The new release allows AI agents to view the customer’s screen, interpret on-page elements, spot friction points, guide users with on-screen annotations, and hand off to human agents with complete contextual history.

It combines these capabilities with enterprise-grade redaction, auditing, and privacy controls, positioning the solution as a major leap in safe, context-driven AI support.

Indeed, Corbrowse believes that only after outlining these capabilities does the core problem become clear.

As Zac Scalzi, Director of Sales at Cobrowse, told CX Today:

“AI agents transformed how customers communicate, but they still lack the context required to actually solve problems. Until AI can see what the user sees, every answer is an educated guess.”

The “Context Gap” Holding AI Back

Large language models allow virtual agents to understand intent and deliver increasingly natural conversations.

But as Scalzi notes, “they still don’t see the actual user interface, what the user is doing, what errors are shown, or the UI obstacles encountered.”

This lack of grounding is what Cobrowse calls the “context gap”: the fundamental reason AI often sounds helpful yet fails to deliver meaningful resolution.

Customers end up repeating themselves, agents resort to guesswork, and support escalations pile up.

Cobrowse’s official product page argues that “LLMs can interpret and relay information, but without visual context they cannot reason. They behave like a searchable knowledge base, not an intelligent support agent.”

The vendor is emphatic that solving this gap is essential for businesses to thrive in the next era of agentic AI.

What Cobrowse AI Actually Brings to Virtual Agents

The new Cobrowse AI platform introduces capabilities traditionally reserved for human-assisted cobrowsing, but now fully integrated with automated support flows. These include:

Real-Time Visibility into UI State

Virtual agents can observe the customer’s web or mobile session, enabling them to identify errors, locate confusing elements, and understand exactly where a user is stuck.

Situation-Aware Guidance

Cobrowse notes that AI can now “visually direct customers with drawing and annotation tools,” giving step-by-step guidance instead of generic instructions.

Intelligent Analysis of Friction

The product interprets UI behavior and friction points in real time, giving AI agents the context needed to provide precise and timely instructions.

Seamless Escalation

If a case requires human intervention, the AI hands over with full visual and conversational history – eliminating the need for the customer to restate the issue.

Enterprise-Grade Safeguards

The platform includes redaction controls, audit logging, and deployment options tailored for regulated industries.

Scalzi described the solution’s ambition clearly, stating:

“Cobrowse AI elevates existing AI strategies by giving agents the context they need to reason. It shifts AI from relaying information to resolving issues autonomously.”

Why This Release Matters for the Broader CX Landscape

Even as enterprises invest heavily in AI assistants and copilots, many remain disappointed by low containment and inconsistent accuracy.

According to Scalzi, that frustration stems from over-reliance on data inputs that lack situational understanding.

“Most companies try to feed AI more information, such as FAQs, documentation, and logs, but without visual grounding, the AI is still guessing,” he said.

Teams often attempt to patch the issue by building custom APIs that expose product state to the AI. Yet, according to Cobrowse, these approaches are “engineering-intensive, fragile, and often introduce privacy risks.”

Cobrowse AI aims to eliminate these workarounds, giving virtual agents the context they need without custom engineering or risky integrations.

Expected Outcomes for Support Organizations

Like any AI solution, it all boils down to whether or not the tool can really deliver measurable results.

Cobrowse highlighted the following areas where it believes adopters can expect meaningful gains:

• Higher containment: More issues resolved entirely by AI.
• Greater accuracy and understanding: Virtual agents can interpret intent with UI awareness rather than assumptions.
• Improved CSAT: Customers experience interactions that feel relevant and confident.
• Higher FCR: AI agents can complete end-to-end resolutions rather than pushing users through multiple steps.
• Better digital adoption: Users learn the product as they’re guided through real workflows.

The company’s website claims that Cobrowse AI “gives your virtual agents the context they need to guide, resolve, and drive digital confidence.”

A Step Toward AI That Truly Understands Customers

Cobrowse’s latest release delivers something that conversational systems have historically lacked: the shared visual context that makes human-to-human support efficient and intuitive.

The company argues that this advancement is essential for agentic AI, as it enables the technology to not only speak like a human, but also think and respond like one.

In discussing the broader context of AI evolution, Scalzi summed up this point nicely:

“Without context, AI is little more than a smart FAQ. With visual intelligence, it can finally operate with real understanding.”

For organizations seeking to scale automation without sacrificing quality, this release may signal a new path forward.

It’s a fascinating concept. The idea isn’t to replace the human on the phone but to give them a teammate that actually does things.

Before we get into the details, if you are trying to keep track of everything happening this week, you can find our full AWS re:Invent 2025 Event Guide here and our full re:Invent hub of news here.

The headline news centers on ‘Agentic AI’—a term you are likely familiar with by now. Amazon Connect is rolling out 29 new capabilities designed to show it’s more than just a buzzword. Unlike the rather rigid chatbots that would get confused if you phrased a question the wrong way, these agents can reason, look up accounts, and process requests. The goal is to handle the drudgery, the notes, the summaries, the form-filling, so the human agent can focus on being, well, human.

“We’re now entering an era of agentic AI in Connect.” — Pasquale DeMaio

Here is what that actually looks like for the people doing the work.

Agents Get Instant Access to Enterprise Knowledge

There is nothing worse than being on the phone and not knowing the answer. It’s awkward for everyone.

To fix this, Amazon Connect is connecting its AI agents directly to enterprise knowledge bases via Amazon Bedrock. It means the AI can pull accurate answers instantly during a conversation.

They have also added support for the Model Context Protocol (MCP). It sounds technical, but it essentially means the AI can talk to other systems—like inventory databases or order management platforms—without a fuss. And for those already invested in the ecosystem, these AI features now extend seamlessly into the Salesforce Contact Center.

Amazon Connect Just Made Proactive Outreach Easier for Service Teams

Ideally, you fix the problem before the customer has to call you.

The new “Journeys” feature allows businesses to design multi-step experiences that adapt based on what the customer does. Combined with new predictive insights, the system can spot churn risks or purchasing interests and suggest reaching out proactively.

They have also added WhatsApp support for outbound campaigns. Given how much of the world lives on that app, it feels like a necessary addition.

Amazon Connect Delivers Complete Visibility for Teams Trusting Autonomous AI

Handing control over to an AI can feel a bit risky. To calm those nerves, Amazon Connect has introduced “enhanced observability.”

You can now see exactly why the AI made a decision, what tools it used, and how it got there. It provides a level of transparency that has been missing. They have also added tools to simulate thousands of interactions, so you can test how the AI behaves before you let it loose on real customers.

Global Brands Get Genuinely Human Voice Interactions

Robotic voices are usually a bit odd. They kill the mood.

Amazon Connect is launching “Nova Sonic” voices. These are designed to sound genuinely human, with the ability to handle interruptions gracefully and understand different accents.

If you prefer other flavors, they have also opened the platform to third-party speech tools like ElevenLabs and Deepgram. It gives businesses a choice, which is always nice.

• Read Amazon Nova Sonic: The End of the “Robot Pause” in CX?

Amazon Connect Removes Analytics Headaches for Managers with Natural Language Queries

If you have ever stared at a complex dashboard wondering why call volumes are spiking, this might appeal to you.

Amazon Connect is introducing an AI assistant for managers that lets you ask questions in plain English. You can simply ask, “Which agents need coaching on product knowledge?” or “What is causing the spike in call volume today?”

The AI digs through the data and gives you an answer. It removes the friction of needing to be a data scientist just to run a contact center, which seems like a rather sensible move.

What’s Next?

I began to wonder if we are moving toward a world where the “superhuman” agent is the standard, not the exception. It is a lot to digest.

We will be digging into this all week. Stay tuned for exclusive video interviews and a few more scoops from the event floor here on CX Today.

To kick things off, there are six predictions that all examine what might be the most tangible shift taking shape in customer experience: agentic AI moving from underwhelming chatbots into systems that can actually handle real work.

After years of disappointing automation projects, the technology has reached a point where self-service might finally live up to its billing.

