AI PERFORMANCE / PRODUCTION DATA

AI results
on real books.

Expert-reviewed production results.

PRODUCTION DATA · METHOD AND LIMITS

0 AI QUALITY CHECKS LOGGED PRODUCTION · AS OF {{ asOf }}

0 COMPANIES MEASURED PRODUCTION · AS OF {{ asOf }}

0 TRANSACTIONS ANALYZED FY2025 AUDIT-LOG DATASET

01 / QUALITY CHECKS

Every AI quality check is logged.

Each quality check records its type, company, and outcome in the production workflow. The measurements come directly from the control stream used to move close-blocking work toward resolution.

M-01 PRODUCTION QUALITY CHECKS VALIDATED · JUN 2026 EXTRACTION

0 QUALITY CHECKS LOGGED

0 DISTINCT CHECK TYPES

0 COMPANIES COVERED

603,922 · JUN 2026 0 CUMULATIVE QUALITY CHECKS

MAR 2024 JAN 2025 JAN 2026 JUN 2026

TREND LINE ILLUSTRATIVE · MONTHLY DATA PENDING · ENDPOINT MEASURED

METHOD Count of AI quality checks logged in the production control workflow, grouped by check type across active companies. This measures the scale of the control layer. The analysis does not estimate how much any individual check changes book quality.

N 1,053 COMPANIES PERIOD THROUGH 2026-06

PRODUCTION DATA · METHOD AND LIMITS AS OF {{ asOf }}

02 / REVIEW

Expert review changed 10.8% of proposed treatments.

The FY2025 audit log compares each proposed transaction treatment with the treatment recorded after expert review. Across 1.55 million transactions, experts made a substantive change to 10.8%.

Changed treatments moved $72.5 million into deductible expense and $52.0 million out. The median net movement per company was +$4,282 into deductible expense.

+$0 MEDIAN NET MOVEMENT INTO DEDUCTIBLE EXPENSE, PER COMPANY · FY2025 MEDIAN LEADS · DISTRIBUTION IS SKEWED, MEAN REPORTED ONLY IN THE PAPER

M-02 EXPERT REVIEW CHANGES VALIDATED · INTERNALLY PEER-REVIEWED

0% OF 1.55M TRANSACTIONS SUBSTANTIVELY RE-TREATED IN EXPERT REVIEW

CONFIRMED · 85.7% CHANGED · NON-SUBSTANTIVE · 3.5% SUBSTANTIVELY RE-TREATED · 10.8%

GROSS CATEGORY MOVEMENT · FY2025

INTO DEDUCTIBLE +$72.5M

OUT OF DEDUCTIBLE −$52.0M

METHOD Every FY2025 journal line was reconstructed from the audit log and compared before and after expert review. The analysis separates substantive changes from non-substantive changes; 14.3% changed in any way. The final category is an operating proxy, not an independent audit of correctness.

N 1,505 COMPANIES · 1.55M TXNS PERIOD FY2025

PRODUCTION DATA · METHOD AND LIMITS AS OF {{ asOf }}

03 / INDUSTRY

Review rates vary by industry.

In the FY2025 dataset, substantive re-treatment rates ranged from 15.0% for construction companies to 8.3% for financial-services companies, compared with 10.8% across all companies. This comparison shows where review rates were highest and lowest; it does not establish why they differed.

M-03 SUBSTANTIVE RE-TREATMENT RATE BY INDUSTRY · FY2025 VALIDATED · INTERNALLY PEER-REVIEWED

CONSTRUCTION 15.0% HIGHEST

ALL COMPANIES 10.8% BASELINE

FINANCIAL SERVICES 8.3% LOWEST

0% SCALE MAX 16%

METHOD Same FY2025 audit-log analysis as M-02, grouped by customer industry. This page shows the highest rate, overall rate, and lowest rate. The methodology paper includes the full table.

N 1,505 COMPANIES · WITH INDUSTRY PERIOD FY2025

PRODUCTION DATA · METHOD AND LIMITS AS OF {{ asOf }}

M-04 AI-ASSISTED THROUGHPUT VALIDATION IN PROGRESS

1.28×–2.06× THE THROUGHPUT OF MATCHED MANUAL CONTROLS · SELECTED IN-APP WORKFLOWS

1.0× MANUAL 1.5× 2.0× 2.5×

SELECTED IN-APP WORKFLOWS

1.28× 2.06×

JOURNAL-EDIT WORKFLOWS · VS 12,626 HUMAN PEER PACKETS

1.70× 2.51×

METHOD 30-day production retrospective covering 1,486 durable accounting units across 117 packets and 492.49 app-active minutes, compared with matched manual controls. This measures application-action throughput. It does not measure labor cost, autonomy, or outcome quality.

FIGURES WITHHELD

Throughput figures publish once validation completes. The card frame ships now; the numbers ship when they clear.

N 1,486 UNITS · 117 PACKETS PERIOD 30-DAY RETROSPECTIVE · 2026

PRELIMINARY · FINAL FIGURES AFTER VALIDATION AS OF {{ asOf }}

04 / THROUGHPUT

Selected AI-assisted workflows showed higher in-app throughput.

In a 30-day production retrospective, selected AI-assisted workflows showed 1.28 to 2.06 times the application-action throughput of matched manual work. The analysis is preliminary and validation is in progress.

This metric measures work completed per active minute in the application. It does not measure labor savings, ROI, autonomy, or outcome quality.

Preliminary throughput results, with the limits shown.

05 / LIMITS

What these results do not show.

Each result above maps to a documented query, time period, and owner. The cards below show what the current data does not support, along with the evidence required or Uplinq's operating position.

The limits are part of the result.

B-01 Autonomous close rate This page does not report a straight-through-processing rate. The first analysis used different denominators for rule-posted and AI-posted work, so it was withheld. TO CLAIM IT Recalculate both on one shared, engine-labeled denominator.

B-02 AI sign-off A person signs off on every close. The routing measurement tracks how work reaches the right reviewer. TO CLAIM IT Not pursued. Human sign-off remains part of every close.

B-03 Audited correctness The final category after review is an operating proxy. Independent audit evidence is not part of this analysis. TO CLAIM IT An independent audit of a sampled ledger.

B-04 Realized labor savings & net ROI This page reports application-action throughput. Customer savings and ROI depend on staffing, process, and deployment context. TO CLAIM IT A longitudinal customer study.

B-05 Tax correctness The analysis tracks movement into and out of deductible categories. It does not test the accuracy of filed returns. TO CLAIM IT A filed-return outcome study.

B-06 Causal improvement from quality checks The 603,922 checks occurred in workflows that resolved close-blocking work. The analysis measures that association, not whether the checks caused the outcome. TO CLAIM IT A controlled rollout design.

GOVERNANCE · EVERY RESULT MAPS TO A DOCUMENTED QUERY, TIME PERIOD, AND OWNER · COUNTS REFRESHED AT BUILD TIME

06 / RESEARCH