Here's how it starts. The CEO sees the support budget: $1.2 million per year. That's 15 agents at $80K loaded cost. Someone suggests AI. "We could cut that in half." A chatbot vendor promises 70% deflection. The contract gets signed.

Three months later, the chatbot deflection rate is indeed 70%. The dashboard looks great. Costs are down. Everyone's happy.

Six months later, churn is up 8%. NPS dropped 12 points. The subreddit has a pinned post titled "How to actually get help from [company]." Customer acquisition cost went up because word of mouth went negative.

Nobody connects the dots for another quarter. By then, the company has lost more in churned revenue than it saved on support.

This is the doom loop.

How It Works

Step 1: Company deploys AI to reduce support costs. The goal is usually a specific number: "reduce ticket volume by X%" or "handle Y% of queries without a human."

Step 2: AI hits the metric. It handles X% of queries. The dashboard shows success. Headcount is reduced or not replaced as agents leave.

Step 3: But the queries AI "handled" aren't actually resolved. Some percentage of customers gave up. Some got wrong answers. Some couldn't find a human and just stopped using the product.

Step 4: Customer satisfaction drops. But CSAT doesn't drop immediately because the survey goes to resolved tickets. The people who gave up never got to the survey.

Step 5: Churn increases. Slowly at first, then faster. The customers who couldn't get help are leaving. They're not filing tickets about it. They're just gone.

Step 6: Revenue impact becomes visible. By now, the original cost savings are dwarfed by lost customer lifetime value. But the AI deployment is "working" by its own metrics. Getting buy-in to fix it is an uphill battle because the dashboard says everything is fine.

Real-World Examples

Airlines were among the first to see the backlash. During the 2023-2024 holiday travel disruptions, multiple carriers pushed customers to chatbots for rebooking. The bots couldn't handle complex itinerary changes, multi-leg rebookings, or compensation claims. Customers waited hours on hold after the chatbot failed. Social media exploded. Congressional hearings followed.

Telecom companies have faced similar backlash. Xfinity's virtual assistant became a meme for being unhelpful. Customers reported being unable to cancel services or report outages through the bot. The FCC received thousands of complaints.

Financial services companies learned the hardest lesson. Customers who can't access their money or resolve billing errors through a chatbot don't just churn. They file regulatory complaints, request chargebacks, and post on social media. The CFPB now actively monitors AI support in financial services.

These aren't small companies with bad implementations. These are billion-dollar companies with million-dollar AI contracts. The problem isn't execution. It's the approach.

The Metric Problem

The doom loop persists because the metrics that track AI support success don't capture what matters.

"Deflection rate" counts every query the AI handled. It doesn't distinguish between "AI resolved the issue" and "customer gave up." If a customer types "I want to talk to a human" five times and eventually closes the chat, some systems count that as a successful deflection.

"Containment rate" measures what percentage of conversations stayed within the AI. Again, containment doesn't mean resolution. A customer who is "contained" but unsatisfied is worse than one who escalates to a human.

"Average handle time" measures how long each interaction takes. AI is fast. But a fast wrong answer doesn't help.

The metric that matters is "customer effort score after AI interaction." How hard did the customer have to work to get their problem solved? If they had to go through the bot, fail, call the phone number, wait on hold, and re-explain everything to a human, the effort was enormous even though the AI metrics looked great.

Breaking the Loop

The fix isn't "don't use AI." AI in support works. When it's done right, it's faster and cheaper than human-only support with equal or better customer satisfaction.

The fix is changing what AI does.

Bad AI: tries to resolve everything, blocks access to humans, optimizes for deflection, ignores resolution quality.

Good AI: resolves what it's confident about, routes what it isn't, always offers a human option, optimizes for resolution quality.

The specific implementation matters. A purpose-built classifier that understands customer intent and routes accurately is different from a chatbot that generates responses and hopes they're right.

With classification, you know what the customer wants in under 200 milliseconds. If it's a simple, automatable request (password reset, order status, business hours), the AI resolves it. If it's anything complex, ambiguous, or emotional, it goes to a human immediately. No loops. No fake empathy. No wall.

The deflection rate is lower. Maybe 40 to 50% instead of 70%. But the resolution rate is higher. The customer satisfaction is higher. The churn impact is zero or positive.

The Math That Matters

A support team of 15 agents costs $1.2 million/year and handles 50,000 tickets/year. Cost per ticket: $24.

AI-first approach (bad): deflects 70% of tickets. Saves $840K. But 8% churn increase on a $10M ARR business loses $800K. Net savings: $40K. And that's before accounting for increased acquisition costs and brand damage.

AI-assisted approach (good): handles 40% of tickets automatically, assists agents on 30% more. Saves $480K in labor. Churn is unchanged or slightly improved (faster response times). Net savings: $480K.

The second approach saves less on paper but more in reality because it doesn't create the doom loop.

The Signal to Watch

If you've deployed AI support and want to know if you're in the doom loop, watch one number: the percentage of customers who contact support more than once for the same issue.

In a healthy support system (human or AI), repeat contacts for the same issue run 10 to 15%. If your AI is driving that above 20%, customers aren't getting their problems solved on the first try. They're coming back. And some percentage of them won't come back at all.

Track it weekly. If it's rising, your AI is deflecting, not resolving. That's the loop starting.