Meet the experts:

• Simon Thorpe, Director of Global Product Marketing for Customer Service & Sales Automation at Pegasystems
• Kishan Chetan, EVP and GM of Agentforce Service at Salesforce
• Matt Price, CEO of Crescendo
• Hakob Astabatsyan, Co-Founder & CEO of Synthflow AI
• Matthias Goehler, CTO in the Europe region for Zendesk
• Zeus Kerravala, Principal Analyst at ZK Research

The Chatbot Problem Nobody Wants to Talk About

Simon Thorpe, Director of Global Product Marketing for Customer Service & Sales Automation at Pegasystems, isn’t mincing words about where self-service has been.

“Look, everyone is talking about AI right now. And for good reason,” he says.

“But the thing that I’m really excited about is the fact that we can finally deliver self-service that actually works for our customers. You know, self-service that can get real work done. It’s able to resolve issues, complete tasks, deflect work from our centers and it’s self-service that our customers are actively going to want to use.”

Customers actively wanting to use self-service has been the unicorn of CX for the better part of two decades. Chatbots and IVRs promised a lot but mostly delivered frustration.

Simple queries? Sure. Anything remotely complicated? Straight back to the queue.

Thorpe sees agentic AI changing that dynamic because it can reason, adapt, and understand natural language in ways that rigid scripting never could.

He explains “What once took months is now going to take weeks, which is tremendously exciting.”

But there’s a catch. Speed without structure creates chaos, particularly in regulated industries where processes can’t just be improvised by an AI agent with good intentions.

“Without governance and workflow or workflow backbone, AI agents can go rogue. They can ignore processes. They can introduce risks.”

His solution is pairing agentic AI with enterprise-grade workflows that act as guardrails, ensuring “your rules, your regulations, your standards are consistently applied every single time.”

AI Agents Move from Pilot Projects to Production

Kishan Chetan, Salesforce’s EVP and GM of Agentforce Service, believes 2026 is when AI agents move away from experimentation toward becoming infrastructure.

“For me, the CX prediction for next year, the biggest one, is far more mainstream of AI agents,” he says.

“Companies across the board will use AI agents in their customer experience, and they’ll use that for different processes, and that’ll work seamlessly with their human service reps.”

The emphasis on working alongside humans rather than replacing them reflects how the conversation around AI has matured. Early hype suggested automation would eliminate jobs.

The reality is messier and more interesting: AI handles volume and repetition, humans manage complexity and judgment.

When AI Outperforms the Average Agent

Matt Price, CEO of Crescendo, makes a prediction that’s bound to spark debate: AI agents will become more empathetic and more efficient than humans in 2026.

“On average across all of the interactions between service agents and AI, AI will perform better because on average, AI assistants are able now to have great language, detect tone and respond appropriately to customers in the moment and have full access to all of the information that they need in order to give customers what they want, which is an answer.”

Price isn’t suggesting every AI interaction will beat every human one. Top-tier agents will still outperform AI. But AI doesn’t have bad days, doesn’t forget context, and doesn’t struggle with tone on the 200th repetitive call of the day. That consistency matters.

There’s also a perception angle here, as Price notes that “a lot of the time for clients, it’s not necessarily just how well you serve them, but how much effort you put in.

“And there’s nothing better than showing the amount of effort that’s been put in, than putting a human in the loop rather than an AI agent.”

So even as AI gains emotional intelligence, there will still be moments where customers want to know a person is involved.

The Innovation Slowdown (That’s Actually Good News)

Hakob Astabatsyan, Co-Founder & CEO at Synthflow AI, predicts 2026 will see a decline in forward-looking innovation and instead focus on making AI work at scale.

“My prediction for 2026 is that we will be seeing less groundbreaking innovation that we have experienced in the last two years and more ROE and value delivery to the customers, to enterprises.

“What I mean by that is more scalable, more reliable platforms that allow the enterprises to go into production and deploy agents, voice agents, but also chat, omnichannel chat and text agents into production and scale them to millions of calls.”

That might sound boring compared to the breathless pace of the last couple of years, but it’s what enterprises actually need. Over-the-top innovations don’t matter if the technology can’t handle production traffic without breaking.

From Reactive to Proactive

Matthias Goehler, CTO in the Europe region for Zendesk, sees AI shifting from solving problems to preventing them before customers even notice.

“My biggest prediction for 26 when it comes to CX is that AI will move from automation to anticipation,” he says. “Instant resolution still remains the biggest expectation of customers. But on top of that, customers also more and more expect personalized engagement.

“And then even on top of that, if companies could start to become more proactive and reach out to customers instead of customers having always to reach out to companies, I think then we’re really talking about the gold standard in service.”

That’s a higher bar than most organizations have reached, but the technology to get there has begun its early stages.

Customers Will Actually Prefer Virtual Agents (For Simple Tasks)

Zeus Kerravala, Principal Analyst at ZK Research, predicts that for straightforward requests, customers will start choosing virtual agents over humans.

“My CX prediction for 2026 is that virtual agents get so good that for simple requests, people start to prefer the virtual agent over humans,” he says.

“And you might think that this is contrary to everything we believe, but if you look back at the early days of online banking and restaurant reservations online, people said that back then that no one would prefer a computer over a person. And in both cases that certainly wasn’t true.”

Kerravala draws a parallel to other initiatives that originally faced skepticism but eventually became preferred options once they proved faster and more reliable.

“Virtual agents can do things faster and more accurately than people now for complicated tasks. We’re still going to do prefer to a human, but in 2026 the quality of virtual agents will get so good that for simple tasks we’re going to prefer machines over people.”

The prediction doesn’t suggest humans become obsolete. It suggests customer preferences will align with the strengths of each channel.

What This Means for 2026

The common thread across these predictions is straightforward: AI agents are maturing from disappointing novelties into reliable tools that can handle real customer service work.

Self-service that actually works, agents that operate alongside humans without replacing them, and systems that anticipate problems rather than just reacting to them.

These aren’t abstract possibilities anymore. They’re becoming baseline expectations.

The recent CX Today report on AI in Customer Experience found 46% of businesses now invest more in AI for customer service than they do for sales, marketing, or commerce. Additionally, more than 90% of CX teams are planning on centralizing customer-facing solutions to ensure AI applications can influence every part of the customer journey.

The tools are already everywhere from virtual assistants, AI copilots, auto-reply apps, to smart routing systems. But customer satisfaction levels still aren’t improving as much as expected.

In fact, Forrester’s Customer Experience Index hit a record low in 2024. So, while AI is certainly changing how things get done, it’s still unclear whether it’s making things better for the people on the other end of the call. So, what can CX leaders learn from all this, and what’s next?

The Current State of Generative AI in the Contact Center

It’s wild to think about how quickly generative AI in the contact center went from interesting concept to “already everywhere”. Just two years ago, most contact centers were still playing around with basic chatbots. Now, nearly 80% have implemented some form of generative AI, from agent copilots to auto-summarization tools.

For most companies, the early go-to was AI copilots, built to sit alongside agents during live interactions. These tools help with drafting replies, pulling up relevant articles, and summarizing calls afterward. They’re fast, low-risk, and popular. 82% of CX teams say they’re already using some kind of copilot, and three out of four say it’s delivering real value.

But the tech is moving fast. Many companies are now experimenting with autonomous AI agents, tools that don’t just assist agents but actually handle full conversations on their own.

Some can answer customer questions using internal documentation. Others can triage issues and escalate them if needed.

Despite some very public AI misfires (remember the delivery bot that started swearing at customers?), the confidence in these tools is surprisingly high. According to the same CX Today report, 79% of CX leaders say they’d trust an AI agent to talk to customers without any prior training.

That’s interesting, when 66% of businesses still admit their customers prefer talking to humans over bots. And 61% of industry pros believe the government should step in and guarantee people the right to speak to a human if they want to.

Core Use Cases of Generative AI in the Contact Center

So, what’s generative AI actually being used for in contact centers today? Quite a lot. Some of the most common use cases, as mentioned above, focus on helping agents move faster, avoid admin work, and make fewer mistakes. The CX Today report highlights a few popular options:

Writing replies for agents

This is the big one. Half of all contact centers are using generative AI to draft replies for agents. The AI figures out the customer’s intent, pulls relevant info from internal systems, and writes a suggested message. Agents can tweak it or send it as is. It saves time and helps teams stay consistent in tone and language.

