How to Structure Human + AI Agent Teaming in Your Contact Centre

Matt Clare at UJET explores why the future of customer service depends on building smarter partnerships between AI and human agents, rather than replacing one with the other.

70.3% of companies use AI for customer interactions, with another 20.3% planning to adopt it, thus making customer/contact centre interactions one of the top deployment areas for AI.

But the most successful contact centres won’t be the ones deploying AI with the intent to replace humans, but the ones building the most intelligent division of labor between AI and human agents.

Metrigy’s AI for Business Success study of 697 companies found that AI-powered agent assist reduces Average Handle Time by 27.2% and improves customer ratings by 20.5% – but only when the AI layer is designed to augment human agents, not replace them.

Gartner predicts that by 2029, agentic AI will autonomously resolve 80% of routine customer service issues – which means the human interactions remaining will be disproportionately complex, emotional, and high-stakes.

The answer isn’t more AI or fewer humans. It’s smarter teaming.

This can be called the Layered Intelligence Model: a structured approach to assigning interactions based on complexity, emotional weight, and business risk – with a persistent AI layer supporting every human touchpoint throughout the interaction lifecycle.

Below is the operational playbook for building it.

Key Takeaways

• 70.3% of contact centres use AI, but AI-only conversations still fail to reach full resolution – adoption and effectiveness are not the same metric.
• The Layered Intelligence Model assigns interactions based on complexity, emotional weight, and business risk across three layers – and Layer 2, AI-augmented human resolution, is where most contact centres critically underinvest.
• Contextual continuity – the preservation of full interaction context across every AI-to-human handoff – is an architectural requirement, not a feature. It cannot be achieved on fragmented stacks.
• As AI handles more routine interactions, the remaining human interactions become more complex, more emotionally demanding, and more consequential. The human role is not shrinking – it is evolving.
• There are metrics that matter. Metrics such as resolution rate, CSAT, churn reduction, revenue contribution, customer lifetime value, etc. If you are only measuring deflection, you are optimizing for the wrong outcome.

Why Most AI Deployments Are Failing

The number is striking: 70.3% of contact centers now run some form of AI. And yet, if our experience as customers tell us anything, most AI-only conversations still end without full resolution.

This is a structural gap: organizations deployed AI as a layer on top of their existing contact center model – chatbots in front of queues, sentiment tools monitoring calls, knowledge base search embedded in agent desktops.

These tools made individual tasks faster. They did not change the fundamental architecture of how interactions are handled.

The result is a contact center that is simultaneously over-automated and under-resolved. AI handles the easy deflections.

Humans handle everything else. And the handoff between the two is manual, context-destroying, and frustrating for the customer.

The organizations pulling ahead have stopped thinking about AI and humans as separate channels. They have started thinking about them as a single, layered system.

Introducing the Layered Intelligence Model

The Layered Intelligence Model is a framework for structuring the relationship between AI and human agents in a modern contact centre. It has three layers:

Layer 1 – AI Resolution Layer

Handles interactions that are high-volume, low-complexity, and well-defined. Password resets, order status checks, appointment scheduling, basic billing inquiries.

Agentic AI resolves these end to end without human involvement. Containment rates of 60-90% are achievable on well-scoped use cases.

Layer 2 – AI-Augmented Human Layer

Handles interactions that require human judgment, empathy, or authority – but where AI dramatically improves human performance. The human agent leads the conversation.

AI provides real-time context, suggested responses, next-best-action guidance, and automated workflow execution in the background. The agent focuses on the relationship. The AI handles the systems.

Layer 3 – Human-Led Complex Resolution Layer

Handles interactions that are high-stakes, emotionally charged, legally sensitive, or genuinely novel. Escalations involving fraud, medical issues, executive complaints, or situations requiring creative problem-solving. AI supports with documentation and post-interaction summarization, but human judgment drives the outcome.

Layer 2 is Where Most Contact Centres Underinvest.

Organizations focus on Layer 1 automation targets and Layer 3 escalation protocols while leaving the middle layer – the majority of daily interactions – without a coherent AI strategy.

This is where the gap between AI adoption and AI effectiveness lives.

The AI layer should never fully disappear. Even in Layer 3, AI is pulling context, summarizing prior interactions, flagging sentiment shifts, and writing the case notes after. The human isn’t working alone. They’re working with an AI that’s already done the prep work.

