The new phase of AI in contact centers is all about control
Customer experience has become the most honest thermometer for any artificial intelligence strategy running in a real-world environment. And that thermometer is exactly what is forcing a major shift among companies using AI in their contact centers.
For years, the goal was simple: get the AI up and running as fast as possible. Show off automation, cut costs, present pretty numbers to the C-suite. But that cycle is over. Now, the questions on the table are much harder and much more important.
Is the AI actually working? For which customers? In which situations? And how would we even know if it wasn’t?
These questions have pushed two topics to the center of the debate: observability and governance. Not as isolated technical concepts buried in the IT team, but as real operational pillars with a direct impact on the people using the service every day.
What you will find here is an overview of how companies are tackling this new moment, what heavyweight experts are saying about it, and why controlling AI in production might be the most underrated competitive advantage in the industry right now. 🎯
The invisible problem nobody wanted to see
For a long time, contact centers treated artificial intelligence like an off-the-shelf solution. You hire the service, set up the flows, train the model on the company FAQs, and flip the switch. Simple enough, at least in theory. In practice, things played out very differently: the model responded well in expected scenarios but stumbled badly on anything that went off-script. And worst of all, nobody knew exactly when it was happening or why.
The customer would hang up frustrated, and the company would sit there staring at dashboards showing seemingly healthy resolution rates. It was a false sense of control that cost, and still costs, a fortune in terms of customer experience.
Michelle Brigman, contact center principal at Quantum Metric, sums up the transition nicely. According to her, a year or two ago, the pressure was to get the bot live, automate call types, and show cost reduction. Now the harder questions are showing up: is it working? For which customers? Under what conditions? And how would I know if it wasn’t?
Brigman adds that these questions are no longer about adoption, they are about control. And she makes an observation that hits the central point of the discussion dead-on: customers don’t experience your AI strategy, they experience a single interaction. And when that interaction produces a confidently wrong answer or a loop they can’t escape, they don’t blame the technology, they blame the brand.
This scenario became even more apparent as models grew more complex. With the arrival of large language models, response capability grew enormously, but predictability decreased at the same rate. A model that can chat about virtually any topic can also drift into inappropriate, inconsistent, or flat-out wrong answers, without a single alert going off on the operations dashboard.
Receive the best innovation content in your email.
All the news, tips, trends, and resources you're looking for, delivered to your inbox.
By subscribing to the newsletter, you agree to receive communications from Método Viral. We are committed to always protecting and respecting your privacy.
This is where the concept of observability enters with full force: the ability to see what is happening inside the system in real time, to understand the model’s behavior not just in success cases, but especially in cases of failure, drift, and ambiguity.
What sets observability apart from basic metric monitoring is exactly this depth. Monitoring is counting how many interactions were resolved. Observing is understanding how they were resolved, which paths the model took, where it hesitated, what it prioritized, and what it ignored. For contact centers, this represents a massive paradigm shift. It is no longer enough to know that the bot handled ten thousand calls. You need to know whether it handled them well, for whom, in which contexts, and with what consistency over time.
Scaling AI without losing the reins
Bhaskar Jayakrishnan, senior vice president of engineering and customer experience at Cisco, explains that there has been a clear shift: the focus has moved from simple deployment to operating at scale. According to him, as AI becomes embedded in real customer engagement, reliability and trust matter just as much as the model’s technical capability.
Jayakrishnan points out that running AI in dynamic, real-world customer service environments demands new disciplines. These disciplines include observability, rigorous testing, and clear governance to understand how systems behave in production. He warns that AI applications cannot run without a full set of observability and security built in, especially when they function as mission-critical infrastructure.
In his view, organizations need visibility, accountability, and human oversight baked into AI systems, ensuring they continue to meet business objectives and regulatory expectations. And when those guardrails are implemented the right way, they don’t limit innovation. On the contrary, they allow AI to scale responsibly and with confidence.
This point is essential because it debunks a myth that still circulates in many companies: the idea that control and speed are at odds. In practice, organizations that invest in robust control structures can actually scale faster, because they have the safety net to experiment, iterate, and course-correct quickly. Meanwhile, those operating without these structures end up paralyzed by the fear of breaking something nobody fully understands.
Governance is not bureaucracy, it is strategy
When the topic is governance of artificial intelligence, there is a natural resistance in many companies. The word carries the baggage of slow processes, committees, endless approvals, and documentation nobody reads. But that view is wrong, and more and more technology and operations leaders are realizing it.
Governance, in the context of AI in production, is essentially about defining who decides what, based on which data, and against which quality criteria. It is about creating a framework that allows AI to evolve without losing control over what it is doing and how it is impacting the business, especially the customer experience.
Suresh Kumar Bennet, executive vice president and global head of business process services at Hexaware, adds that governance is becoming increasingly integrated into day-to-day operations rather than being treated as a separate function. According to him, organizations need clear visibility into how AI systems make decisions, especially in customer-facing scenarios where accuracy, fairness, and transparency are critical.
Bennet further emphasizes that at the same time, it is necessary to ensure these systems consistently meet regulatory requirements and customer expectations around trust and experience. This is driving a more structured and proactive approach to oversight, with greater alignment between risk management, compliance frameworks, and real-time operational performance, rather than relying on periodic reviews or static policies.
