Account reconciliation
Prepare and align two datasets, apply three agreed matching rules, investigate up to ten exception types, produce source-linked notes and verify balances/conclusions. No autonomous journal posting. Savings include configured deterministic matching delivered in this project.
Evidence: New named 2026 cash-reconciliation example reports 180 to five-to-ten minutes; adjacent current-model trade validation reports about thirty to under four. Both are vendor cases with incomplete review and timing methods. They support a potentially large narrow-workflow reduction, not empirical endpoints for the catalog scope.
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 |
|---|---|---|---|
| Prepare and align datasets; check completeness and totals | 20 | 3–15 | 5–17 |
Match records and identify genuine differencesAssumption and referencesRules execute over the entire population. Lower retains ten hours to resolve ambiguous candidates and inspect unmatched records. Upper retains one hour to initiate/check the matching result; no one manually compares already matched rows. This is mainly deterministic automation, not an attributed LLM reasoning effect. | 35 | 1–10 | 25–34 |
Investigate exceptions and draft source-linked notesAssumption and referencesLower: most exceptions need external follow-up and AI mainly organizes records. Upper: a small recurring set of timing, reference and split-payment differences is resolved from the two connected datasets; staff check proposed causes and handle the remaining genuinely unresolved cases. Twenty-five to six is an assumed routine-exception budget, not a measured published stage result. | 25 | 6–20 | 5–19 |
Independently verify balances, evidence and resolution decisionsAssumption and referencesThe same reviewer controls and final decisions remain. Upper saves five hours from having matched items, control totals and source-linked evidence organized together; professional judgment and unresolved differences still require people. Sources: 4 | 20 | 15–20 | 0–5 |
| Reference workload total | 100 | 25–65 | 35–75 |
Hours refer to this reference workload. Saved = before − after; negative saved hours add work. The fast case has fewer remaining hours. 35–75% saved before outward rounding to 35–75%. Displayed hours are rounded to 0.1. These are conditional scenarios, not confidence limits.
Reference scenario conditions
Constrained case · 35% saved · 65 hours remain
- Starting reconciliation still includes significant manual matching
- Two reliable datasets and three reusable, deterministic matching rules
- Upper case has routine exception types with causes traceable within connected records
- Unchanged balances, materiality, evidence requirements and reviewer sign-off
- Every manual repair, unresolved exception and reviewer correction counted
- Preparation fifteen + matching ten + investigation twenty + independent verification twenty = sixty-five human-hours, a thirty-five-hour saving.
Fast case · 75% saved · 25 hours remain
- Starting reconciliation still includes significant manual matching
- Two reliable datasets and three reusable, deterministic matching rules
- Upper case has routine exception types with causes traceable within connected records
- Unchanged balances, materiality, evidence requirements and reviewer sign-off
- Every manual repair, unresolved exception and reviewer correction counted
- Stable feeds, automatic rule matching and mostly recurring supported exceptions leave three + one + six + fifteen = twenty-five human-hours, a seventy-five-hour saving.
Citations
OpenAI: Franchisee reconciles cash in 94% less time with ChatGPT Work
Vendor-authored named customer case · 2026-10-20
Ted Taylor compares cash transactions with bank deposits monthly across three restaurants. Reported time falls from three hours to five to ten minutes; the article says he reviews results before implementing.
Evidence period: Current customer use described in October 2026; exact deployment/measurement dates and model version unstated. No known older experiment is identified.
- No transaction count, exception mix, timed sample, error rate, independent reviewer or breakdown of preparation/review is disclosed. It is unclear whether five to ten minutes includes every retained check. The catalog's two-dataset controlled reconciliation has a more explicit review standard.
OpenAI: Chatham scales its capital markets expertise with OpenAI
Vendor-authored named customer early measurement · 2026-10-02
Trade-validation review reportedly falls from about 30 minutes to under four; evidence is assembled, terms compared and discrepancies flagged.
Evidence period: Early measurement reported October 2026; exact dates unstated; GPT-5.6 generation explicitly identified.
- Trade validation differs from ledger reconciliation. No sample, error rate or timing protocol; comparison against experienced reviewers is ongoing. It does not justify removing independent review.
Microsoft: Empowering Finance with an AI Assistant in Microsoft 365 Copilot
Product announcement · 2025-10-20
ERP-grounded draft customer replies for staff review and Excel reconciliation matching/exception assistance.
Evidence period: Product availability in October 2025; unnamed pilots have unstated measurement periods.
- Days-to-hours pilot wording has no auditable human-time denominator. Supplies no task percentage or endpoint.
Netherlands AFM: 12 Building blocks for controlled use of auditing tools
Regulator guidance · 2025-12
No productivity result. Requires reliable input, verifiable outputs and continuing auditor responsibility.
Evidence period: 2025 supervisory guidance; older firm-selection data are unused.
- No numerical effect; none used.
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
- New named 2026 cash-reconciliation example reports 180 to five-to-ten minutes; adjacent current-model trade validation reports about thirty to under four. Both are vendor cases with incomplete review and timing methods. They support a potentially large narrow-workflow reduction, not empirical endpoints for the catalog scope.
- Use two comparable historical periods, then a shadow production run with the same rules and exception severity
- Measure preparation, matching oversight, investigation and reviewer labor separately
- Count all falsely matched, missed, duplicated and unsupported-explanation cases; time reviewer re-performance and manual fallback
- Replace assumed task shares with observed time and rerun the exact arithmetic
Reviewed 2026-10-25 · Reference model data · Research notes