Auto-QA and coaching

Quality assurance is another go-to. Around 45% of teams use AI to automatically review conversations, flag good and bad moments, and even generate coaching tips. It’s a lot faster than handling manual reviews, and it’s helping managers spot trends they might miss otherwise.

Creating knowledge articles on the fly

Tools like NICE Enlighten AI are excellent for this. Instead of waiting for someone to write up a new help article, the AI listens to real conversations and creates articles based on what agents are already doing. Around 39% of teams are using this to keep their knowledge bases fresh and useful.

Post-call summaries and CRM updates

Fewer teams are using AI here (about 38%), but demand is still high. Instead of spending a few minutes writing a call summary, the AI does it instead. It fills out CRM fields, updates tags, and gets the agent ready for the next call in seconds.

Virtual assistants and copilots

Over 80% of contact centers now use some kind of copilot. These tools help agents find answers, suggest next steps, or guide them through processes. They’re everywhere, and for good reason: most teams say they’re working well.

Full-service AI agents

This is where generative AI in the contact center is really starting to evolve. Some businesses are testing fully autonomous agents that don’t just assist, they act as digital team members. They can answer customer questions, escalate complex cases, and even improve themselves over time by learning from past interactions.

Additional Use Cases

Of course, lots of other generative AI use cases are beginning to emerge too, particularly as AI continues to be baked into endless tools and platforms. Companies are exploring:

• Real-time language translation for global support teams.
• AI that understands sentiment and escalates emotional or frustrated customers faster.
• Bots that talk to other bots as “machine customers” become more common.
• Compliance-checking tools that flag risky language or policy violations.
• Predictive tools that spot early signs of churn and trigger personalized retention efforts.

Lessons to Learn from Generative AI in the Contact Center

There’s no shortage of excitement around generative AI in the contact center. But companies aren’t just focusing on the hype anymore, either. They’re paying attention to the reality. Now that the tech is living in thousands of contact centers, patterns are starting to emerge, and they’re not all good.

We’re learning that:

More AI doesn’t automatically mean better CX

Despite all the investment, customer satisfaction scores aren’t going up. At least not consistently. In fact, Forrester’s CX Index hit its lowest level ever in 2024. Throwing AI at the problem without a clear strategy doesn’t guarantee positive results.

A Gartner study found 64% of customers would actually prefer it if companies didn’t use AI in their service strategy at all. That doesn’t necessarily mean AI doesn’t have a place in the contact center, but it might not always need to be customer-facing.

Another major issue is that AI and humans aren’t always working together effectively. Bot-to-human handoffs are still messy. Some people still have to repeat themselves when they get to a live agent. If a bot hands over a conversation, the agent needs the full backstory, otherwise the customer experience falls apart.

Companies Need to Build the Right Foundations

For all the talk about AI transforming customer service, many contact centers are still skipping the groundwork. One common issue is rushed investments, often driven more by C-suite pressure than by actual customer needs. When the goal is just to cut costs or ride the AI trend, the resulting deployments tend to underdeliver or backfire.

But even well-intentioned projects can stumble without a clear understanding of customer demand. Many teams still don’t fully grasp why customers are reaching out. Without good journey mapping or conversation analytics, there’s a real risk of automating the wrong things.

On top of that, governments are starting to enforce guardrails. Spain’s three-minute response rule and California’s AI Act are just the beginning. Brands will soon need to prove that their AI systems are fair, accessible, and respectful of customer rights. Future-proofing means thinking about compliance from the very beginning, not scrambling after a policy change.

Get the Data Right, or Everything Falls Apart

There’s a saying in tech: garbage in, garbage out. And that applies heavily to generative AI in the contact center. Even the best generative AI won’t deliver results if it’s working with bad or incomplete data.

Outdated knowledge articles, siloed customer histories, and missing context can all lead to poor recommendations or wrong answers. That erodes both trust and efficiency.

This is why centralizing and enriching customer data has become a top priority. The best-performing contact centers are investing heavily in unified CRMs, real-time data pipelines, and knowledge management tools that keep everything current. Because when AI has access to the right data at the right moment, its value skyrockets.

Containment is a Flawed KPI

Many companies still measure the success of their generative AI bots by measuring containment rates, how many users stay with a bot and never reach a human. But containment doesn’t always mean that an issue was solved.

Contact centers need to back up their AI strategy with clear, business-relevant metrics. Think resolution rates, average handle time, deflection to human, customer satisfaction (CSAT), and even Net Promoter Score (NPS).

While high containment rates or automation percentages might look good in a dashboard, they rarely tell the full story. What actually matters is whether the customer walked away with their problem solved, and whether they’d come back again next time.

AI Is a Living System, Not a One-Time Project

One of the biggest misconceptions about AI is that it’s plug-and-play. That’s particularly true now that there are so many pre-built bots and AI solutions that seem so easy to use.

Realistically, though, generative AI in the contact center needs to evolve alongside the business. That means tracking performance, reviewing what’s working (and what’s not), and constantly refining prompts, workflows, and escalation rules.

The contact centers seeing the most success place their AI tools in a continuous improvement loop. They create an engine that gets better over time with the right inputs. It’s not just about the tech either. Cross-functional collaboration is also essential.

When service, marketing, sales, and IT teams align on data, goals, and customer experience design, AI deployments become more consistent, scalable, and effective. In some cases, organizations are creating new roles, like Chief Experience Officers, to keep everyone focused on the full customer journey, not just isolated fixes.

The Rise of Agentic AI: What’s Next?

Interest in generative AI in the contact center hasn’t disappeared completely, but business leaders are shifting their attention. While GenAI usually focuses on content, like writing replies, summarizing calls, and generating knowledge, the new era of AI focuses on action.

Agentic AI is stepping into the spotlight. Major companies, from Salesforce, to IBM, Microsoft, Zoom, and even Adobe are investing in a new era of flexible agents.

These agents don’t rely on humans for constant handholding and prompts. They follow multi-step workflows autonomously, adjust dynamically, make decisions based on context, and even access connected tools. Real world examples include:

• AI agents that handles a customer issue from start to finish, triaging the request, checking the knowledge base, resolving the problem, and closing the case.
• Sales-focused AI agents that joins calls with new reps, offers real-time coaching, and updates the CRM automatically afterward.
• Marketing agent that audits landing pages, suggests stronger calls to action, and A/B tests messaging, without human input.

These tools are becoming just as accessible as generative AI, with kits like Salesforce’s Agentforce, Microsoft Copilot Studio, and even Zoom’s customizable AI Companion.

Agentic AI will undoubtedly unlock new value for contact centers, but it also raises the stakes. With more autonomy comes a bigger need for transparency, explainability, and trust.

Looking Ahead: Preparing for the Next Age of AI

For companies investing in generative AI in the contact center or forward-thinking brands exploring agentic AI, there’s still a lot of work to be done. Every organization will need to clean up its data strategy, adjust how it monitors metrics and KPIs, and think carefully about how it will manage human-AI collaboration going forward.

GenAI definitely brought both innovation and disruption to the contact center. Now, companies have numerous lessons to learn as they move forward into the next age of autonomous agents and intelligent growth. The businesses that thrive will be the ones that don’t just focus on reducing costs or speeding up tasks but use AI to enhance customer experiences.

As agentic AI takes hold, the expectations will rise. Customers will want fast, smart answers, but still expect empathy and control. Agents will rely on AI more heavily, but still need tools they can trust. And CX leaders will need to justify not just the cost of AI, but its long-term value.

The communications platform disclosed its Q3 earnings on Monday, highlighting a strong growth from its customer experience portfolio. 

Zoom has also revealed its plans to grow product revenue further by enhancing its existing products with additional AI capabilities to drive AI-first customer experiences. 

During the earnings call, Zoom announced that the platform would be evolving from its traditional customer experience platform to an AI-focused one, aiming to drive productivity and relationships. 

Eric Yuan, CEO and Founder of Zoom, revealed that after its strong quarterly results, Zoom would be able to move forward with this vision. 

He said: “This performance reflects the durability of our business driven by the growing value we are delivering for customers as we evolve from a communications leader to an AI-first platform for work and customer experience. 

“Our vision is to be the AI-first work platform for human connection.” 

