error.

Why is the threshold setting not static?

The feedback loop between HITL review and the confidence threshold makes the setting self-improving rather than static.

How does the lower cutoff treat a clause-level score?

The lower cutoff auto-accepts a clause-level score at or above 0.90.

Do the cutoffs represent measured error rates?

The cutoffs are auto-accept confidence cutoffs for indemnification-clause extraction, not measured error rates.

Quick answers

Also worth reading: Stanford: NLP vs Manual Clause Review: 82% Faster, 94% Accurate: Stanford: NLP vs Manual Clause · McCulloch v Maryland (1819) How the Necessary and Proper Clause Shaped Modern Contract Law: McCulloch v Maryland (1819) How · Oregon Divorce Settlement Review: $350 Clause-Cut vs $2,800 Escalate: Oregon Divorce Settlement Review: $350