# Four specialized Grok Bots reconcile bank vs invoices; sample of 3,412 lines finished in 41 minutes leaving only 16 unprovable exceptions

Grok · X (Twitter) · Amount undisclosed (time-saving workflow case, not a revenue case; author reports sample of 3,412 lines finished in 41 minutes vs six working days by hand)

## Business summary

Author runs four specialized Grok Bots (Reader loads & cleans, Matcher pairs only on strict rules, Snagger tags failures, Closer writes the close note and stops) to reconcile bank lines against invoices each month. Manual work previously took six days; on a sample March of 3,412 open lines the process took 41 minutes and handed back only 16 unprovable lines (9 duplicates, 4 splits, 2 missing, 1 swapped total). Rules: same vendor, amount exact to the cent, dates within 3 days. It never forces a match and the ledger stays untouched until the human has seen the diff. Readers can reuse the “single-responsibility bots + strict match rules + exception-only handback + human final yes” month-close pattern.

## Business value

Author runs four specialized Grok Bots (Reader loads & cleans, Matcher pairs only on strict rules, Snagger tags failures, Closer writes the close note and stops) to reconcile bank lines against invoices each month. Manual work previously took six days; on a sample March of 3,412 open lines the process took 41 minutes and handed back only 16 unprovable lines (9 duplicates, 4 splits, 2 missing, 1 swapped total). Rules: same vendor, amount exact to the cent, dates within 3 days. It never forces a match and the ledger stays untouched until the human has seen the diff. Readers can reuse the “single-responsibility bots + strict match rules + exception-only handback + human final yes” month-close pattern.

## Workflow

Reader loads every line, cleans vendors and currencies, keeps source row

Matcher pairs only on same vendor, exact amount to the cent, dates within 3 days

Snagger tags each failure with the broken rule and both amounts

Closer writes close note, saves exceptions.md and stops; waits for human yes

## Preparation and limitations

Results and numbers are as reported by the original post author (sample March: 3,412 lines, 41 minutes, 16 exceptions; six working days by hand). No multi-month results disclosed. Workflow steps, freePrompt, and trial I/O are AgentFlow reconstructions and were not live-tested here. The post links to the author's separate article about Grok Bot memory, not the reconciliation setup.

## Starter prompt

````text
You are one specialized Bot inside a month-close reconciliation system. Execute only your assigned single role; do not exceed it.

Variables (user must supply):
- BANK_LINES: bank statement export (date, amount, vendor/description, source row).
- INVOICE_LINES: invoice list (date, amount, vendor, source row).
- ROLE: one of Reader | Matcher | Snagger | Closer.

Task (by ROLE):
- Reader: load every line, clean vendor names and currencies, keep original source row, output cleaned set.
- Matcher: pair only lines that satisfy ALL of: same vendor, amount exact to the cent, date difference ≤ 3 days. Leave anything unprovable unpaired.
- Snagger: tag every failure with the rule it broke and both amounts.
- Closer: write the close note, save exceptions.md, then stop. Never modify the ledger.

Output format:
- Structured list or markdown (cleaned lines / pair results / tagged exceptions / exceptions.md).
- Clearly mark which items are proved and which remain for human review.

Missing / conflict handling:
- If a line lacks vendor, amount or date, mark as unprovable and place in exceptions; do not guess.
- On amount-unit or currency conflict, stop that line and flag it.
- Never force a match on any pair that fails any rule.

Confirmation boundaries:
- You may only read the provided lines and produce notes/exceptions.md.
- Do not modify any ledger, bank record or invoice system.
- Final close requires the owner's explicit yes; until then the ledger remains untouched.
- Any payment, send or external write is forbidden unless the owner separately and explicitly authorizes it.

Untested. Reconstructed from original post description.
````

## Costs and pitfalls

Results and numbers are as reported by the original post author (sample March: 3,412 lines, 41 minutes, 16 exceptions; six working days by hand). No multi-month results disclosed. Workflow steps, freePrompt, and trial I/O are AgentFlow reconstructions and were not live-tested here. The post links to the author's separate article about Grok Bot memory, not the reconciliation setup.

## Source and attribution

kvinsi (@kv1nsiii)
https://x.com/kv1nsiii/status/2120198563120464112
2026-11-11
Results come from the source author. AgentFlow adds analysis and proposed workflows; outcomes are not guaranteed.
