Dr. Gleb Tsipursky, a behavioral scientist called the “Office Whisperer” by The New York Times, helps tech-forward leaders stop overpaying for AI while boosting engagement and innovation. He serves as the CEO of the AI consultancy Disaster Avoidance Experts, and wrote eight books, including The Psychology of AI Adoption at Work: From Resistance to Results (Georgetown University Press, 2026).
Long Beach is about to get a vivid look at how quickly artificial intelligence can move from a chat window into the customer-service front line. On Aug. 19, a hands-on workshop at Ballast Point Brewing invited customer-experience and technology leaders to build a functional AI agent that handles routine interactions and support operations.
That is useful experimentation. It also highlights a governance question that deserves an answer. Before these systems become the norm, when exactly must the agent stop and hand the customer to a person?
The risk is not simply that an AI agent will produce an obviously wrong answer. Here’s a dilemma: what if there’s an agent that works well on routine cases, earns enough confidence to expand and then encounters a dispute, a vulnerable customer, an unusual refund, an accessibility need or a factual claim outside its lane.
What does the AI do then? At that point, a vague promise of “human oversight” is not enough.
Every customer-facing AI workflow should have a written handoff rule with four parts in this order:
- Define the actions the agent is allowed to take: Answering store hours or retrieving an order status is different from approving a refund, changing a contract term or making a commitment that creates financial or legal consequences.
- Define stop conditions: Uncertain identity, repeated customer disagreement, requests involving safety, legal rights, financial hardship or exceptions to policy should trigger escalation rather than another automated answer.
- Name the human owner: “A person will review it” is not an operating process. A specific role should own the decision when the agent reaches its boundary.
- Set an escalation deadline: A handoff that sends someone into an unmonitored queue is not really a handoff. Customers should know when someone will respond and should not have to repeat the entire case after the transfer.
Long Beach has reason to get this right beyond the private sector. The City’s current list of upcoming technology contracts includes AI integration for call centers and AI-based tools. If public-facing systems are eventually procured, handoff design should be evaluated alongside features, price and speed. Vendors should be able to show what gets escalated, what is logged and who remains accountable for the final action.
Local entrepreneurs have a chance to build this discipline early. California State University Long Beach’s Apostle Incubator launched its fall programming on Aug. 25 to help campus and community entrepreneurs turn ideas into businesses. Startups that define decision boundaries while their workflows are still small will have an easier time than companies trying to shift accountability onto automation after customers are already relying on it.
Businesses should also measure what happens after the first answer. For 30 days, track how much time the agent saves, then subtract the time people spend correcting responses, reopening cases, handling escalations and repairing customer confusion. An agent that deflects more contacts but creates more rework may look efficient on a dashboard while making the whole service process worse.
Long Beach does not need to choose between using AI agents and protecting human judgment. It needs to make the exit ramp part of the system before traffic gets heavy.
