A report landed on my desk last week that stopped me cold. According to Redwood Software's October 2026 benchmarks study, 77% of enterprises have already experienced negative consequences from ungoverned AI agents. Unexpected cloud costs. Operational disruptions. Compliance issues.
The same research found that only 15% of contact centres successfully orchestrate agentic AI across customer touchpoints, while the rest struggle to show real business returns.
That's a sobering number for an industry that's spent the last two years chasing the agentic AI dream.
What went wrong?
The problem isn't enthusiasm. 89% of organisations increased their agentic AI deployment in the past year. 54% report they're at "scaling or mature" adoption levels.
The problem is infrastructure. 82% of respondents agreed that existing systems are bigger obstacles to scaling AI than a lack of strategy. Only 42% have a centralised registry of deployed agents. Just 51% have real-time monitoring of agent activity.
In other words: companies are deploying AI agents faster than they can govern them.
The orchestration gap
Here's the thing about agentic AI that gets lost in the excitement: an AI agent isn't particularly useful on its own. The value comes when multiple agents work together across a customer journey—handling intake, checking policies, updating systems, escalating when needed.
That coordination layer is called orchestration, and it's where most contact centre AI projects fall apart.
Consider a typical customer query about a billing discrepancy. An effective agentic system might involve:
- An intake agent that understands the query and pulls relevant account data
- A policy agent that checks what adjustments are permitted
- A billing agent that can actually make changes
- An escalation agent that knows when to involve a human
Without proper orchestration, these agents make conflicting decisions. One might approve a refund while another flags the same case for fraud review. The Redwood research found this exact pattern: agents doing their individual jobs correctly, but no coordinator ensuring they work together coherently.
What the 15% do differently
The contact centres that successfully orchestrate agentic AI share three characteristics:
They start with a single process, not a platform. Rather than deploying "AI agents" as a general capability, they pick one workflow—say, appointment rescheduling or payment arrangement—and build orchestration around that specific path. Only after proving it works do they expand.
They treat legacy integration as the main project. 62% of organisations pointed to better observability and real-time visibility as what would help them scale with confidence. 58% cited the ability to integrate AI agents with legacy systems. The contact centres getting results budgeted more time for connecting agents to existing CRM, billing, and telephony systems than for building the agents themselves.
They keep humans in the loop for consequential decisions. This isn't about distrust of AI—it's about governance. Every action that moves money, changes customer data, or represents the company externally gets routed through an approval step. The AI does the preparation; a human makes the final call.
The governance question
Gartner predicts that over 40% of agentic AI projects will be cancelled by 2027, primarily due to rising costs, unclear business value, and weak risk controls.
That last point—risk controls—is where contact centres have an advantage if they use it. Customer service operations already have compliance frameworks, quality monitoring, and escalation procedures. The organisations succeeding with agentic AI are adapting those existing governance structures rather than inventing new ones.
Practical steps that work:
- Scope each agent's permissions narrowly. An agent that can read customer data shouldn't automatically be able to write to billing systems.
- Log every agent action. When something goes wrong—and it will—you need to trace exactly what happened.
- Set hard limits on autonomous behaviour. Turn caps, spend limits, and mandatory escalation thresholds.
- Test with real cases, not demos. The demo always works. Production is where agent coordination breaks down.
Where this leaves contact centre leaders
The agentic AI opportunity is real. Automating multi-step customer journeys, reducing average handling time, improving first-contact resolution—these outcomes are achievable.
But the path there isn't "deploy agents and watch the magic happen." It's methodical work: choosing the right processes, building robust integration with existing systems, establishing governance before scaling, and measuring against clear baselines.
The 15% of contact centres getting this right aren't using different technology than everyone else. They're just being more disciplined about how they deploy it.
If you're planning an agentic AI initiative, start with one process, one journey, one measurable outcome. Prove orchestration works there before expanding. And budget at least as much time for integration and governance as for the agents themselves.
That's not exciting advice. But it's the advice that actually works.
Considering agentic AI for your contact centre? Hostcomm's Persona platform provides AI voice agents with built-in orchestration, governance, and integration capabilities designed for enterprise contact centre environments. Get in touch to discuss your specific requirements.