Verizon's AI gamble backfires: Project 624 brings worse support and higher bills

Verizon gives customers another reason to be angry

Verizon's Project 624 launched in late June to improve customer service with AI, 24/7 coverage, and faster resolutions. The plan: let AI handle simple tasks and route complex cases to humans. On paper, that strikes the right balance. In practice, customers say they're stuck with bots and the basics aren't getting done.

What Project 624 promised

Verizon positioned AI as a support multiplier, not a replacement for human cognition. The company said its tools run on Google's Gemini technology with specialized small language models for customer scenarios and claimed more than 90 percent accuracy.

Google's Gemini can be strong at structured tasks. But accuracy in a lab isn't the same as real-world task completion with edge cases, policy nuance, and billing exceptions.

What customers report instead

Subscribers say it's hard to reach a person, and bots are failing to solve issues. "it's trash and the level one support people gaslight the hell out of you" - StrictSchedule3113, Reddit user, September 2025.

Bot handoffs are a common complaint: "It's awful. 11 different bots tried to help me. I kept being passed around. Issue still not resolved." - neverenough14, Reddit user, September 2026.

Price hikes amplify the frustration: "Bill keeps getting bigger and bigger, yet we are paying for them to have a robot not a person." - Dunnomyname1029, Reddit user, September 2025.

How customers voted

Should customers always be served by human agents?

• Yes, bots can never replace human agents: 78.24%
• Not always, but whenever needed: 17.37%
• If Verizon wants bots to take over, train them better: 4.39%

Total votes: 501.

Why this matters to support leaders

Price increases plus higher effort and unresolved cases signal churn risk. Verizon has been losing customers and says Project 624 wasn't a reaction to that. Either way, outcomes are what count.

AT&T, T-Mobile, and Verizon all claim "best network," but most customers only notice small differences in coverage. Support experience and overall value win loyalty. If AI blocks access to humans or drags out resolutions, you lose the room.

Practical takeaways for support teams

• Set hard guardrails for AI containment. Escalate to a human when sentiment drops, a customer repeats themselves twice, there's account risk (billing disputes, cancellations), or time-in-flow exceeds a threshold (e.g., 6 minutes without progress).
• Make human fallback fast and obvious. Offer "talk to a person" or callback options up front. Promise an escalation SLA and show a live countdown or queue position.
• Measure what matters, not "accuracy." Track task completion rate, first contact resolution (FCR), customer effort score (CES), repeat contact rate, transfer loops per case, and abandonment. Report exceptions weekly.
• Stop the bot pinball. Use a single orchestration layer that owns state, not multiple uncoordinated bots. Preserve context across channels so customers don't repeat themselves.
• Keep models narrow. Small language models can work well on tight intent sets. Use retrieval for policy and billing rules. Don't force AI into edge cases-route them.
• Use AI to assist agents, not block customers. Summaries, next-best actions, and auto-filled forms reduce handle time and improve quality. Let humans decide when to deploy them.
• Align pricing with service. If you raise prices, show clear service gains: shorter wait times, specialized teams, or guaranteed callbacks. Otherwise, trust erodes.
• Run a weekly red team. Review top 20 failed intents, fix prompts, add guardrails, update knowledge, and retest. Publish before/after metrics.
• Design for choice. Voice: "Press 0 to reach a person." Chat: persistent "Speak to a human" button. Email: one-click schedule-a-call link.
• Protect high-value moments. Cancellations, outages, fraud, and billing disputes should default to human-first handling with priority routing.

If you're implementing AI support next quarter

• Start with one intent with clear success criteria (e.g., SIM swap or address change). Ship, measure, iterate, then expand.
• Define exit criteria before launch: if FCR drops 5 points or transfer loops exceed 1.2 per case, auto-scale human routing.
• Label failure modes (policy ambiguity, identity verification, missing data) and fix the system, not just prompts.
• Pilot with a control group and publish a scorecard to your frontline weekly.

Want structured training for support teams deploying AI safely? See practical, role-based options at Complete AI Training - Courses by Job.

The bottom line

Project 624 set the right intent: faster help, humans on the hard stuff. The execution-bot mazes, scarce human access, and rising prices-breaks trust.

Make AI invisible and useful, keep humans available, and measure by outcomes customers feel. We have asked Verizon to comment and will update the article if we get a response.