Zoom expects to accomplish this transformation by following its three strategic priorities: enhancing its existing products with AI, driving growth in AI products, and scaling AI-first customer experiences. 

Enhancing Existing Products 

During Zoomtopia 2025, the communications platform unveiled AI Companion 3.0, an updated version of AI Companion that utilizes agentic AI not only to respond, but also to act, advising on tasks such as meeting preparations, freeing up time, and call follow-ups. 

Zoom has embedded various AI capabilities and tools, including AI Companion, across its platform foundation, including: 

• Zoom Meetings: Zoom’s AI Companion, a proactive AI assistant tool, offers meeting summaries, follow-ups for next steps, and drives work forward. 
• Team Chat: Rising by 20% in active monthly users year over year, AI Companion supports the messaging product by providing customers with chat summaries, composition tools, and simplified search options for higher productivity. 
• Zoom Phone: This tool now offers Voice Intelligence for call transcription, summaries, noise cancellation, call routing, and analytics and insights for customer data collection, with over 10 million users now paying for Zoom Phone as of early Q3. 
• Zoom Contact Center: Working as Zoom’s cloud-based contact center solution, this platform has adopted AI tools such as Virtual Agent, an agentic AI chatbot offering complex tasks and responses for customers, and AI Expert Assist, allowing agents to utilize AI support in real-time with summaries and translations and offer possible agent responses during customer interactions. 

In fact, AI Companion usage has grown four times year-on-year, revealing that these AI features are seeing value from user activity, resulting in rapid adoption. 

By adding AI to these already-established products, customers are more likely to accept these capabilities once they’ve been integrated into the software. 

Driving Growth in AI Products 

By moving beyond its core communication tools and investing in greater agentic abilities, Zoom offers its customers further access to its AI tools to personalize them to their needs. 

This allows Zoom the chance to drive AI product revenue with product monetization, generating financial growth rather than just adding tools to products. 

In fact, 90% of Zoom’s top CX deals involve paid AI features to contribute to product revenue, offering both subscription and consumption models to suit the customer. 

This includes the development of AI tools such as Custom AI Companion, a paid version of the standard AI Companion model targeted towards enterprise-tier customers, allowing businesses to customize the tool to meet specific demands and policies. 

This also includes similar products such as Virtual Agent and AI Expert Assist, as well as Zoom’s recent acquisition of BrightHire. 

Scaling AI-First Customer Experiences 

Through utilizing tools such as Virtual Agent and AI Expert Assist, Zoom is using AI to transform interactions between customers and enterprises by expanding these products across the platform for automated workflows. 

These tools will involve automating routine requests and advise agents during workflow automation, voice, chat, and video calls for faster results. 

Zoom has also implemented a feature that allows enterprises to install either Zoom’s or a third-party’s AI tool, encouraging them to become familiar with AI usage while tailoring it to their needs. 

This strategy will also involve Zoom working with its largest customers to move AI agents into deployment; however, this may prove difficult. 

During the earnings call, Zoom noted that despite this upsurge in AI tool adoption, its net dollar expansion rate stayed at 98%, 2% lower than expected, likely suggesting that large customers had not been spending as much as hoped on Zoom’s products, with renewals on larger accounts proving difficult to resume. 

Zoom Key Earnings Results 

Zoom’s earnings results showed some strong areas of performance across enterprise and cashflow revenue results 

• Zoom’s total revenue reached $1.23BN, up 4.4% year-on-year 

• Its enterprise revenue grew 6.1%, totalling 60% of Zoom’s total revenue 

• Average monthly churn increased by 2.7%, similar to Q3 2024 

• Its operating cash flow increased to $629MN, up 30% year-on-year 

For many enterprises, the toughest part of navigating the agentic AI wave is determining whether the technology actually solves a real problem. Danny Gunn, Head of Workforce Planning at Bet365, put it candidly:

“Part of the understanding of AI is the use case. It may sound great from the sales pitch, but does it actually work? There are quite a few [solutions] where we tried them and they don’t actually work, whether that’s because we’re not ready and our backend processes can’t use all of that or the technology isn’t quite as good as the sales pitch that we get to see.”

For organizations operating under financial constraints, the hype can create tension to turn to AI as a cure-all to deliver cost savings. Kim Baker, Head of Operational Support Services at UK housing association L&Q, noted that leaders are under “huge pressure to not spend too much money and save as much as we can.”

“Everyone just says AI as if it’s now the answer to everything, but I don’t think people really fully understand what AI is and what it might be able to do for them.”

Baker added a critical reminder that any organization considering agentic AI for automation needs to address “simple truths” before jumping on the bandwagon:

“There’s no point launching AI if your data is not right in the first place, because where’s it going to look to answer these questions?”

Without reliable data, even the most advanced agentic AI implementation will deliver inconsistent results, and undermine trust in the technology.

Understanding Where AI Truly Adds Value

The pressure to “have AI” is evident across the industry, often overshadowing the need for alignment with real organizational challenges. Keith Griffin, Cisco Fellow VP, noted how easily organizations default to AI without planning or frameworks: “It is very much about ‘we need to have some AI capability’ but not think deeply about where there’s evidence of where it [gets results] and some of the reasons why AI adoption scores.”

“Mismatched use cases, expecting AI to do things that it’s not very good at, or assuming that it can do more than possible,” all result in failed implementations, Griffin added. “People are getting caught up with which AI models should be used, and it really doesn’t matter… as long as it’s safe to use and an appropriate use for the organization.”

Chris Rainsforth, Director of Learning & Innovation at contact center industry body The Forum, noted that AI has become “a catch all” for any operational challenge and highlighted the pitfall of rushed deployments:

“What we’ve seen a lot of examples of, unfortunately… people trying to deploy something without understanding the problem they’re trying to solve. In the first instance, they spend a lot of money, they spend a lot of time, spend a lot of effort doing something, and then it doesn’t get the results.”

Leadership often grows frustrated when a costly AI project fails to deliver results, questioning why the investment isn’t paying off. Untangling those underlying issues then becomes a difficult and time-consuming process.

But encouragingly, more organizations are beginning to pause and reassess, Rainsforth said.

“On the flip side, we are starting to see more people take a more considered approach, going, ‘what are the outcomes? What am I trying to solve? Let’s then work back from that to understand what technology can enable us to deliver it.’ And AI might not be the answer to every problem. It might be something else.”

“People are starting to have those conversations be a bit more kind of thoughtful about that approach, rather than just wasting time and effort and money,” Rainsforth said.

Putting the Customer First in Tech Decisions

Ultimately, when leaders pay close attention to what their customers need, rather than what the market is hyping, they gain a clearer sense of which tools will genuinely improve experiences and which investments aren’t worth pursuing, several speakers emphasized.

Listening to customers provides the grounding needed to make purposeful, informed choices about where and how to deploy new technology, whether that’s agentic AI or other systems.

Frontline experience shows that customers don’t tend to share the industry’s fixation on AI-driven speed and automation; they’re simply looking for problems to be solved efficiently. As David Holmes, Director of Sales at UK utility SSE observed:

“I don’t have customers tell me, ‘I hope you hurry up with that AI’. I don’t have customers saying, ‘I hope you handle my call quicker’; customers care about the resolution, they care about the time on the phone. They do care about simplicity and I think most sales could benefit from all of that, and that’s where technology can help.”

Across discussions on the show floor, that theme consistently resurfaced. The path to meaningful AI adoption starts with understanding customer needs. When enterprises anchor technology decisions in real-world pain points rather than hype cycles, they’re more likely to avoid missteps and deliver measurable improvements.

With global enterprises like BT, Deutsche Telekom, and EE already seeing results, the discussion explores how hybrid human-AI teams, composable experience design, and data-driven automation are shaping the contact center of 2026 — and how a unified CX platform can future-proof operations and drive measurable business value.

In this CX Today interview, Rob Scott sits down with Shrenik Jain, VP of Product Management for CCaaS at Sprinklr, to explore what’s next for contact centers and how Sprinklr is taking a platform-first approach to transform customer experience.

Jain explains why Sprinklr didn’t follow the legacy voice-first route, how unification enables smarter AI, and what the shift from assistant AI to agentic AI means for tomorrow’s CX workforce. If you’re rethinking your contact center strategy for 2026, this one’s worth a watch.

