AI-Assisted Reconciliation: Where It Helps and Where It Is Dangerous

// AI-assisted reconciliation: find gaps between DB and Paystack settlement
async function findReconciliationGaps(settlementCsv) {
  // Parse settlement CSV
  var settlementRefs = new Set(settlementCsv.map(row => row.reference));

  // Get our DB transactions in the same date range
  var dbTransactions = await db.transactions.findSuccessful({ from: settlementCsv[0].paid_at, to: settlementCsv.at(-1).paid_at });
  var dbRefs = new Set(dbTransactions.map(t => t.reference));

  var gaps = {
    inDbNotSettled: dbTransactions.filter(t => !settlementRefs.has(t.reference)), // we think it's paid, Paystack didn't settle
    settledNotInDb: [...settlementRefs].filter(r => !dbRefs.has(r)),              // Paystack settled but we have no record
  };

  // Now feed gaps to LLM for explanation
  var explanation = await callLLM({
    prompt: 'Here are reconciliation gaps: ' + JSON.stringify(gaps) + '. Explain what each type of gap might mean and what action the finance team should take. Do not create any accounting entries — only explain.',
  });

  return { gaps, explanation }; // human reviews and takes action
}

The LLM provides plain-English explanation and suggested actions. A human finance officer reviews, verifies, and executes any corrections. AI never touches the accounting system directly.

• Do not let AI create journal entries — journal entries in Xero, QuickBooks, or your GL must be created by humans. AI can draft the entries for human review, but not execute them.
• Do not let AI mark discrepancies as "resolved" — "resolved" in a financial context means a human has verified the resolution. AI marking something resolved hides real problems.
• Do not accept AI explanations as fact — LLMs can hallucinate plausible-sounding reasons for discrepancies. All AI explanations should be verified against actual Paystack API data or Paystack support.
• Do not run AI reconciliation without a clear audit trail — log what the AI was given, what it output, who reviewed it, and what action was taken.

See audit trails for AI financial actions for logging requirements.

Sign up for the McTaba newsletter