More AI Agents, More Agent Managers, But Still the Same Problem

This is Part five of Making Every Conversation Count, an ongoing video series featuring 8x8 CEO Sam Wilson and Liz Miller, VP and Principal Analyst at Constellation Research. Miller covers customer engagement, CX strategy, and AI-driven business outcomes.

An emergent AI trend of the last couple of years, to name one of many, is the proliferation of agents. Agents that function as receptionists, SDRs, schedulers, and support and data analytics (to arguably differing levels of success). For many organizations, the original plan was: "let's deploy a few smart agents to handle discrete tasks and do some of the more repetitive heavy lifting."

Fast forward to today, and Gartner predicts that by 2028, the average global Fortune 500 company will have over 150,000 agents in use, up from less than 15 in 2025. That kind of sprawl generates unprecedented IT complexity and adds agent wrangling efforts.

Just like with any growing workforce, adding an agent management layer feels like the intuitive next step. But I'm here to argue that the reflex to add yet another orchestration layer (without thinking through the details) will create more problems than it solves.

Agentic spread definitely needs to be managed, don't get me wrong. But the core problem I want to discuss here, in the context of customer experience, is that many agents who don't know anything about the customer can't hand things off cleanly, creating more work for the humans left holding the conversation.

That's an issue that can't be solved by adding a management layer. If agents are stateless, context-blind, and siloed, handoffs between agents (or people) are often where information disappears, forcing customers to repeat themselves and creating more problems than the agent's deployment ever solved.

They don't always know who the customer is, what they've already tried, or what was promised last time. Staff waste time untangling what the agent broke or missed. Or they get stuck managing the managers rather than doing work that creates tangible value.

AI Agent Governance Is Already Behindlink to this section

A different but no less important side of the proliferation issue is governance. According to Deloitte, only one in five organizations has a real AI governance model. That's despite the fact that 80% of them are also already deploying agentic AI.

Chances are, you have no idea if your agents are hallucinating, oversharing data, making unauthorized calls, or simply doing nothing useful.

At the shop-floor level, poor or nonexistent governance models can have a significant business impact. For example, if you don't have clear boundaries for agents that define which decisions they can make independently versus which require human approval, you've possibly handed off the power to make business-impacting decisions to a bot that wasn't created for that purpose.

Also consider that, without unified governance, there are no real-time monitoring systems to track agent behavior and flag anomalies, and no way to know whether the agent is misbehaving or malfunctioning.

Finally, many supposed agent managers lack audit trails that capture the full chain of agent actions, which helps ensure accountability. No trails means no way to identify where things have gone wrong in the event of a negative outcome, like a bad review or a churned customer.

Without proper governance, agents could be negatively impacting the customer journey or your bottom line, and no one would even know. More orchestration for its own sake can multiply ignorance rather than enhance knowledge. This is especially true if individual agents don't share context or customer status. At best, organizing and dispatching those agents more elegantly just routes ignorance more efficiently.

Start With the Customer, Not the Agent Countlink to this section

Here’s my advice: if you’re considering finding a solution for your heaps of multiplying agents:

1. Prune the rosebush. Be deliberate about what you choose to keep: use fewer, tie them more directly to deeper context, and make sure to create clean handoffs from the start rather than cobbled-together ones after the fact.
2. Think customer-first, not agent-first, and definitely not agent manager-first. Ask: Are these agents designed to benefit the customer, or were they created for some other purpose? If pressed, could we even name that purpose?
3. Management infrastructure should align with agent design choices while keeping purpose at the forefront. Creating sprawl and then solving for the management issue is way less effective than avoiding sprawl in the first place.

Sprawl is a symptom. The real question, the one worth asking before you add another layer, is whether your agents are actually serving the customer. That answer should drive everything else.

The principle worth building around: intelligence that runs on the same foundation as the conversation, not bolted on afterward. Context is kept, and governance spans agents and humans alike, built in from the start.

If it's not built on the conversation, it's not close enough to it to be useful.