In the most mature contact centers on the market, governance is already treated as part of the product, not as an extra layer added after the fact. This means that before any new model or feature goes into production, there is a clear set of criteria that must be met:
- Testing with adversarial scenarios
- Response validation across different user profiles
- Analysis of potential bias in responses
- Clear boundaries for what the AI can and cannot decide on its own
This is not paranoia, it is operational responsibility. And the companies doing this well are seeing concrete results: fewer unnecessary escalations, higher customer satisfaction, and a much greater ability to identify and fix problems before they turn into a crisis. 🔍
Another central element of governance is traceability. In a regulated environment like financial services or healthcare, for example, it is not enough for the AI to make the right decision. You need to be able to explain why it made that decision, based on what information, and following which rules. This has a technical name, explainability, and it is becoming a requirement that is not just regulatory but also commercial. Clients and business partners are beginning to demand transparency about how artificial intelligence systems operate. Ignoring this means leaving a massive vulnerability wide open in the business.
Operations redefined with AI at the core
The change happening is not just about adding new titles to the org chart. Bennet explains that the transformation goes deeper: it is about redefining how operations are run with AI embedded at the core of processes. While roles like AI operations leads and governance managers are emerging, what matters most is the shared accountability forming across data, technology, and operations teams around how AI performs in real-world environments.
According to Bennet, this includes accountability for outcomes, ensuring that systems don’t just respond correctly, but actually resolve customer needs. He adds that this requires tighter coordination between AI agents and human agents, with clear escalation paths, human-in-the-loop controls, and continuous feedback loops.
In Bennet’s view, responsibilities are evolving from simply deploying AI to actively managing and optimizing how work is done between AI systems and humans, with accountability tied to outcomes rather than just implementation.
Brigman adds to this perspective by noting that accountability for how AI performs in real-world customer service environments still tends to fall into a gap between IT, operations, digital, and the contact center itself. In her view, the shift that needs to happen is collective: every function that touches the customer journey has a stake in defining what production-ready means, and in demanding the visibility and controls to validate that on an ongoing basis, not just at launch.
Metrics that actually matter
Choosing the right metrics to evaluate whether an AI system is operating reliably is another topic that has gained prominence. Brigman suggests starting with failure demand: the percentage of contact volume that comes from customers who gave up because something didn’t work on their first attempt. If that number keeps climbing as AI handles more interactions, you are solving surface-level problems while making the deeper ones worse.
She also highlights repeat contacts for the same issue as a critical indicator. And she brings up journey completeness, meaning whether the customer actually managed to do what they came to do. It is possible to have a seemingly clean and efficient interaction but still have a customer who gave up halfway through. And that is where trust in the brand starts to erode.
Bennet adds that the most effective metrics connect system performance directly to customer outcomes, not just to the quality of the isolated interaction. He recommends that contact centers prioritize:
- Resolution rate, including both AI-contained and AI-assisted outcomes
- Time to resolution, to measure end-to-end effectiveness
- Containment and escalation rates, but only in the context of whether customer needs were actually resolved and how transitions to human agents were handled
- Experience metrics such as CSAT or customer effort, measured specifically for AI-driven interactions
According to Bennet, together these indicators provide a clearer picture of how reliably AI systems are resolving customer needs and improving overall operational performance.
Testing as a continuous discipline, not a checkpoint
Bennet makes a point of emphasizing that testing needs to become a continuous discipline, not just a pre-deployment step. According to him, companies should combine structured validation with ongoing monitoring of live interactions to identify gaps and refine system performance.
This includes using real-world data to stress-test systems, validate edge cases, and ensure consistent behavior over time. This approach allows organizations to evolve their AI systems in a controlled way, maintaining alignment with customer experience expectations and minimizing unintended risks.
In practice, this means that AI validation never ends. Every new interaction is a learning opportunity, and every behavioral drift that gets caught is a chance to adjust before the impact reaches the customer. Companies that adopt this mindset are creating a virtuous cycle where AI improves continuously, guided by real data rather than internal assumptions alone.
Why this matters for customer experience
At the end of the day, the entire discussion about observability and governance converges on a single point: the customer. Customer experience in a contact center is shaped by every interaction, and when artificial intelligence sits in the path of those interactions, it has the power to either elevate or wreck that experience depending on how it is controlled.
A customer who gets an inconsistent answer, has to repeat their request multiple times, or gets misrouted by the automated system is not necessarily going to complain about the AI, but they will complain about the company. And that negative perception builds up over time.
What the most successful organizations are learning is that treating customer experience as an input to AI decisions, rather than just an expected outcome, completely changes the dynamic. When you collect signals of dissatisfaction, hesitation, or confusion during interactions and use them to continuously adjust the model’s behavior, AI starts becoming genuinely more useful over time. This doesn’t happen by accident, and it doesn’t happen without observability. You need to see what is happening in order to make it better.
The good news is that this virtuous cycle, where governance and observability feed continuous improvement of the customer experience, is entirely achievable with the tools and methodologies available today. What is missing in most cases is not the technology, it is organizational commitment to this operating model.
Companies that understand this now and place these pillars at the center of their artificial intelligence strategy for contact centers will get ahead in a way that will be very hard to catch up to later. Getting AI live was just the beginning. Making it work well, consistently, and under real oversight is the game that is playing out right now. 🚀
