AI Troubleshooting

30% of AI tool 'problems' are actually service outages. AI companies have reliability issues more often than most software because their inference infrastructure is under enormous load. Status pages: - OpenAI (ChatGPT, API): status.openai.com - Anthropic (Claude): status.anthropic.com - Google (Gemini): workspace.google.com/status - Midjourney: discord.gg/midjourney (check #status-updates channel) - Microsoft Copilot: admin.microsoft.com (M365 admin center) or portal.office.com/servicestatus Also check: downdetector.com for user-reported issues, and the company's official Twitter/X for real-time outage announcements. If the service is partially degraded (not a full outage), you might see intermittent responses, slower than normal generation, or specific features failing. Partial degradation is harder to spot on status pages because companies are slow to update them. If there's a confirmed outage: stop troubleshooting and wait. Nothing you do will fix someone else's server problem.

The second most common cause of AI tool failures is account-level issues, not technical ones. Check in order: 1. Subscription active? Log in to the platform. Confirm your plan shows as active. A failed payment silently downgrades you to free tier -- which then hits its limits quickly. 2. Usage limits hit? Free tiers have hard daily limits. ChatGPT free has message caps. Claude free cuts off after a certain volume. Gemini free limits advanced features. If you're hitting limits, the tool doesn't error clearly -- it just slows down or stops responding. 3. API key valid (if using the API)? API keys can be rotated, expired, or revoked. In the OpenAI dashboard: API → API Keys → confirm your key is still listed as Active. Anthropic: console.anthropic.com → API Keys. 4. Payment method current? If your credit card expired, the subscription continues for a grace period then cuts off. Check billing settings. 5. Account flagged or restricted? This is rare but happens -- especially if automated usage triggered abuse detection. Check for any account notices or warning emails. Most 'the AI stopped working' problems resolve at this step.

'It's not working' covers a dozen different problems. Narrow it down: Not responding at all: Requests go in, nothing comes back. Usually: network issue, service outage, or rate limit with no error message. Error 429 -- Too Many Requests: You've hit a rate limit. Wait and retry. For API users: implement exponential backoff in your code. For chat users: slow down. Error 401 -- Unauthorized / Invalid API Key: Your API key is wrong, expired, or not authorized for this endpoint. Regenerate the key, update it everywhere you've used it. Error 500 -- Internal Server Error: The AI service has a server-side problem. Not fixable from your end. Try again in a few minutes or hours. Context length exceeded: You've sent more text than the model can handle in one conversation. For GPT-4: 128K token limit. For Claude: 200K. Start a new conversation with less context, or summarize earlier content. Degraded output quality: The model is responding but outputs seem worse than usual. Possible causes: model was updated, the conversation accumulated too much context, or the prompt changed. Automation/integration not triggering: For Zapier/Make setups -- the trigger isn't firing, or the action is failing silently. Check the platform's task history for the specific error. Model returning unexpected format: For API/automation use -- the model is generating text but not in the format you specified. This is a prompt issue, not a connection issue.

Troubleshooting is the process of eliminating possibilities until one remains. Remove variables one at a time. For chat tool issues: 1. Open a new conversation (rules out accumulated context issues) 2. Try a simple prompt: "Say the word hello." If that works, the problem was your original prompt. 3. Try a different browser (rules out browser extension conflicts -- ad blockers frequently break AI chat UIs) 4. Try an incognito/private window (rules out cookies and cached data) 5. Try the mobile app if available (rules out desktop app or browser issues) For API issues: 1. Test with the simplest possible API call (just 'Hello' as a message) using the API dashboard's built-in playground 2. If playground works but your code doesn't: the API is fine, your code is wrong 3. If playground fails too: the API key, account, or service has an issue For automation issues (Zapier/Make): 1. Manually trigger one run and watch it step by step in the scenario editor 2. Check the input data at each step -- is the data actually what you expect? 3. Remove the AI step temporarily and confirm the rest of the automation works 4. Then add the AI step back and check what it's actually receiving and returning Isolating variables is not glamorous, but it's the only reliable way to find problems in complex systems.

If the AI is responding but outputs are wrong, inconsistent, or not in the format you expect -- the problem is almost always the prompt. Outputs are too long or too short: Add explicit length instructions. "In exactly 3 bullet points" or "In 2 sentences maximum." Outputs are ignoring your format: Add "Respond with only [format]. No explanations, no other text." at the end of the prompt. Models are trained to be helpful by adding context -- you need to explicitly turn that off. Outputs are inconsistent across runs: The same prompt produces different results because LLMs are probabilistic. Lower the temperature setting (API) or add more constraints ("Always respond in the same structure: [structure]"). The model keeps forgetting instructions: The instruction is getting lost in a long conversation. Put the most important constraints at the very beginning AND the very end of your prompt -- models attend more to the start and end of long contexts. The model is refusing to complete the task: Rephrase. Refusals are often triggered by word choices, not intent. "Analyze why this email might feel aggressive" works differently than "write an aggressive email." For automation prompts specifically: if the model is returning JSON with markdown formatting (code blocks around it), add: "Return valid JSON only. No markdown. No code fences. No surrounding text."

If the AI itself is fine but your integration with it isn't working: OAuth/account disconnection: In Zapier or Make, go to the connection and click "Reconnect" or "Reauthorize." OAuth tokens expire. This is the most common integration failure. Webhook not receiving data: Confirm the webhook URL is correct in both the sending app and the receiving automation. Test with a tool like webhook.site -- paste that URL as the destination, trigger the event, and confirm data arrives. If data arrives at webhook.site but not at your automation, the URL in your automation is wrong. API key rotation: If you rotated an API key in OpenAI or Anthropic, you must update it in every automation, app, and code file that uses it. A key rotation that updates in one place and misses another place creates partial, confusing failures. SSL certificate issues: If you're self-hosting n8n or another tool, expired SSL certs cause connection failures. Check cert expiration dates. CORS errors (for browser-based tools): This is a development issue -- if you're building a tool that calls AI APIs from the browser, you'll hit CORS restrictions. Calls to AI APIs should go through your backend server, not directly from the browser. Rate limit vs. quota exceeded: Rate limits are temporary (wait and retry). Quota exceeded is a billing limit (add payment or upgrade). They look similar but have different fixes.

Some problems are genuinely easy to diagnose and fix. Others require access to logs, configuration visibility, or experience with the specific error pattern that makes self-diagnosis unreliable. Get help if: - You've been troubleshooting for more than 90 minutes without progress - The error message returns no useful results when searched - The problem is intermittent and you can't reproduce it reliably - Multiple things stopped working at the same time (usually points to account, network, or infrastructure) - You're seeing data loss or unexpected data being sent/received Before contacting support (for any platform): Collect your account email, the exact error message, the time it happened, what you were trying to do, and what you've already tried. This cuts resolution time dramatically. For OpenAI issues: help.openai.com For Anthropic issues: support.anthropic.com For Zapier issues: zapier.com/help Or book a Geek. We've seen most of these error patterns before and usually know the fix faster than support queues.