Key Discussion Points:

• The real reason CCaaS is saturated: Why traditional cloud migration is no longer enough, and what customers now expect.
• Connected intelligence over channel sprawl: How Sprinklr integrates social, voice, messaging, and AI into a single platform.
• The rise of agentic AI: What moving from reactive assistants to autonomous agents means for skills, roles, and workforce planning.
• Actionable insights, not just dashboards: How AI is shifting from descriptive to prescriptive, and even autonomous decision-making.

To explore Sprinklr’s unified CXM platform and how it enables smarter, scalable, AI-powered customer engagement, visit sprinklr.com

The problem usually isn’t the model. It’s the data behind it. When knowledge bases are out of date, fragmented, or inconsistent, even the smartest AI will confidently generate the wrong answer. This is why knowledge base integrity and RAG governance matter more than model size or speed.

The urgency is clear. McKinsey reports that almost all companies are using AI, but only 1% feel they’re at maturity. Many also admit that accuracy and trust are still major barriers. In customer experience, where loyalty is fragile, a single hallucination can trigger churn, compliance headaches, and reputational fallout.

Leading enterprises are starting to treat hallucinations as a governance problem, not a technical one. Without governed data, AI becomes a liability in CX. With it, organizations can build automation that actually strengthens trust.

What Are AI Hallucinations and What Causes Them?

When customer-facing AI goes off-script, it usually isn’t because the model suddenly turned unreliable. AI hallucinations in CX happen when the system fills gaps left by bad or missing data. Picture a bot telling a customer they qualify for same-day refunds when the actual policy is 30 days. That’s not creativity, it’s a broken knowledge base.

Hallucinations tend to creep in when:

• Knowledge bases are outdated or inconsistent, with different “truths” stored across systems.
• Context is missing, for example, an AI forgetting a customer’s purchase history mid-conversation.
• Validation checks are skipped, so the bot never confirms whether the answer is still correct.

The risks aren’t small. 80% of enterprises cite bias, explainability, or trust as barriers to using AI at scale. In CX, inaccuracy quickly turns into churn, complaints, or compliance headaches.

There are proven fixes. Enterprises just need to know what to implement before they go all-in on agentifying the contact center.

The Real-World Impact of AI Hallucinations in CX

The stakes around AI hallucinations in CX translate directly into lost revenue, churn, and regulatory risk. A bot that invents refund rules or misstates eligibility for a benefit doesn’t just frustrate a customer – it creates liability.

Some of the impacts seen across industries:

• Retail: Misleading warranty responses trigger unnecessary refunds and drive shoppers to competitors.
• Public sector: Incorrect entitlement checks leave citizens without services they qualify for.
• Travel: Fabricated policy details can mean denied boarding or stranded passengers.

The financial burden is real. Industry analysts estimate that bad data costs businesses trillions globally each year, and the average cost of a single data-driven error can run into millions once churn and remediation are factored in.

Case studies show the impact, too. Just look at all the stories about ChatGPT, creating fictitious documents for lawyers, or making up statements about teacher actions in education. Every hallucination is a reminder: without knowledge base integrity and RAG governance, automation introduces more risk than reward. With them, AI becomes a growth driver instead of a liability.

Why Hallucinations Are Really a Data Integrity Problem

It’s tempting to think of AI hallucinations in CX as model failures. In reality, they’re usually symptoms of poor data integrity. When the information feeding an AI is out of date, inconsistent, or fragmented, the system will confidently generate the wrong answer.

Knowledge base integrity means more than just storing information. It’s about ensuring accuracy, consistency, and governance across every touchpoint. Without that, CX automation is built on sand.

Common breakdowns include:

• Outdated articles: A policy change goes live, but the bot still cites the old rules.
• Conflicting records: Multiple “truths” for the same customer, leading to contradictory answers.
• Ungoverned logs: Data pulled in without privacy controls, creating compliance exposure.

Some organizations are already proving the value of treating hallucinations as governance problems. Adobe Population Health saved $800,000 annually by enforcing stronger data controls, ensuring agents and AI systems pulled only from validated knowledge sources.

Building the Foundation: Clean, Cohesive Knowledge

Solving AI hallucinations in CX starts with building a solid data foundation. No model, no matter how advanced, can perform reliably without knowledge base integrity. That means every system, from the CRM and contact center platform to the CDP – has to point to the same version of the truth.

A few steps make the difference:

• Unified profiles: Use CDP to connect IDs, preferences, and history across systems. Vodafone recently reported a 30% boost in engagement after investing in unified profiles and data quality.
• Agent-ready records: Golden IDs, schema alignment, and deduplication stop bots from improvising. Service accuracy depends on knowing which record is the right one.
• Data freshness: Expired knowledge is one of the fastest routes to hallucination. Setting SLAs for update frequency ensures AI doesn’t serve answers that are weeks, or years, out of date.
• Governance layers: Microsoft’s Purview DLP and DSPM frameworks, for example, help enforce privacy boundaries and ensure sensitive data is never exposed to customer-facing AI.

Clean, governed data is what allows automation to scale safely. In fact, Gartner notes that automation without unified data pipelines is one of the leading causes of failure in AI deployments.

The lesson is clear: AI only works if the underlying knowledge is accurate and consistent. RAG governance begins not at the model layer, but in how enterprises treat their data.

Choosing Your LLM Carefully: Size Isn’t Everything

When automating CX workflows, the assumption that “bigger means better” often backfires. In fact, purpose-built, smaller language models can outperform broad, heavyweight counterparts, especially when they’re trained for specific customer service tasks.

Here’s what’s working:

• Smaller, tailored models excel at soft-skill evaluations. In contact center hiring, they outperform general-purpose LLMs simply because they understand the nuances of human interaction better.
• Efficiency is a major advantage. Smaller models require fewer computational resources, process faster, and cost less to run, making them ideal for real-time CX workflows.
• They also tend to hallucinate less. Because they’re fine-tuned on targeted data, they stay focused on relevant knowledge and avoid the “overconfident bluffing” larger models can fall into.
• Distillation, teaching a smaller model to mimic a larger “teacher”, is now a common technique. It delivers much of the performance without the infrastructure cost.

Choosing the right model is a strategic decision: smaller, domain-specific models support RAG governance and knowledge base integrity more effectively, without blowing your budget or opening new risks.

RAG Governance: Why Retrieval Can Fail Without It

Retrieval-augmented generation (RAG) has become a go-to strategy for tackling AI hallucinations in CX. Companies like PolyAI are already using RAG to make voice agents check against validated knowledge before replying, cutting down hallucinations dramatically.

Instead of relying only on the model’s training data, RAG pulls answers from a knowledge base in real time. In theory, it keeps responses grounded. In practice, without proper RAG governance, it can still go wrong.

The risks are straightforward:

• If the knowledge base is outdated, RAG just retrieves the wrong answer faster.
• If content is unstructured, like PDFs, duplicate docs, or inconsistent schemas, the model struggles to pull reliable context.
• If version control is missing, customers may get different answers depending on which copy the system accessed.

That’s why knowledge base integrity is critical. Enterprises are starting to use semantic chunking, version-controlled KBs, and graph-RAG approaches to make sure AI agents retrieve the right data, in the right context, every time.

Vendors are also moving quickly. Google Vertex Agent Builder, Microsoft Copilot Studio’s RAG connectors, and open-source projects like Rasa’s extensions are designed to enforce cleaner retrieval pipelines. Companies like Ada are proving that governed RAG can cut down false answers in sensitive workflows like background checks.

RAG is powerful, but without governance, it risks becoming a faster way to spread bad information. Grounding AI in trusted, validated sources, through structured retrieval and strong RAG governance, is the difference between automation that builds trust and automation that erodes it.

The Model Context Protocol for reducing AI hallucination

Even with RAG governance, there’s still a missing piece: how the model itself connects to external tools and data. That’s where the Model Context Protocol (MCP) comes in. MCP is emerging as a standard that formalizes how AI systems request and consume knowledge, adding a layer of compliance and control that CX leaders have been waiting for.