Which Interactions Belong to AI and Which Belong to Humans?

The right assignment depends on three variables: interaction complexity, emotional weight, and business risk. Here is how those variables map across the most common contact center interaction types:

Interaction Type	Complexity	Emotional Weight	Business Risk	Layer
Password reset/account unlock	Low	Low	Low	Layer 1 – AI Transactional, no judgement required
Order status/shipment tracking	Low	Low	Low	Layer 1 – AI High Volume, zero ambuguity
FAQ/policy lookup	Low	Low	Low	Layer 1 – AI static knowledge, fast resolution
Appointment scheduling	Low	Low	Low	Layer 1 – AI rule-based, system-executable
Basic billing inquiry	Low	Low	Medium	Layer 1 -> 2 AI initiatives; escalates if customer pushes back
product troubleshooting (complex)	Medium	Medium	Medium	Layer 2 – AI + Human Multi-step; human judgement adds value
Complaint handling	Medium	High	Medium	Layer 2 – AI + Human emotional intelligence required
Cancellation/churn risk	Medium	High	High	Layer 2 – AI + Human Business risk; human retention value
Billing dispute	Medium	High	High	Layer 2 -> 3 Escalates based on amount or customer tone
Healthcare/sensitive personal data	High	Very High	Very High	Layer 3 – Human compliance + empathy requitements
Crisis/distress situation	High	Very High	Very High	Layer 3 – Human No AI should own this conversation
High-value relationship calls	High	High	Very High	Layer 3 – Human revenue and trust on the line
Novel/edge-case complaints	Variable	Variable	Variable	Layer 3 – Human AI has no reliable training data
Legal/regulatory inquiries	High	Medium	Very High	Layer 3 – Human liability; requires expert judgement

This table isn’t static. As AI systems mature and accumulate more interaction data, what belongs in Layer 1 today may confidently expand. What matters is having a principled framework for making that call – and revisiting it regularly.

The most common mistake: assigning Layer 2 interactions to Layer 1. Attempting full automation on interactions that carry emotional weight or business risk is what produces the chatbot frustration that drives customers to demand a human.

What Does the ‘Handoff’ Look Like in a Well-Built System?

Contextual continuity is the defining characteristic of a well-built human + AI system. It means that when an interaction moves from AI to human – or from human back to AI – no context is lost and no effort is repeated.

In a poorly designed system: the customer explains their issue to a chatbot, gets transferred, and explains the same issue again to a human agent who has no record of the prior conversation. This is the single most common source of customer frustration in AI-enabled contact centers.

In a well-designed system, the handoff works like this:

What Does the ‘Handoff’ Look Like in a Well-Built System?
Contextual continuity is the defining characteristic of a well-built human + AI system. It means that when an interaction moves from AI to human – or from human back to AI – no context is lost and no effort is repeated.

In a poorly designed system: the customer explains their issue to a chatbot, gets transferred, and explains the same issue again to a human agent who has no record of the prior conversation. This is the single most common source of customer frustration in AI-enabled contact centers.

In a well-designed system, the handoff works like this:

Before Escalation

The AI agent builds a real-time interaction summary – issue description, customer history, sentiment indicators, attempted resolutions, and recommended next steps – and delivers it to the human agent before they say hello.

During the Human Interaction

The AI layer stays active in the background. It surfaces relevant knowledge base articles, flags compliance requirements, suggests responses, and executes backend workflows as the human agent directs – without the agent navigating multiple systems.

After Resolution

The AI layer handles post-interaction work automatically – summarizing the conversation, updating CRM records, filing follow-up tasks, and flagging the interaction for quality review if relevant criteria are met.

Contextual continuity is not a feature. It is an architectural requirement. It requires a unified data layer that connects AI agents, human agents, CRM, and back-office systems in real time. Contact centers running on fragmented stacks cannot deliver it regardless of which AI tools they deploy on top.

The Klarna Lesson: What Happens When You Remove the Human Layer Entirely

In 2023, Klarna went all in on AI customer service – replacing 700 human agents with an AI chatbot that handled two-thirds of all customer inquiries. The efficiency numbers looked impressive. The customer experience numbers did not.

By 2025, CEO Sebastian Siemiatkowski acknowledged publicly in an interview with Bloomberg that cost had been “a too predominant evaluation factor” and that the AI-only approach had produced lower-quality service. Klarna began rehiring human agents.

