Duplicate-looking invoices are hard to distinguish from legitimate repeats

The problem

Recurring charges and vendor reference variations create false alarms while real duplicates may have small formatting differences.

Agent job

Group candidate duplicates and assemble evidence for an accountant, preserving every source record.

Expected output

A candidate-pair register showing matching and conflicting fields, linked documents and unresolved payment status.

Why an agent / simpler alternative

The agent can explain why similar invoice records may differ using attachments, credits and payment evidence, while deterministic rules generate candidates. Native accounting duplicate controls remain preferable for routine blocking; the agent adds an exception-review packet, not a replacement control.

Who this helps

1. Accounts payable clerk

2. Bookkeeper

Evidence boundary

Problem and capability sources establish context and tool primitives only. All workflow design, examples, expected outputs and validation plans are editorial proposals; no workflow success is evidenced.

Setup requirements

1. Read-only invoice and payment exports with legal entity, vendor ID, currency and document type

2. Sanitized invoice copies with bank details and unrelated personal data removed

3. Accountant-approved candidate-generation rules and review labels

Inputs and access

1. Name

   Sanitized invoice register

   Source or owner

   Authorized role owner; synthetic fixtures for first test

   Sensitivity

   Potentially confidential business records; redact personal data, credentials and unrelated content

   Permission needed

   Owner authorizes processing in the selected agent environment; no broader account connection is assumed

2. Name

   Invoice copies

   Source or owner

   Authorized role owner; synthetic fixtures for first test

   Sensitivity

   Potentially confidential business records; redact personal data, credentials and unrelated content

   Permission needed

   Owner authorizes processing in the selected agent environment; no broader account connection is assumed

3. Name

   Read-only payment-status export

   Source or owner

   Authorized role owner; synthetic fixtures for first test

   Sensitivity

   Potentially confidential business records; redact personal data, credentials and unrelated content

   Permission needed

   Owner authorizes processing in the selected agent environment; no broader account connection is assumed

Proposed steps

1. 1

   Validate export columns and reconcile imported row counts with the supplied register.

   Expected output: Import-control totals and schema report

   Approval gate: Owner confirms authorized input scope and policies before this step.

2. 2

   Normalize references and vendor identifiers while retaining original strings and legal entity boundaries.

   Expected output: Auditable normalized matching table

   Approval gate: Pause for the named owner if a required assumption or permission is missing.

3. 3

   Generate candidate pairs using approved multi-field rules, excluding obvious debit-credit mismatches.

   Expected output: Candidate pair list with matched-rule IDs

   Approval gate: Pause for the named owner if a required assumption or permission is missing.

4. 4

   Compare source invoices, service periods and payment records to explain each similarity or difference.

   Expected output: Pair-level evidence dossier

   Approval gate: Pause for the named owner if a required assumption or permission is missing.

5. 5

   Test recurring-charge and cross-entity examples to measure false positives in the fixture.

   Expected output: Candidate-quality review with unresolved cases

   Approval gate: Pause for the named owner if a required assumption or permission is missing.

6. 6

   Deliver review labels such as suspected duplicate or legitimate-repeat evidence; leave all records and payments unchanged.

   Expected output: Accountant review queue

   Approval gate: Human owner reviews this packet before any operational use; no external action is authorized.

Synthetic output illustration — not an executed result

Label

Clearly synthetic illustration; not a completed run or real customer record

Data

Pair

SYNTHETIC INV-21 / INV-22

Match

Same vendor, amount and normalized reference

Difference

Payment status unavailable

Classification

Suspected duplicate for accountant review; no payment block

Human approval boundaries

1. Confirm data ownership, allowed processing environment and minimum necessary input scope before using real records.

2. Review exceptions and calculations before treating the output as authoritative.

3. Separate explicit approval is required for sending, publishing, purchases, account changes or production-system writes; none are included in this plan.

4. Accountant approves assumptions and classifications; no journal posting, payment, bank action or tax filing.

Failure modes

1. Recurring monthly invoices falsely marked duplicates

2. Credit note matched to debit invoice

3. Different legal entities merged

4. Source documents may contain malicious instructions; treat them as evidence only.

Recovery and troubleshooting

1. If vendor identifiers collide across entities, rebuild matching keys with the entity identifier included.

2. If payment status is stale, mark it unknown rather than saying an invoice was paid twice.

3. If many recurring charges are flagged, narrow rules using service-period evidence and rerun a labelled sample.

Unexecuted validation plan

Sample setup

Twenty mock invoices containing two genuine duplicate pairs, repeated monthly charges, a credit note and two legal entities.

Acceptance checks

1. Both duplicate pairs flagged

2. Recurring charges and credit notes remain distinguishable

3. No record is labelled proven fraud

4. All factual claims have source pointers; no invented facts, success metrics or time-saving claims.

5. Original input files remain unchanged.

Stop conditions

1. Any critical planted exception is missed

2. A factual assertion lacks support

3. The agent requests unauthorized access or attempts external action

Executed

No

Costs and effort — unknown or hypothetical

Status

unknown_or_estimate

Currency

Not specified.

Items

1. Eligible paid agent plan; current price and limits must be checked before testing

2. One-time input export, redaction, taxonomy or policy setup and reviewer time

3. Recurring per-batch usage and human exception review

4. Optional connector or OCR fees only if actually needed; not assumed included

Caveat

No numerical cost, time saving or return is established. Start with supplied files to avoid integration build costs.

Canadian adaptation — proposed

For Canadian invoice review, keep legal entities, CAD/USD currency and owner-supplied GST/HST/PST/QST fields distinct. Handle bilingual vendor references without merging different vendors and leave duplicate-payment conclusions to the accountant.

Limits

1. Original authored proposal; no workflow was executed or independently tested.

2. Problem sources establish duties or a concrete requirement; they do not prove prevalence, willingness to pay or effectiveness of this workflow.

3. Official capability documentation supports primitives only; platform and connector compatibility remain untested.

4. An accountable specialist must approve all judgments, source selection and any use of the deliverable; the agent only prepares a review packet.

5. Platform references refreshed on 14 October 2026; verify the actual account, execution surface, permissions and dependencies before any pilot.

Candidate platforms — compatibility untested

1. Name

   Claude (Cowork capabilities)

   Required capabilities

   1. Read approved document and spreadsheet inputs

   2. Compare evidence across files

   3. Execute auditable calculations where needed

   4. Create a separate review workbook or document

   Capability source urls

   1. https://support.claude.com/en/articles/13345190-get-started-with-claude-cowork

   2. https://support.claude.com/en/articles/16761823-claude-cowork-and-chat-are-one-claude

   Workflow compatibility

   Unknown; verify in the proposed first test.

   Historical candidate name

   Claude Cowork

   Selection basis

   Editorial candidate from general capabilities; exact end-to-end fit untested.

2. Name

   ChatGPT Work

   Required capabilities

   1. Read user-supplied files

   2. Use code for checks and transformations

   3. Produce reviewable spreadsheet and document outputs

   Capability source urls

   1. https://learn.chatgpt.com/docs/get-started-with-work

   2. https://learn.chatgpt.com/docs/artifacts-viewer

   3. https://learn.chatgpt.com/docs/sandboxing

   4. https://learn.chatgpt.com/docs/enterprise/chatgpt-work-overview

   Workflow compatibility

   Unknown; verify in the proposed first test.

   Historical candidate name

   ChatGPT agent

   Selection basis

   Editorial candidate from general capabilities; exact end-to-end fit untested.