Without MCP, connectors can still pull in unreliable or non-compliant data. With MCP, rules can be enforced before the model ever sees the input. That means:

• Version control: AI agents only access the latest, approved policies.
• Schema validation: Data must meet format and quality checks before it’s used.
• Integrity enforcement: Broken or incomplete records are automatically rejected.

This is particularly relevant in regulated industries. Financial services, healthcare, and the public sector can’t risk AI fabricating eligibility or compliance-related answers. MCP provides a structured way to prove governance at the system level.

Vendors are already moving in this direction. Salesforce’s Agentforce 3 announcement positioned governance and compliance as central to its next-generation agent framework. For CX leaders, MCP could become the difference between AI that “sounds right” and AI that is provably compliant.

Smarter Prompting: Designing Agents to Think in Steps

Even with clean data and strong RAG governance, AI hallucinations in CX can still happen if the model is prompted poorly. The someone asks a question shapes the quality of the answer. That’s where smarter prompting techniques come in.

One of the most effective is chain-of-thought reasoning. Instead of pushing the model to jump straight to an answer, prompts guide it to reason through the steps. For example, in a travel entitlement check, the AI might be told to:

• Confirm eligibility rules.
• Check dates against the customer record.
• Validate exceptions before giving a final response.

This structured approach reduces the chance of the AI skipping logic or inventing details to “sound confident.”

Other strategies include:

• Context restating: Have the model summarize customer inputs before answering, to avoid missing key details.
• Instruction layering: Embedding guard phrases like “If unsure, escalate” directly into prompts.

Better prompting changes how the AI reasons. Combined with knowledge base integrity and retrieval grounding, thoughtful prompt design is one of the simplest, most cost-effective ways to cut hallucinations before they ever reach a customer.

Keeping Humans in the Loop: Where Autonomy Should Stop

AI is getting better at handling customer requests, but it shouldn’t be left to run everything on its own. In CX, the cost of a wrong answer can be far bigger than a frustrated caller. A single AI hallucination in CX around something like a loan decision, a medical entitlement, or a refund policy can create compliance risks and damage trust.

That’s why most successful deployments still keep people in the loop. Routine questions like order status, password resets, and warranty lookups are safe to automate. But when the stakes rise, the system needs a clear off-ramp to a human; no company should try to aim for limitless automation.

There are simple ways to design for this:

• Flagging low-confidence answers so they’re routed to an agent.
• Escalating automatically when rules aren’t clear or when exceptions apply.
• Training models with reinforcement from human feedback so they learn when to stop guessing.

Real-world examples prove the value. Ada’s work with Life360 showed that giving AI responsibility for repetitive queries freed agents to focus on tougher cases. Customers got faster answers when it mattered most, without losing the reassurance of human judgment for sensitive issues.

The lesson is straightforward: automation should extend, not replace, human service.

Guardrail Systems: Preventing AI hallucination

AI can be fast, but it still needs limits. In customer service, those limits are guardrails. They stop automation from giving answers it shouldn’t, even when the data looks clean. Without them, AI hallucinations in CX can slip through and cause real damage.

Guardrails take different forms. Some block responses if the system isn’t confident enough. Others make sure refund rules, discounts, or eligibility checks stay within company policy. Many firms now add filters that catch bias or toxic language before it reaches a customer.

The goal isn’t perfection. It’s layers of protection. If one check misses an error, another is there to catch it. Tucan.ai showed how this works in practice. By adding guardrails to its contract analysis tools, it cut the risk of misinterpreted clauses while still saving clients time.

For CX teams, guardrails aren’t about slowing automation down. They’re about trust. Customers need to know that the answers they get are safe even when they come from a machine.

Testing, Monitoring, and Iterating

AI systems drift. Policies change, data updates, and customer expectations move quickly. Without regular checks, those shifts turn into AI hallucinations in CX.

Strong CX teams treat testing and monitoring as part of daily operations. That means:

• Running “red team” prompts to see how an agent handles edge cases.
• Tracking hallucination rates over time instead of waiting for customer complaints.
• Comparing different prompts or retrieval methods to see which reduces errors.

Enterprises are starting to put this discipline into place. Retell AI cut false positives by 70% through systematic testing and feedback loops. Microsoft and others now offer dashboards that log how models use data, making it easier to spot problems early.

The principle is straightforward. AI is not a one-off project. It’s a system that needs continuous oversight, just like a contact center workforce. Test it, measure it, refine it.

The Future of AI Hallucinations in CX

Customer experience is moving into a new phase. Contact centers are no longer testing basic chatbots. They are rolling out autonomous agents that can manage full interactions, from checking an order to triggering a refund. Microsoft’s Intent Agent, NICE’s CXone Mpower, and Genesys’ AI Studio are early examples of that shift.

The upside is clear: faster service, lower costs, and better coordination across systems. The risk is also higher. A single AI hallucination in CX could mean a compliance breach or a reputational hit that takes years to repair. Regulators are watching closely. The EU AI Act and ISO/IEC 42001 both push for stricter rules on governance, transparency, and accountability.

The market is responding. Salesforce’s move to acquire Convergence.ai and NiCE’s purchase of Cognigy show how major vendors are racing to build platforms where governance is built in, not added later. Enterprises want systems that are safe to scale, not pilots that collapse under risk.

The reality is that hallucinations won’t disappear. Companies will need to learn how to contain them. A strong knowledge base, integrity, tight RAG governance, and frameworks like MCP will differentiate brands that customers trust from those they don’t.

Eliminating AI Hallucinations in CX

The risk of AI hallucinations in CX is not going away. As enterprises scale automation, the cost of a wrong answer grows, whether that’s a compliance breach, a lost customer, or a dent in brand trust.

The good news is that hallucinations are not an unsolvable problem. They’re usually data problems. With a strong knowledge base, integrity, clear RAG governance, and frameworks like MCP to enforce compliance, organizations can keep automation reliable and safe. Guardrails, smarter prompting, and human oversight add further protection.

Together, these measures turn AI from a liability into an asset. Companies that treat governance as central will be able to roll out advanced agents with confidence. Those that don’t risk being left behind.

But customers? They’re not exactly throwing confetti about it.

In fact, Gartner found 64% of people would rather companies didn’t use AI for service, full stop. Plus, only 24% of customers trust AI with anything messy, like complaints, policy decisions, those emotionally-charged “your system charged me twice” moments. So there’s this weird split: companies rushing forward, customers dragging their heels, and everyone quietly hoping the bots behave.

That’s the real issue, honestly. AI doesn’t warn you it’s going wrong with an error code, at least not all the time. It goes sideways in behavior. A chatbot invents a refund rule. A voice assistant snaps at a vulnerable caller. A CRM-embedded agent quietly mislabels half your complaints as “general enquiries.”

This is why AI behavior monitoring and real AI governance oversight are really becoming the only guardrails between scaling CX AI, and watching it drift into places you really don’t want headlines about.

Why AI Governance Oversight Is Critical for CX Success

Most CX teams think they’re rolling out “smart automation,” but what they’re actually doing is handing decision-making power to systems they don’t fully understand yet. That’s just where the industry is right now. The tech moved faster than the manuals.

This is exactly why AI behavior monitoring, AI governance oversight, and all the messy parts of CX AI oversight are suddenly showing up in board conversations. It’s pattern recognition. The problems are becoming glaringly obvious. We’ve all seen a bot make a weird decision and thought, “Wait… why did it do that?”

Ultimately, we’re starting to bump against a very real trust ceiling with AI and automation in CX.

KPMG’s global study found 83% of people expect AI to deliver benefits, but more than half still don’t trust it, especially in markets that have seen its failures up close.

Unfortunately, business leaders aren’t making it easier to trust these systems either.

Here’s where things get dicey. PwC’s 2025 research shows only a small fraction of companies feel “very effective” at AI risk monitoring or maintaining an inventory of their AI systems. That’s not just making customers skeptical, it’s opening the door to countless problems with security, data governance, and even AI compliance.

What Off-the-Rails AI Looks like in CX

It’s funny, when people talk about AI risk, they usually imagine some Terminator-style meltdown. In reality, CX AI goes off the rails in more subtle ways:

Hallucinations & fabricated information

Hallucinations sound like this mystical AI thing, until your bot confidently invents a cancellation policy that’s never existed and suddenly, you’re handing out refunds like coupons.