As Siemiatkowski put it: “From a brand perspective, a company perspective, I just think it’s so critical that you are clear to your customer that there will always be a human if you want.”

The lesson is not that AI failed. Klarna’s AI still handles two-thirds of inquiries and has driven measurable improvements in response time. The lesson is that removing the human layer entirely – rather than restructuring it – produced a contact centre that was fast but not trusted.

The Layered Intelligence Model is not a path to a human-free contact center. It is a path to a contact center where every interaction is handled at the right layer.

Why Is the Human Role Actually Expanding, Not Shrinking?

This is the counterintuitive reality of agentic AI in the contact center: as AI handles more routine interactions, the average complexity and emotional weight of human-handled interactions increases.

When AI resolves 60-80% of routine inquiries autonomously, what remains for human agents is not the same work at lower volume.

It is harder work – more emotionally demanding, more complex, more consequential. Fraud investigations. Medical concerns. Executive escalations. Retention conversations with high-value customers.

This shift has two implications most contact center leaders haven’t fully absorbed:

Agent skill requirements are rising – The agents who thrive in an AI-augmented contact centre are not the ones who are fastest at navigating systems. They are the ones with the strongest communication skills, emotional intelligence, and judgment. The job is changing from information retrieval to relationship management.

Agent support requirements are rising with them – If human agents are handling the hardest interactions, they need better real-time support – not less. AI-powered coaching, real-time sentiment analysis, next-best-action guidance, and one-click workflow execution are not luxuries in this model. They are requirements.

The contact centers that understand this are investing in their human agents alongside their AI capabilities. The ones that don’t are discovering that automating the easy work while leaving humans unsupported on the hard work produces worse outcomes than either approach alone.

How to Know If Your Human + AI Balance Is Wrong

Six signals that your current model needs restructuring:

1. High AI Containment, Low Resolution

AI is capturing interactions it can’t finish. You’re deflecting, not resolving. This is the most common misalignment signal and the one most often hidden by containment rate reporting. If your CSAT is flat or declining while your containment rate is rising, this is the diagnosis.

2. Your Human Agents are Handling a High Volume of Repeat Contacts

Customers are returning because AI resolved the surface issue without addressing the root cause. This is a Layer 1 / Layer 2 assignment problem – interactions with underlying complexity are being routed to full automation when they need human judgment on the first contact.

3. Context is Lost at Every Handoff

Customers are re-explaining their issues when transferred. You do not have contextual continuity. This is an architecture problem, not an AI problem – and it cannot be fixed by swapping one AI vendor for another without addressing the underlying data layer.

4. Your Human Agents Report That AI Tools Add Work Rather Than Reduce It

Agents are navigating 4-6 applications before they can meaningfully engage with a customer. Your Layer 2 augmentation is informational rather than action-capable. Agents need click-to-execute workflow automation, not read-and-act recommendations.

5. Your AI Deployment is Uniform Across All Interaction Types

Every program or interaction type runs the same AI model with the same routing logic. You have not differentiated by complexity, emotional weight, or business risk.

You are applying Layer 1 logic to Layer 2 and Layer 3 interactions – and your resolution rates and CSAT scores on complex interactions will show it.

6. You Are Measuring Deflection Rate Instead of Resolution Rate

Deflection rate tells you how many customers didn’t reach a human. Resolution rate tells you how many customers got their problem solved.

These are not the same number – and organizations that optimize for deflection often find that CSAT and repeat contact rates move in the wrong direction simultaneously.

A quick note on Resolution Rate vs. Deflection Rate:

Deflection rate measures how many interactions AI kept away from human agents – it is a volume metric. Resolution rate measures how many customer issues were actually solved – it is an outcome metric.

A contact center can have a high deflection rate and a low resolution rate simultaneously, which means AI is capturing interactions it cannot complete.

Resolution rate is the metric that reflects actual customer experience quality. Deflection rate is an operational efficiency metric that tells you nothing about whether the customer’s problem was solved.

The Layered Intelligence Model is designed to maximize resolution, not containment.

This blog post has been re-published by kind permission of UJET – View the Original Article

For more information about UJET - visit the UJET Website

Call Centre Helper is not responsible for the content of these guest blog posts. The opinions expressed in this article are those of the author, and do not necessarily reflect those of Call Centre Helper.

Author: UJET
Reviewed by: Robyn Coppell