How to Reconcile a Messy AR Spreadsheet With a Browser AI Agent (2026)

Lead

Point a browser AI agent at your accounts receivable spreadsheet and it cross-checks customer emails against your CRM, matches past-due invoices to payments, and flags stale records,all in 20 minutes instead of half a day. The agent writes every change to a sidecar column, logs every action, and stops before overwriting anything.

• Use a browser agent to read multiple SaaS tabs (CRM, billing portal, email, master sheet) at once
• Write proposed fixes to a sidecar column, not live data
• Build a structured audit log with timestamp, source URL, and reviewer signature

Verified against Claude in Chrome and the Claude Agent SDK on 2026-05-04.

Prerequisites

How ScaleIT handles this: See AI Quick-Start

• Accounts receivable spreadsheet (Google Sheets, Excel, or similar)
• Access to 2+ systems: billing portal (Stripe, QuickBooks), CRM (HubSpot, Salesforce), email inbox
• Claude in Chrome extension installed (browser-based) OR Claude Agent SDK with Playwright (unattended)
• 30–60 minutes to build issue list and scope before first run

Steps

1. Build the Issue List

Create a CSV file listing invoices you want fixed. This is your scope boundary,the agent only fixes what's listed.

1. Open your AR spreadsheet (master ledger)
2. Create new spreadsheet: ar-issues.csv
3. Add columns: invoice_number, customer_name, issue_type, expected_value, notes
4. Issue types: "stale_email", "status_drift", "missing_po", "currency_mismatch"
5. Populate 50–100 rows of invoices you know are wrong (agent runs 15–25 minutes)
6. Save CSV to shared location (Google Drive, Dropbox, local folder)

Example row: INV-2025-0412, Acme Corp, stale_email, acme-new@company.com, "Bounced; sales gave me new address"

2. Build the Agent Prompt

Write a single prompt telling the agent the goal, issue types, and tabs to open.

1. Create text file: ar-agent-prompt.txt
2. Include: goal (fix X invoices), 4 issue types with "check this tab, write to sidecar column" instructions
3. List exact tabs to open: CRM URL, billing portal URL, email link, AR sheet link
4. Add critical rules: "NEVER delete row, NEVER overwrite original columns, ALWAYS write to agent_fix column (sidecar), ALWAYS log to audit_log sheet"
5. Save to same folder as CSV

3. Set Up Sidecar Column and Audit Log

Add columns to spreadsheet to capture proposed changes without touching live data.

1. Open AR master spreadsheet
2. Add three columns: agent_fix (agent writes here), reviewed (you mark "approved"/"rejected"), applied_by (your name or date)
3. Create new sheet tab: audit_log with columns: timestamp, invoice_number, url_checked, field, value_read, value_proposed, status

Example audit log: 2025-05-04 14:22:15 | INV-2025-0412 | https://app.hubspot.com/.../acme | email | old@acme.com | new@company.com | approved

4. Run the Agent (Claude in Chrome)

Use browser agent to work through issue list.

Interactive (Claude in Chrome extension):

1. Install Claude in Chrome from Chrome Web Store
2. Open AR spreadsheet in one tab, ar-issues.csv in another
3. Click Claude icon → paste your prompt from Step 2
4. Ask: "Start with row 1. For each row, check the relevant tabs and write proposed fixes to agent_fix column. Log every action to audit_log sheet."
5. Let agent run 20–30 minutes; it will ask for clarification if needed
6. Review work in real time and course-correct

Unattended (Claude Agent SDK + Playwright):

1. Install: npm install @anthropic-ai/sdk @playwright/browser
2. Write agent script that reads issue CSV, calls Claude with your prompt, controls browser, writes audit log
3. Run on schedule: 0 2 * node ar-reconciliation-agent.js (nightly)

5. Review and Approve Changes

Agent finishes and writes proposed changes to sidecar column. You review and copy approved values to live columns.

1. Open agent_fix column in AR sheet
2. For each proposed change: mark reviewed = "approved" or "rejected"
3. For approved rows, copy value from agent_fix into live column (email, status, po_number, etc.)
4. Enter your name or date in applied_by column
5. Time estimate: 20–30 minutes for 100 rows

Verify

1. Confirm sidecar was used: Check live columns (status, email, amount),they should be unchanged. All proposed changes should be in agent_fix column.
2. Confirm audit log is complete: Check audit_log sheet. Every row agent touched should have timestamp, URL, and proposed value.
3. Spot-check one approval: Pick an approved row and confirm it applied correctly (e.g., email column shows new email address).

Troubleshooting

Agent gets stuck on "Customer not found in CRM"

• Cause: Customer name in ar-issues.csv doesn't match CRM name (typo, business vs. legal name).
• Fix: Update customer_name in CSV to match CRM exactly. Re-run that row or fix manually.

Agent overwrites live column instead of using sidecar

• Cause: Agent ignored instruction to use sidecar column.
• Fix: Stop run immediately. Restore from backup. Revise prompt: "NEVER modify live columns. ALWAYS write to agent_fix column." Re-run.

Audit log exists but agent never writes to it

• Cause: Agent doesn't know to log. Tab exists but not mentioned in prompt.
• Fix: Add to prompt: "After every fix, write a row to audit_log sheet with: [timestamp], [invoice], [url], [field], [old], [new], [status]." Re-run.

Agent marks rows "needs_human" without clear reason

• Cause: Agent hit edge case (two matching POs, ambiguous record) and correctly stopped.
• Fix: This is working as designed. Handle manually or ask agent for clarification. Mark as "rejected" in review column.

Work is wrong or incomplete

• Cause: Agent lacked context or prompt was ambiguous.
• Fix: Review agent_fix and audit_log sheets to see what agent saw. Revise prompt to clarify ambiguous cases, add examples, narrow issue types. Prepare second issue list with failed rows and re-run.

CTA

Reconciling AR by hand costs hours every month. A well-configured agent run cuts that to 30 minutes and produces a clean audit trail for close. Contact us for a free consultation if you'd like help designing a browser-agent workflow for your AR process.