2025 observability research keeps pointing to the same pattern: hallucinations usually come from messy or contradictory knowledge bases, not the model itself. A tiny change in wording, an outdated policy page, and suddenly the AI “helpfully” fills in the blanks.

This is where AI drift detection becomes so important. Hallucinations often creep in after small updates to data pipelines, not major system changes.

Tone errors, “cold automation” & empathy failures

Efficiency without empathy doesn’t win customers.

Brands aren’t losing customers because AI is wrong, they’re losing them because the AI feels cold. It encourages negative response. Research found 42% of Brits admit they’re ruder to chatbots than humans, and 40% would pay extra just to talk to a real person during a stressful moment.

Tone errors don’t even have to be outrageous, just off-beat. This is absolutely part of CX AI oversight, whether companies like it or not.

Misclassification & journey misrouting

Smart routing can absolutely transform CX. It might even be the secret to reducing handling times. But if your intent model falls apart:

• Complaints get tagged as “general enquiries.”
• Cancellation requests bounce between departments.
• High-risk customers get routed to low-priority queues.
• Agents spend half their time rewriting what the AI misread.

When companies adopt agentic systems inside CRMs or collaboration platforms (Salesforce, Teams, Slack), misclassification gets even harder to catch because the AI is now initiating actions, not just tagging them. Behavioral drift in these areas builds up subtly.

Bias & fairness issues

Bias is the slowest-moving train wreck in CX because nothing looks broken at first.

You only notice it in patterns:

• Certain accents triggering more escalations,
• Particular age groups receiving fewer goodwill gestures,
• Postcode clusters with mysteriously higher friction scores.

A survey last year found 63% of consumers are worried about AI bias influencing service decisions, and honestly, they’re not wrong to be. These systems learn from your historical data, and if your history isn’t spotless, neither is the AI.

Policy, privacy & security violations

This is the failure mode that’s getting more painful for business leaders:

• A bot accidentally quoting internal-only pricing.
• A Teams assistant pulling PII into a shared channel.
• A generative agent surfacing sensitive case notes in a CRM suggestion.

None of these will necessarily trigger a system alert. The AI is technically “working.” But behaviorally, it’s crossing lines that no compliance team would ever sign off on.

Drift & degradation over time

Here’s the thing almost nobody outside of data science talks about: AI drifts the same way that language, processes, or product portfolios drift. Gradually. Quietly.

Models don’t stay sharp without maintenance. Policies evolve. Customer context changes. And then you get:

• Rising recontact rates,
• Slowly dipping FCR scores,
• Sentiment trending down month over month.

Organizations that monitor drift proactively see significantly higher long-term ROI than those who “set and forget.” It’s that simple.

Behavior Monitoring Tips for AI Governance Oversight in CX

AI is making decisions, influencing outcomes, and shaping journeys, yet for some reason, companies still aren’t paying enough attention to what goes on behind the scenes. It takes more than a few policies to make AI governance oversight in CX work. You need:

A Multi-Layer Monitoring Model

With AI, problems rarely start where you’d think. If a bot is rude to a customer, it’s not a chat app that’s usually the problem, it’s something underneath. That’s why you need to monitor all the layers:

• Data layer: Here, you’re watching for data freshness, schema changes, versioning of your knowledge base, inconsistent tags across channels, and omni-data alignment across channels. Poor data quality costs companies billions a year, but unified data reduces service cost and churn.
• Model layer: At this level, useful metrics include things like intent accuracy, precision/recall, hallucination rate, and AI drift detection signals like confidence over time. Think of this as your AI’s cognitive health check.
• Behavior layer: Here, you’re looking at escalation rates, human override frequency, low-confidence responses, weird tool-call chains, anomaly scores on tone, sentiment, and word patterns.
• Business layer: This is where you see how AI activity correlates to results like CSAT/NPS scores, re-contact rate, churn levels, cost-per resolution, and so on.

The Right CX Behavior Metrics

If you forced me to pick the non-negotiables, it’d be these:

• Hallucination rate (and how often humans correct it)
• Empathy and politeness scores
• Sentiment swings inside a single conversation
• FCR delta pre- and post-AI deployment
• Human override and escalation rates
• Percentage of interactions where the AI breaks policy
• Cost-per-resolution

If you only track “containment” or “deflection,” you’re not monitoring AI properly.

A Holistic Approach to Observability

The teams doing this well have one thing in common: end-to-end traces that show the whole story.

A trace that looks like this: Prompt → Context → Retrieved documents → Tool calls → Model output → Actions → Customer response → Feedback signal

If you can’t replay an interaction like a black box recording, you can’t meaningfully audit it, and auditing is core to AI ethics and governance, especially with regulations tightening.

You also need:

• Replayable transcripts
• Decision graphs
• Versioned datasets
• Source attribution
• Logs that a regulator could read without laughing

If your logs only say “API call succeeded,” you’re not looking deep enough.

Alerting Design & Behavior SLOs

Most orgs have SLOs for uptime. Great. Now add SLOs for behavior, that’s where AI governance oversight grows up.

A few examples:

• “Fewer than 1 in 500 interactions require a formal apology due to an AI behavior issue.”
• “0 instances of PII in AI-generated responses.”
• “No more than X% of high-risk flows handled without human validation.”

Alerts should trigger on things like:

• Sharp drops in sentiment
• Spikes in human overrides
• Unusual tool-call behavior (especially in agentic systems)
• Data access that doesn’t match the pattern (teams/slack bots can be wild here)

Instrumentation by Design (CI/CD)

If your monitoring is an afterthought, your AI will behave like an afterthought.

Good teams bake behavior tests into CI/CD:

• Regression suites for prompts and RAG pipelines
• Sanity checks for tone and policy alignment
• Automatic drift tests
• Sandbox simulations (Salesforce’s “everse” idea is a great emerging model)
• And historical replay of real conversations

If you wouldn’t deploy a major code change without tests, why would you deploy an AI model that rewrites emails, updates CRM records, or nudges refund decisions?

AI Governance Oversight: Behavior Guardrails

Monitoring AI behavior is great, controlling it is better.

Behavior guardrails are a part of AI governance oversight that transform AI from a clever experiment into something you can trust in a live customer environment.

Let’s start with some obvious guardrail types:

• Prompt & reasoning guardrails: You’d be amazed how much chaos disappears when the system is told: “If unsure, escalate.” Or “When conflicted sources exist, ask for human review.”
• Policy guardrails Encode the rules that matter most: refunds, hardship cases, financial decisions, vulnerable customers. AI should never improvise here. Ever.
• Response filters: We’re talking toxicity, bias, PII detection, brand-voice checks, the things you hope you’ll never need, but you feel sick the moment you realize you didn’t set them up.
• Action limits Agentic AI is powerful, but it needs clear boundaries. Limits like maximum refund amounts or which CRM fields it can access matter. Microsoft, Salesforce, and Genesys all call this “structured autonomy”, so freedom in a very safe box.
• RAG governance guardrails: If you’re using retrieval-augmented generation, you have to govern the source material. Versioned KBs. Chunking rules. Off-limits documents.
  Use connectors (like Model Context Protocol-style tools) that enforce: “Use only verified, compliant content. Nothing else.”

The Automation / Autonomy Fit Matrix

The other part of the puzzle here (aside from setting up guardrails), is getting the human AI balance right. Before any AI touches anything customer-facing, map your flows into three buckets:

• Low-risk, high-volume: FAQs, order status, password resets, shipping updates. This is where automation should thrive.
• Medium-risk: Straightforward refunds, address changes, simple loyalty adjustments. Great fit for AI + guardrails + a human-on-the-loop to catch outliers.
• High-risk / irreversible: Hardship claims. Complaints with legal implications. Anything involving vulnerable customers. Here, AI is an assistant, not a decision-maker.

To keep these AI governance oversight boundaries solid, implement a kill-switch strategy that includes when to turn off an agent, pause a queue or workflow, or freeze updates to avoid further damage.

The Role of Humans in AI Governance Oversight

There’s still this strange myth floating around that the endgame of AI in CX is “no humans required.” I genuinely don’t know where that came from. Anyone who’s watched a real customer interaction knows exactly how naive that is. AI is remarkable at scale and speed, but when a conversation gets emotional or ambiguous or ethically tricky, it still just acts like software. That’s all it is.

