Tax monitoring
Monitor five official sources for one tax regime across two jurisdictions, including checks with no relevant change; triage and read updates, extract dates, draft summaries and obtain expert applicability approval. Same update volume/coverage and approved output. No filing, autonomous advice or policy implementation.
Evidence: Late-2025 self-reported tax-research survey plus named 2026 tax-research implementation account; no measured end-to-end tax monitoring labor result. Conditional estimate.
Reference task model
The reference workload starts with 100 manual staff-hours. Organization settings adjust the hours below at the same work volume. Task shares and case budgets are assumptions. Checking, corrections and manual fallback are included.
| Operating work | Before staff-hours | After staff-hours fast–constrained | Saved staff-hours constrained–fast |
|---|---|---|---|
Check the five official sources, including no-change checks; detect and triage updatesAssumption and referencesAnalyst operational assumptions, without an empirical scan-time anchor. Low: feeds/readable exports cover 80% of baseline scanning work, 90% succeeds and accepted automation removes 80% of its human time. High: all five sources work reliably, 98% succeeds, accepted scan/triage needs 5% of its former time. Human spot checks and false-positive review remain in residual scan time; no-change checks belong to the denominator. Sources: 3 | 30 | 2.1–12.7 | 17.3–27.9 |
Read relevant updates, extract publication/effective dates and draft cited summariesAssumption and referencesApply the adjacent reported 32% research saving to the entire 30-hour read/extract/draft block in the constrained case. The fast case assumes 65% less time, consistent with the conditional same-output interpretation of Copeland Buhl’s 2–3-times-faster research claim. Both are scope transfers to monitoring, not observed monitoring effects. No additional coverage or success discount is applied to these real-world research results. | 30 | 10.5–20.4 | 9.6–19.5 |
Check legal applicability, effective dates and completeness; correct and approve summariesAssumption and referencesRetain all expert review and existing correction labor. Forty hours is an assumed monitoring-workflow share, not a measured result. The PLAT benchmark supplies accuracy context rather than a review-hours factor. No filing or autonomous advice is included. Analyst assumption; no published stage timing. | 40 | 40 | 0 |
Additional failed-attempt, false-positive and AI-specific correction laborAssumption and referencesAssumed extra human work per 100 baseline hours, including investigating missed or misclassified updates and checking failures. Unsuccessful/uncovered work also keeps its full baseline labor above. Ongoing source-feed maintenance is separate. Analyst assumption; no published stage timing. | 0 | 1–3 | -3–-1 |
| Reference workload total | 100 | 53.6–76.1 | 23.9–46.4 |
Hours refer to this reference workload. Saved = before − after; negative saved hours add work. The fast case has fewer remaining hours. 23.9–46.4% saved before outward rounding to 20–50%. Displayed hours are rounded to 0.1. These are conditional scenarios, not confidence limits.
Reference scenario conditions
Constrained case · 23.9% saved · 76.1 hours remain
Scanning has 80% coverage, 90% success and 80% accepted-task saving; read/draft work uses the reported 32% research effect without another coverage discount. Retain all 40 expert-review hours and add 3 correction hours.
Fast case · 46.4% saved · 53.6 hours remain
All source checks are covered, with 98% success and 95% accepted-task saving; the research block assumes 65% less time. Retain all 40 expert-review hours and add 1 correction hour.
Citations
CoCounsel Tax: The AI tax assistant for in-house teams
Vendor user survey · 2026-06-29
Users report an average 32% less tax-research time.
Evidence period: Self-reported user survey November–December 2025.
- Sample size, task mix and review boundary not disclosed; not independently timed or verified. Research differs from scanning five sources and obtaining approval for every relevant update.
From hours to minutes: How CoCounsel reimagined tax workflow efficiency at Copeland Buhl
Named vendor/customer implementation account · 2026-03-09
Research projects reported 2–3 times faster; reviewers save 15–30 minutes per research memo. 67 active licenses.
Evidence period: 14,000+ messages over six months, exact months unstated. CoCounsel Tax & Audit launched March 2025 per separate dated first-party announcement.
- No timed sample, matched task baseline, exact dates or average review duration. 2–3 times faster implies roughly 50–67% shorter research only if the same task and output are compared. Memo-review minutes do not justify eliminating expert tax applicability approval.
How Amazon Finance streamlines regulatory inquiries by using generative AI on AWS
First-party technical deployment account · 2026-05-12
Finds historical documents and precedents and supports regulatory-response preparation. Parallel retrieval falls from 10 seconds to under 2 seconds.
Evidence period: May 2026 account names Claude Sonnet 4.5; exact go-live/observation dates unstated.
- Retrieval number compares system configurations; no human-hours result. Regulatory inquiries differ from ongoing tax monitoring.
Taxation Perspectives from Large Language Models: A Case Study on Additional Tax Penalties
Tax reasoning benchmark · 2025-03-05; revised 2026-02-08
Continuing difficulties applying tax law to facts and conclusions.
Evidence period: Recent 2025 model snapshots in revised benchmark; exact runs unstated.
- 300 Korean tax-penalty questions; no human-time study.
How to interpret this range
Conditional planning scenarios, not measured population bounds or confidence intervals.
The same work is completed manually using existing systems and approved inputs. Sources and calculation models are usable. These are workflow reference cases, not company types.
- The workflow is in steady-state use and passes checks for equivalent output quality.
- Volume, coverage and output remain unchanged. Saved capacity is not assumed to reduce staffing.
- Human review, corrections, exceptions and existing approvals are included.
- Baseline task shares and transferred task-saving rates are analyst assumptions, not measured industry allocations.
Savings include conventional extraction, matching, retrieval and integration delivered with each scoped project. The separate marginal contribution of GenAI is not established.
Separate costs: Implementation and training, Separate project/platform upkeep, Upstream remediation outside the project, Additional output and newly worthwhile work.
Already-automated work, low adoption, difficult exceptions or unreliable output can produce smaller savings, zero or additional work. The positive lower endpoint is not a guaranteed minimum.
- Baseline task hours, eligible coverage and remaining human-work budgets are scenario assumptions.
- Published task results inform the scenario; no source establishes the complete workflow range.
- Same work volume and required output; manual fallback and corrections are included. Separate upkeep is excluded.
Convert to monthly hours: Monthly operating hours saved = baseline monthly workflow hours × saving percentage. Net capacity released = operating hours saved − annual project upkeep / 12 − allocated shared upkeep / 12. Shared upkeep is allocated once across the portfolio.
What would improve the estimate
- Late-2025 self-reported tax-research survey plus named 2026 tax-research implementation account; no measured end-to-end tax monitoring labor result. Conditional estimate.
- Source failures, weak archives, heavy legal interpretation or long review/correction can put results below 20% or below zero.
- More than 45% is possible when routine source checks/drafting dominate baseline time and review is a smaller share; 40% retained review here creates a scenario-specific limit, not a universal tax-monitoring cap.
- Time scan work on all five sources, including no-change days and irrelevant updates; compare against the actual existing alert workflow.
- Use the same historical and live update sets and require equivalent recall; missing a relevant update is not a saving.
- Time human triage, date extraction, drafting, applicability approval, rejected summaries and corrections separately.
- Track sources/updates that fall back to manual processing and record the extra failed-attempt labor.
- Measure source-feed upkeep, test updates and classification maintenance separately.
Reviewed 2026-10-25 · Reference model data · Research notes