Finance information support
Find metric definitions/approved values, locate datasets, retrieve policies and precedents, prepare cited answers, check authority/current version and applicability, and resolve unsupported questions. Count all requester and finance-owner human work once. No downstream transactions or decisions; ongoing knowledge maintenance is separate.
Evidence: A 2026 knowledge-search case reports saved minutes without a baseline duration. Reference lookup duration, task mix, coverage and answer-success rates are explicit assumptions; no complete finance-knowledge workflow is measured.
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 |
|---|---|---|---|
Locate authoritative material: metric definitions/approved values, datasets, policies and precedentsAssumption and referencesFM estimates 6–10 minutes saved per search, without a baseline duration. This reference assumes a successful manual lookup takes 10 minutes and accepted assisted lookup takes 2 or 0.5 minutes: 8 or 9.5 minutes saved. Those timings, coverage (70%/95%) and success on covered work (85%/95%) are assumptions. They are not FM-measured percentages or precision/recall conversions. Uncovered and unsuccessful work retains its baseline time; all existing checking is separate. Sources: 1 | 60 | 8.6–31.4 | 28.6–51.4 |
| Assemble and communicate a concise cited answer | 20 | 3.8–12.9 | 7.1–16.2 |
Check source/version and applicability; resolve unsupported questions and approve answersAssumption and referencesPreserve all existing human checking, judgment and escalation labor. This is an assumed baseline share, not an observed finance service time study. Count requester and finance-owner work once. Analyst assumption; no published stage timing. | 20 | 20 | 0 |
Additional prompting, failed attempts and AI-specific correctionAssumption and referencesExplicit assumed extra labor per 100 baseline human hours, including attempts that fall back to the original work. In the factors above, unsuccessful/uncovered work retains its full baseline labor. Separate corpus maintenance is excluded. Analyst assumption; no published stage timing. | 0 | 1–4 | -4–-1 |
| Reference workload total | 100 | 33.3–68.3 | 31.7–66.7 |
Hours refer to this reference workload. Saved = before − after; negative saved hours add work. The fast case has fewer remaining hours. 31.7–66.7% saved before outward rounding to 30–70%. Displayed hours are rounded to 0.1. These are conditional scenarios, not confidence limits.
Reference scenario conditions
Constrained case · 31.7% saved · 68.3 hours remain
70% coverage and 85% success on preparation work; accepted lookup 80% faster and cited-answer preparation 60% faster; all 20 existing checking/escalation hours retained; add 4 failure/correction hours. Net 31.7 hours saved.
Fast case · 66.7% saved · 33.3 hours remain
95% coverage and 95% success on preparation work; accepted lookup 95% faster and cited-answer preparation 90% faster; all 20 existing checking/escalation hours retained; add 1 failure/correction hour. Net 66.6875 hours saved.
Other task mixes and faster examples
These additional assumptions show how work mix and coverage change the result. They are separate scenarios, not measured bounds or additions to the main saving.
A well-covered lookup
10 → 1 human minutes; 90% saved.
One well-covered, unambiguous question; current authorized source; short answer; accepted on first attempt; source checking takes 30 seconds in both methods. A 20-minute baseline with the same 1-minute assisted total would save 95%.
90%+ task-level savings are entirely possible under these assumptions. The mixed-workflow scenario is lower because of its assumed work mix, coverage and full retained checking, and does not impose a 70% cap on every service.
Citations
FM empowers 1,500+ engineers with trusted, AI-assisted access to engineering knowledge using Azure OpenAI
Vendor/customer estimate; adjacent engineering knowledge search · 2026-05-28
Estimated 6–10 minutes saved per search; 17,000 queries in the first two production months across more than 1,500 engineers.
Evidence period: First two production months, without calendar dates. Model version and timing protocol unstated.
- No baseline duration or full checking-time boundary. Estimated minutes cannot be converted to a percentage.
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.
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 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
- A 2026 knowledge-search case reports saved minutes without a baseline duration. Reference lookup duration, task mix, coverage and answer-success rates are explicit assumptions; no complete finance-knowledge workflow is measured.
- Poor coverage, failed retrieval, a weak manual baseline or lengthy numerical/source reconciliation can yield less than 30% or negative savings.
- A service dominated by simple well-covered metric/dataset/policy lookups with very short retained checking can exceed 70%; report its measured task mix rather than raising a universal cap.
- Time matched reference questions across metrics, datasets, policies and precedents, including ambiguous, restricted, conflicting and unsupported cases.
- Record total active human minutes, including requester prompting, source verification, finance-owner escalation, failures and corrections; report model wait separately.
- Estimate time-weighted content coverage, successful-answer rate and stage reductions independently, then compare equal answer accuracy/currentness.
- Measure recurring corpus/access/update maintenance separately.
Reviewed 2026-10-25 · Reference model data · Research notes