The Customer Agent Stampede: How Enterprises Can Prepare for Personal AI Agents at Scale

What is the Customer Agent Stampede?

It is a potential pattern in which many personal AI agents pursue customer requests at once, and their follow-up behavior creates more service traffic than the underlying needs require.

A delivery delay, billing error, or service outage could trigger it. Agents may check status repeatedly, retry after timeouts, or move to another channel when they cannot confirm progress. Enterprise automation may add further retries behind the scenes.

I use the term as a planning scenario. The evidence discussed here does not establish how frequently these surges occur today. It does show why leaders should examine how delegated customer requests would behave under pressure.

Are Customers Ready to Delegate Service Interactions to AI?

Avaya’s Customer Experience on Their Terms shows interest in delegating routine work, alongside divided views about AI representation.

When asked which company interactions they would be most likely to give a personal AI assistant, 48% selected checking the status of an order or request, and 39% selected waiting on hold or navigating customer-service systems.  

In a separate scenario, about 40% were comfortable with a personal assistant contacting a company’s AI using authorized account information and taking only pre-approved actions. Meanwhile, 42% were uncomfortable. These are stated preferences under defined conditions, rather than adoption rates.

Gartner’s July 2026 customer survey provides another signal: customers were about three times more likely to use third-party generative AI than company-provided chatbots to resolve service issues. Using AI for assistance doesn't always mean authorizing it to contact a company, but service leaders can no longer assume the journey begins in their own channels.

Could Fewer Human Contacts Still Create More Service Work?

Yes. Delegation changes how much personal time a customer needs to spend pursuing a request. Depending on how agents are designed, follow-up may become more frequent and continue for longer.

Consider an illustrative scenario: 10,000 personal agents each check delivery status six times during an hour. That creates 60,000 status requests for 10,000 underlying customer needs. It is a planning example, not a traffic forecast or survey finding.

A useful status check may prevent a phone call. A redundant check against unchanged information may add processing work without advancing resolution. The distinction depends on what the enterprise tells the agent and what the agent does with that information.

Repeated requests can also reveal real failures: missing updates, unclear ownership, or missed commitments. A traffic reduction target must therefore be paired with successful outcomes and access to help.

Why Can Retries Make a Disruption Worse?

Retries consume resources, and when many clients retry at once, they can make an overloaded service harder to recover. The engineering problem is well established.

AWS documentation on retry behavior explains how increasing the delay between attempts, adding randomized timing, and limiting retries help control repeated traffic. Google’s guidance on handling overload describes retry budgets and warns that retries at multiple layers can multiply the load.

Applying these lessons to customer agents is an architectural inference. An enterprise may face repeated attempts across APIs, web chat, and voice while its own systems retry downstream work. Capacity planning should examine both incoming contacts and the operations each contact triggers.

When the system slows down, a customer should receive a meaningful path forward. Otherwise, trying again elsewhere remains a reasonable response.

What Should an Authorized Customer Agent Receive?

The agent should receive enough reliable information to understand whether work has begun, what remains unresolved, and when another interaction would be useful.

I would design each delegated service journey around a persistent request record. That record should connect the verified customer, the agent’s permitted actions, the current case state, and completed steps. Access should remain limited to what that representative is authorized to see.

The response should distinguish between a request being received, work being in progress, approval being required, and the outcome being completed. It should provide a realistic next update point and a way to escalate when the deadline or circumstances change. Where supported, a subscribed update can reduce repeated polling.

Execution also needs protection. Repeating the same authorized refund request should not issue another refund. A changed request, however, must not be discarded simply because it concerns the same account.

How can Avaya Infinity help enterprises prepare?

Avaya Infinity connects systems, data, AI, people, and workflows in an enterprise-controlled orchestration environment. That architecture supports a practical response to demand for delegated services.

How does shared context reduce unnecessary restarts?

Avaya Infinity helps preserve relevant customer intent, interaction history, and context across channels and between AI and human agents. Connected context lets the next resource understand the work already done and the remaining need.

How can requests move through enterprise systems?

Avaya Infinity can trigger and track workflows across connected front- and back-office systems. Its support for Model Context Protocol, or MCP, provides a standard way to connect AI with enterprise tools and context. This helps link an interaction to the business process needed to complete it.

How can automation stay within appropriate limits?

Avaya Infinity supports enterprise policies, permissions, guardrails, logging, and auditability around AI and workflows. Enterprises must also establish the incoming representative’s delegated authority and implement appropriate traffic and execution controls across their interfaces and backend systems.

What happens when a person needs to take over?

Avaya Infinity coordinates AI and human expertise while preserving context. In Avaya’s research, 64% wanted automatic connection to a person, with their conversation and information carried over, after several unsuccessful AI attempts.  Another 18% wanted the AI to ask whether they would like to be connected with a person.

That finding argues for ending unproductive repetition. During a surge, escalation still needs staffing, prioritization, and a receiving employee who can understand what has happened.

How should enterprises test for a customer agent stampede?

Run a controlled simulation in a test environment around one delegated journey, such as checking an order or correcting a charge. Introduce a shared delay and vary agent behavior: some wait for an update, some retry, and others switch channels. Include customers who contact the company directly.

Measure whether the organization continues resolving authorized needs while protecting access to service:

Request amplification is a proposed diagnostic measure. Set the counting rules carefully and interpret it alongside outcomes. A lower ratio achieved by making help inaccessible would be a service failure.

Then test the changes: better progress information, controlled retries, safeguards against repeated execution, or a prepared human handoff. Identify which changes improve resolution under pressure.

Personal agents could give customers back time spent on routine service tasks. Enterprises can support that benefit by making delegated work understandable, governed, and executable, even when many customers need help together.

Explore the enterprise guide to AI agents acting for customers, read Customer Experience on Their Terms, and see how Avaya Infinity connects the systems and expertise behind the customer journey.

About the research

Avaya’s Customer Experience on Their Terms is part of the Avaya Connected Consumer Research Series. The online survey was fielded in September 2026 among U.S. adults aged 18 and older. All Avaya statistics cited in this article come from questions 21, 23, and 24, each with 558 respondents. The findings provide directional insight into preferences and expectations.