Which answers deserve a closer look?
You are using AI to answer questions about a company’s finances. Some answers may be wrong, but checking every one takes time.
Should you start with the answers the AI doubts most or the ones whose calculations and cited figures raise warnings?
When we cannot check every AI answer, how should we choose which ones to review?
How we tested it
We compared two approaches on 1,127 answers from GPT-5.4 mini, using financial-report questions from FinQA. The AI received report extracts and supplied an answer, calculation, cited figures and confidence score.
- Confidence: ask the AI, “How sure are you?” Review the answers with the lowest confidence first.
- Verification: use rule-based checks to recalculate the answer and look for the cited figures in the source. Prioritize calculation warnings, then citation warnings and unused inputs. Tiny calculation differences within our allowance do not trigger an arithmetic warning.
See a simple example. Neither approach sees the answer key when choosing what to review.
We then compared answers with FinQA’s answer key, accepting small numerical differences. For percentages, answers within 0.005 percentage points counted as matching. 347 answers did not match. Both approaches received the same number of review slots. Matching and automatic-review rules.
Three findings at a glance
- Verification selected more mismatches at all four tested review capacities.
- With 225 review slots, it selected 143 mismatches, versus 103 for confidence.
- Even verification missed most mismatches at that capacity. Better prioritization is not complete protection.
1. Verification made more of the same review budget
Verification selected more mismatches than confidence at every capacity we tested. The gap was largest at 20%: 40 additional mismatches selected, without adding a review slot.
| Review capacity | Responses reviewed | Mismatches caught | |
|---|---|---|---|
| Confidence | Verification | ||
| 5% | 56 | 31 | 35 |
| 10% | 112 | 59 | 88 |
| 20% | 225 | 103 | 143 |
| 50% | 563 | 204 | 211 |
The answers left outside review also matched the answer key more often. At 20% capacity, agreement was 77.4% with verification, versus 72.9% with confidence.
2. How many mistakes did we catch? How many did we miss?
At the 20% capacity, verification selected 143 mismatches in 225 slots. That gives us two useful measures:
- Precision: 63.6%. About 64 of every 100 review slots selected a mismatch, compared with 46 for confidence.
- Recall: 41.2%. We selected about 41 of every 100 mismatches in the batch, compared with 30 for confidence.
That still left 204 of the 347 mismatches unreviewed. Reviewing half the responses raised verification’s recall to 60.8%, leaving 136 mismatches outside review. A better queue helps; it does not make the remaining answers reliably correct.
3. Make warnings reflect what counts as a mistake
Do not automatically give priority to numerical differences your grading rule already accepts. A checker can detect a tiny inconsistency without identifying an answer that fails your matching rule. When review slots are limited, that warning can push a more useful review further down the queue.
That was a practical lesson from this study: align calculation warnings with the numerical allowance used to assess answers. A missing warning still does not prove an answer is correct. How we handled the calculation threshold.
The checks also have a limit: a number can appear in the report and the arithmetic can work, yet the AI may have chosen the wrong year or financial measure. Checking the work is useful, but it is not the same as understanding whether it answers the question.
What would we test next?
Part 2 is now available. We tested four AI configurations on a separately selected set of FinQA questions and a second financial dataset, TAT-QA. Read the follow-up study.
We tested only one model: GPT-5.4 mini. Other frontier models could make different mistakes or judge their own confidence differently, so the better review strategy could change. Next, we would test the same rules with newer frontier models and another dataset to see whether verification’s advantage holds.
We would also investigate checks on the choice of figures and whether combining confidence with verification helps. Neither improvement has been demonstrated here.
The takeaway
On these financial questions, verification selected more answer-key mismatches than confidence at every tested review capacity. These results describe one model and dataset, not a universal winner.
A useful review rule should prioritize the mistakes you want to catch, not every difference it can detect.
What counts as a mismatch matters: our results use the stated numerical allowance, not a measure of financial harm. Matching rules and sensitivity. We measured selection, not whether people corrected the answers or saved time. Limits and reproducibility.
Methods and supporting material
Appendix
A worked example of confidence versus verification
This is an illustration, not a study response. Revenue rises from $100 million to $120 million. The AI answers 25%, writes (120 − 100) ÷ 100 × 100, and reports 99% confidence.
Confidence puts it lower in the queue because the AI says it is sure. Verification recalculates the expression, gets 20%, and raises a warning. Neither uses FinQA’s answer key to select it.
