How to Automate Expense Categorization with Paradime, OpenClaw, and Google Sheets

Stop manually tagging spreadsheet rows. This guide walks you through a repeatable, outcome-driven workflow: read uncategorized transactions from a Google Sheet, classify them by expense category using OpenClaw's AI agent SDK, update the sheet automatically, and schedule everything to run daily with Paradime Bolt—zero human intervention required.

By the end of this article, you'll have a production-grade pipeline that measures uncategorized transactions, identifies the right expense categories (travel, software, meals, etc.), fixes the spreadsheet in place, and validates savings by logging every classification run.

What is Paradime?

Paradime is an all-in-one, AI-native platform that replaces dbt Cloud™ for analytics and data engineering teams. It provides a dbt™-native workspace where teams can develop, schedule, monitor, and debug data pipelines from a single interface.

Three core modules power the platform:

For this tutorial, Bolt is the key module. Bolt lets you schedule any combination of dbt run, dbt test, and custom Python scripts on a cron cadence—complete with environment variable injection, Slack/email/MS Teams notifications, and AI-powered debugging for failed runs.

📖 Paradime Docs — Bolt Scheduling

What is OpenClaw?

OpenClaw is an open-source, self-hosted AI agent framework. Unlike chat-only interfaces, OpenClaw agents can autonomously execute tasks—reading files, calling APIs, classifying data, and writing results back to external systems.

Key characteristics relevant to this build:

• Python SDK (pip install openclaw-python) with a simple Client class for chat completions and skill execution.

• API-key authentication via OPENCLAW_API_KEY environment variable.

• Model-agnostic: route to Claude, GPT-4, local Ollama, or any OpenAI-compatible endpoint.

• Skill system: Markdown-defined instructions that standardize how the agent handles specific task types.

For expense categorization, we'll use OpenClaw's chat_completion endpoint to classify transaction descriptions into predefined categories—with structured JSON output for deterministic parsing.

📖 OpenClaw API Documentation

Architecture Overview

Before diving into code, here's how all the pieces connect:

Figure 1: End-to-end flow — Bolt triggers the Python script on a cron schedule, the script reads uncategorized rows from Google Sheets, classifies each using OpenClaw, and writes categories back.

Setup: openclaw-sdk + Google Sheets API

Step 1: Install Dependencies

Create a pyproject.toml in your dbt™ project root (Paradime Bolt uses Poetry for dependency management):

Step 2: Google Cloud Service Account

1. Go to the Google Cloud Console.

2. Create a project (or select an existing one).

3. Enable the Google Sheets API and Google Drive API.

4. Navigate to IAM & Admin → Service Accounts → Create a service account.

5. Generate a JSON key and download it.

6. Share your target Google Sheet with the service account email (e.g., expense-bot@your-project.iam.gserviceaccount.com).

Step 3: OpenClaw API Key

1. Open the OpenClaw dashboard → Settings → API Keys.

2. Click Generate New Key.

3. Copy and store the key securely.

Step 4: Configure Environment Variables in Paradime

Navigate to Settings → Workspaces → Environment Variables → Bolt Schedules in the Paradime UI, and add:

Figure 2: Environment variable flow — secrets are stored once in Paradime and injected securely at runtime.

📖 Paradime Docs — Bolt Environment Variables

The Classification Script

Create the file scripts/categorize_expenses.py in your dbt™ project:

How the Workflow Maps to the Four-Step Loop

Figure 3: The measure → identify → fix → validate loop runs every day on autopilot.

Environment Variables Reference

Where to set them: In the Paradime UI, go to Settings → Workspaces → Environment Variables → Bolt Schedules section → Add New. Each variable is encrypted at rest and injected into the Bolt runtime.

Access in Python:

📖 Paradime Docs — Bolt Schedule Environment Variables

Bolt Schedule: Cron Daily

Now wire the script into Paradime Bolt so it runs automatically every day.

Option A: Schedules as Code (YAML)

Add this to your paradime_schedules.yml in the project root:

Important: The first command must be poetry install — this creates the virtual environment and installs dependencies (gspread, openclaw-python, etc.) before your script runs.

Paradime auto-syncs paradime_schedules.yml from your default branch every 10 minutes. For immediate pickup, click Parse Schedules in the Bolt UI.

Option B: Bolt UI

1. Navigate to Bolt in the Paradime sidebar.

2. Click + New Schedule → + Create New Schedule.

3. Fill in:

4. Configure notifications (Slack, email, or MS Teams) for failed and sla events.

5. Click Save.

Figure 4: The deployment flow — commit YAML to main, Bolt picks it up, and runs the pipeline daily.