AI governance oversight in CX still needs humans, specifically:

• Humans-in-the-loop (HITL): Any high-risk decision should get a human’s eyes first. Always. HITL isn’t slow. It’s safe. Good AI behavior monitoring will tell you exactly where HITL is mandatory: wherever the AI hesitates, contradicts itself, or hits a confidence threshold you wouldn’t bet your job on.
• Human-on-the-loop (HOTL): Here, the human doesn’t touch everything; they watch the system, the trends, and the anomalies. They’re basically the flight controller. HOTL teams look at anomaly clusters, rising override rates, sentiment dips, and the subtle cues that tell you drift is beginning. They’re the early-warning system that no model can replace.
• Hybrid CX models: We know now that the goal isn’t to replace humans. It’s to let humans handle the moments where trust is earned and let AI tidy up everything that doesn’t require emotional intelligence. Stop striving for an “automate everything” goal.

Another key thing? Training humans to supervise AI. You can build the best monitoring stack in the world, but if your agents and team leads don’t understand what the dashboards mean, it’s pointless.

Humans need training on:

• How to read drift signals
• How to flag bias or tone issues
• How to escalate a behavior problem
• How to give structured feedback
• And how to use collaboration-embedded ai assistants without assuming they’re always right

Embedding AI Governance Oversight into Continuous Improvement

AI behaves like a living system. It evolves, it picks up quirks, it develops strange habits based on whatever data you fed it last week. If you don’t check in regularly, it’ll wander off into the digital woods and start making decisions nobody signed off on.

That’s why continuous improvement isn’t a ceremony; it’s self-defense. Without it, AI governance oversight becomes a rear-view mirror instead of an early-warning system.

Commit to:

• Continuous testing & red-teaming: If you’ve never run a red-team session on your CX AI, you’re genuinely missing out on one of the fastest ways to uncover the weird stuff your model does when nobody’s watching. Red-teamers will shove borderline prompts at the system, try to inject malicious instructions, and stress-test policy boundaries, to show you gaps before they turn into real problems.
• Tying monitoring to predictive CX & customer feedback: If you want to know whether your AI changes are helping or quietly sabotaging the customer journey, connect them to your predictive KPIs. Watch what happens to CSAT, NPS, predicted churn scores, likelihood-to-repurchase, and customer effort.
• Knowledge base integrity review: 80% of hallucinations probably start in the knowledge base, not the model. One policy update slips through without review, or a well-meaning team member rewrites an FAQ with different wording, and suddenly your AI is making decisions based on contradictory inputs. Regular KB governance should become as normal as code review.
• Data quality & lineage checks: The model can only behave as well as the data it’s seeing, and CX data is notoriously chaotic: different teams, different taxonomies, different CRMs duct-taped together over several years. To keep AI honest: consolidate profiles into a CDP with one “golden record,” enforce schemas, and define lineage so you can actually answer, “Where did this value come from?”

The organizations doing this well treat AI like any other adaptable system. They run a full loop: Monitor → Detect → Diagnose → Fix → Test → Redeploy → Report. Simple as that.

AI Governance Oversight: The Only Way to Scale CX AI Responsibly

If there’s one thing that’s become clear while watching CX teams wrestle with AI over the past two years, it’s this: the technology isn’t the hard part. The model quality, the workflows, and the integrations all come with challenges, but they’re solvable.

What really decides whether AI becomes a competitive advantage or a reputational hazard is how well you understand its behavior once it’s loose in the world.

That’s why AI governance oversight, AI behavior monitoring, guardrails, kill switches, and human review models matter more than whatever amazing feature your vendor demoed last month. Those safeguards are what keep the AI aligned with your policies, your ethics, your brand personality, and, frankly, your customers’ tolerance levels.

You can’t prevent every wobble. CX is too complicated, and AI is too adaptive for that illusion. But you can design a system that tells you the moment your AI starts drifting, long before the customer feels the fallout.

CX is just going to keep evolving. Are you ready to reap the rewards without the risks? Read our guide to AI and Automation in Customer Experience.

Will Salesforce’s Informatica Acquisition Make Agentforce Unstoppable?

Salesforce has announced the acquisition of Informatica.

First reported back in May, the purchase has now officially been confirmed for approximately $8BN.

The deal will see Salesforce leverage Informatica’s AI-powered cloud data management capabilities to improve its agentic AI offerings – most notably, the Agentforce platform.

Alongside the data catalog, Salesforce will also gain access to Informatica’s integration, governance, quality and privacy, metadata management, and Master Data Management (MDM) services.

The end goal is to use these capabilities to build a unified data foundation for agentic AI, enabling safe, responsible, and scalable AI agents across the enterprise.

Indeed, in discussing the news, Salesforce Chair and CEO Marc Benioff described data and context as the “true fuel of Agentforce.”

“When companies get their data right, they get their AI right, and Agentforce becomes unstoppable.”

In terms of the specifics, Salesforce also detailed how Informatica’s capabilities will sharpen its Data 360 feature… (Read more).

Cloudflare Outage Disrupts Major Platforms, Payments, and Black Friday Plans

It’s becoming a familiar story: A technical glitch at Cloudflare, one of the biggest internet infrastructure providers, knocked a number of websites and services offline for a few hours on November 18, disrupting customer access and merchant payments.

X (formerly Twitter), ChatGPT, Claude, Perplexity, Spotify and payment giant Square were among those caught up in the fallout.

The trouble began just before 11:48 GMT, when Cloudflare posted that it was dealing with an “internal service degradation” causing intermittent outages across its service network. Users saw error pages, stalled logins, broken APIs, and sites claiming connections were blocked. There were a few conflicting signals about the restoration progress, as at one stage the company reported that services were beginning to recover, but then around 15 minutes later reverted to “continuing to investigate this issue.”

By 13:04 GMT, Cloudflare admitted that one of its fixes involved disabling WARP access in London entirely, temporarily cutting off users from its WARP performance-boosting and VPN service that helps secure and accelerate internet connections:

“During our attempts to remediate, we have disabled WARP access in London. Users in London trying to access the Internet via WARP will see a failure to connect.”

Cloudflare announced a fix five minutes later, but continued to receive “reports of intermittent errors” until close to 17:00 GMT… (Read more).

Five9 Targets CX Inefficiencies with New Genius AI Upgrades

Five9 has introduced a fresh wave of Genius AI updates designed to push the company’s “Agentic CX” vision further into the contact center core.

Announced at the company’s CX Summit in Nashville, the new capabilities span routing, quality management, analytics, and digital engagement, tying them more closely together to help organizations extract greater value from AI at scale.

As many enterprises attempt to take AI from pilot projects into day-to-day operations, fragmentation continues to slow progress.

Disconnected data, inconsistent reporting, and standalone AI experiments often make it difficult to achieve the continuous improvement leaders expect.

Five9’s latest releases aim to combat these challenges by treating AI not as an add-on but as the connective layer running across the environment.

Five9 Chief Product Officer Ajay Awatramani framed the shift as a more fundamental rethinking of how AI should function inside the contact center:

“Our Agentic CX vision is about creating systems that don’t just respond but also help teams better understand and anticipate customer needs.”

So, let’s take a closer look at Five9’s newest features… (Read more).

UJET Acquires Spiral to Address Customer Data Analysis Roadblocks

UJET has announced its acquisition of Spiral to bolster its AI capabilities.

The AI startup will allow UJET to continue its AI roadmap for enhanced customer service solutions.

This partnership will also address customer data analysis issues for UJET’s enterprise customers.

This acquisition is set to further UJET’s AI roadmap vision by bolstering the company’s AI capabilities and addressing customer experience concerns.

By highlighting these issues of visibility between customer and leader, organizations will be able to improve their customer issues before they reach escalation.

In fact, UJET has reported that organizations that are unaware of these individual customer problems are losing approximately $5MN-$30MN in customer churn revenue.

This can be linked to ignored or forgotten negative customer experience complaints, with organizations reportedly gathering only five percent of reported customer issues.

According to UJET CEO, Vasili Triant, customer churn remains a blind spot for many enterprises, arguing that customer interaction analysis is not done effectively… (Read more).