Accuracy, precision and recall
The batch contains 765 matches, 347 numerical mismatches and 15 abstentions. Pairs below are confidence / verification. Precision divides selected mismatches by all review slots. Recall divides them by all 347 mismatches. Released accuracy divides matching answers by all answers left outside review.
| Capacity | Unreviewed | Released accuracy | Precision | Recall |
|---|---|---|---|---|
| 5% | 1071 | 70.49% / 70.87% | 55.36% / 62.50% | 8.93% / 10.09% |
| 10% | 1015 | 71.63% / 74.48% | 52.68% / 78.57% | 17.00% / 25.36% |
| 20% | 902 | 72.95% / 77.38% | 45.78% / 63.56% | 29.68% / 41.21% |
| 50% | 564 | 74.65% / 75.89% | 36.23% / 37.48% | 58.79% / 60.81% |
The random baseline selected 17, 35, 67 and 185 mismatches. It uses one fixed random ranking with the same automatic-review rules, not the average of many random runs. Lines between tested capacities are visual guides, not additional observations.
Setup and numerical matching
GPT-5.4 mini answered with low reasoning effort and no tools. Twenty nonnumeric or nonfinite-reference records were excluded before the original run. A separate 200-question development set selected the original verification ranking. The later threshold adjustment was made after examining test results; it was not validated on that development set or new held-out questions.
Both policies first include 35 mandatory responses: 15 abstentions and 20 unusable derivations. Fourteen have numerical mismatches and six match. These consume slots, rather than being extra reviews. The remaining slots use the confidence or warning ranking. Policies and tie-breaking seed are fixed; the checker adjustment changes arithmetic warnings only.
We compare structured final numbers with FinQA’s qa.exe_ans answer key. Percent outputs are divided by 100 unless the reference program ends in multiplication by 100. Allowances are 0.00005 for converted percent/ratio values and 0.005 in other displayed units, with equality accepted. These are absolute numerical allowances, not validated financial-materiality thresholds. The adapter does not settle every possible unit convention. This is not official FinQA program-execution accuracy.
Calculation warnings: thresholds and lessons
The initial calculation check flagged differences smaller than our grading allowance: above 0.00001 in displayed units. After examining the initial results, we aligned that threshold with the grading allowance and recomputed the review queues using the same saved responses, without changing the answers or grading rule. Confidence selected more mismatches at smaller review budgets under the initial setting; that record is preserved.
The revised checker flags a difference between the stated answer and its recalculated expression above 0.005 in percent and other displayed units, or 0.00005 for ratios. The checker does not consult the answer key. Evidence warnings, unused-input warnings, priority order and mandatory rules are unchanged. Arithmetic warnings fell from 327 to 114. Of the 213 removed, 183 were answer-key matches and 30 were mismatches under the stated scoring rule.
Both answer-key scoring and this checker revision are post-study analyses. The exact revision was fixed before its follow-up run, but motivated by results we had already seen. Testing it on the same saved answers makes this exploratory, not independent validation.
In a separate sensitivity check, stricter numerical matching favored the original verification rule, while a looser allowance favored confidence. That check kept the original queues fixed; it does not establish that the revised checker wins under every scoring rule. Detailed historical results remain in the original analysis record.
Limits, cost and code
This is one model configuration, saved batch and tie-breaking seed. It does not establish statistical superiority on future data. Public benchmark training exposure cannot be ruled out. No people corrected answers; review slots are not measured labor costs.
Recorded model usage across setup, development and final evaluation was approximately $3.60, a usage-based estimate rather than a reconciled invoice. This offline follow-up added $0 external spend. Subscription and human effort are excluded.
Independent review reconstructed all 1,127 responses, all 24 original/revised queues and the scoring sweep. Historical inputs and results remain unchanged. Original study code and the earlier analysis package are on GitHub. The calculation-warning analysis code, tests and aggregate results reproduce the methods behind the updated figures. Exact empirical replay requires the original saved responses and source inputs, which are not publicly distributed. Public tests use synthetic examples.
Further reading
- FinQA: A Dataset of Numerical Reasoning over Financial Data. Chen and colleagues, 2021. The dataset behind our questions and answer key.
- Just Ask for Calibration. Tian and colleagues, 2023. Research on asking language models for confidence; background for our confidence method, not evidence for this financial-review comparison.