📖 Paradime Docs — Schedules as Code

Monitoring and Debugging

Once the schedule is live, Paradime provides three layers of observability:

1. Run History Dashboard

Navigate to Bolt → [your schedule] → Run History. You'll see:

• Status: Success, Error, or Skipped for every execution.

• Execution Time History: A 30-day chart showing duration trends, success/error rates, and total run count.

• Trigger Source: Whether the run was scheduled (cron) or manually triggered.

2. Log Types for Failed Runs

When a run fails, click the Run ID to access three log levels:

3. Notifications

Configure alerts on three event types:

• Failed: Script crashed or returned non-zero exit code.

• SLA: Run exceeded sla_minutes threshold (15 minutes in our config).

• Success: Optional confirmation that the run completed.

Destinations: Email, Slack, and Microsoft Teams. Set these in the notifications block of your paradime_schedules.yml or via the Bolt UI.

4. Bolt System Alerts

Paradime also surfaces workspace-level issues automatically:

• Bolt Parse Errors — YAML syntax issues in paradime_schedules.yml

• OOM Runs — Out-of-memory failures

• Git Clone Failures — Branch not found or auth issues

• 24-Hour Run Timeouts — Runaway processes

📖 Paradime Docs — Viewing Run Log History

Troubleshooting Common Issues

❌ google.auth.exceptions.DefaultCredentialsError

Cause: The GOOGLE_CREDENTIALS_JSON env var is missing, empty, or contains malformed JSON.

Fix:

1. Verify the variable exists in Settings → Environment Variables → Bolt Schedules.

2. Ensure the value is the raw JSON string (not a file path).

3. Test locally:

❌ gspread.exceptions.SpreadsheetNotFound

Cause: The service account doesn't have access to the sheet.

Fix: Share the Google Sheet with the service account email (visible in the JSON key under client_email). Grant Editor access.

❌ 401 Unauthorized from OpenClaw

Cause: Invalid or expired OPENCLAW_API_KEY.

Fix:

1. Regenerate the key in the OpenClaw dashboard → Settings → API Keys.

2. Update the value in Paradime Bolt environment variables.

3. Verify the key format starts with sk-.

❌ json.JSONDecodeError During Classification

Cause: The LLM returned natural language instead of JSON.

Fix:

• Set temperature=0.0 for deterministic responses (already in our script).

• Wrap the parse in a try/except with a fallback to "miscellaneous" (already handled).

• If persistent, add a retry with a more explicit prompt or switch to a more instruction-following model.

❌ poetry install Fails in Bolt

Cause: Missing or malformed pyproject.toml.

Fix:

1. Ensure pyproject.toml is in the root of your dbt™ project (same level as dbt_project.yml).

2. Verify it includes all required dependencies: openclaw-python, gspread, google-auth.

3. Check Bolt Debug Logs for the exact pip/poetry error.

❌ SLA Breach (Run Exceeds 15 Minutes)

Cause: Too many uncategorized rows, or OpenClaw API rate limiting.

Fix:

• Add batching: process in chunks of 50 rows per run.

• Use OpenClaw's rate limit headers (X-RateLimit-Remaining) to throttle.

• Reduce the backlog by running the script manually once (python scripts/categorize_expenses.py) before enabling the cron schedule.

Figure 5: Troubleshooting decision tree — match the error type to the fix, then re-run.

Wrapping Up

You now have a fully automated expense categorization pipeline that:

1. Measures — Reads your Google Sheet daily and identifies uncategorized transactions.

2. Identifies — Sends each transaction description to OpenClaw for AI-powered classification into categories like travel, software, meals, and more.

3. Fixes — Writes the category back to the spreadsheet in a single batch update.

4. Validates — Logs confidence scores, error counts, and elapsed time so you can audit every run.

The entire workflow runs hands-free on Paradime Bolt's cron scheduler, with Slack or email alerts if anything goes wrong.

What You Can Do Next

• Add a confidence threshold: Only write back categories with confidence > 0.8; flag low-confidence rows for human review in a separate column.

• Track spend trends in dbt™: Load the categorized sheet into your warehouse with a dbt™ source, then build models that aggregate spend by category and month.

• Expand categories dynamically: Store your category taxonomy in a separate sheet tab and read it at runtime—no code changes needed.

• Multi-sheet support: Extend the script to iterate over multiple spreadsheets (e.g., per department) by reading a config from a sheets table.

Key Resources

Stop tagging expenses by hand. Let the machines do what they're good at—classifying text at scale—while you focus on what the numbers actually mean.