What is a Successful AI Implementation in the Contact Centre?

When asked about what 'successful' AI in the contact centre looks like, Head of AI & Automation at FourNet, James Brooks suggests that you need to reframe the question entirely. Success isn’t about deploying AI, it’s about knowing “what you want the AI to do. What is the objective? What metric or metrics are we trying to change?” 

Without a defined outcome, you’re essentially implementing technology for technology’s sake. That means getting clear on whether you’re trying to reduce average handling time, improve first contact resolution, support vulnerable customers more effectively, or achieve something else entirely. Only then can you understand “what levers or KPIs need to be pulled in order to get to that metric to change.” 

For organisations in highly regulated environments, success also means taking a lower-risk approach. Director of Operations at MoneyPlus, Christina Albert explains that “selecting the right use case is about taking a lower risk approach and starting to get ourselves comfortable with it.” The goal is finding something that feels “safe” and “beneficial” whilst building “confidence and awareness of how it’s going to work.” 

To know whether AI has been 'successful' or not, you need to start by defining clear objectives tied to specific KPIs. Choose use cases appropriate to your risk appetite. Build in staged success criteria that account for adoption curves and trust-building periods. And ensure you can actually measure whether agents are using the AI, because as Andrew Tucker, Solutions Architect at NiCE notes, “if they’re not using it, I can’t see that it’s being used and I can’t measure that it’s being used. Then the project failed right at the get-go.” 

Organisational readiness extends far beyond having budget or technical infrastructure. Andrew explains that “AI does introduce new data points and data sources for you to look at, which may not necessarily sit under a very traditional metric driven success criteria.” 

With auto-summary, for example, you’ll want to track how many agents are reviewing AI-generated summaries and whether they’re making changes. That introduces “a whole new set of metrics to look at.” The critical question becomes whether “the teams that are measuring success are ready for the new success measurements?” 

Here’s where alignment gets tricky. “Operations might be. They’ll say, ‘we’re ready to go’. But then you’ve got your risk and compliance department that’s saying, way. I’m not there yet.” This tension is particularly acute in regulated businesses where the strategic alignment of internal processes determines whether AI can be safely deployed. 

Clear warning signs that suggest you’re not ready for AI are if “Data is not in good shape or we don’t have good policies and procedures already.” Beyond that, “clear executive sponsorship” is critical, focused on whether “this is moving us towards our strategic ambitions” and whether it “supports our ability to deliver good outcomes for customers.” 

Before proceeding with AI, conduct an honest assessment across operations, IT, risk, compliance, and leadership. Are they aligned on objectives? Do you have the measurement infrastructure for new AI-specific metrics? Is your data clean and accessible? Have you identified or developed the skills needed to support AI systems? If the answer to any of these is no, address those gaps before deployment. 

Selecting the right partners is critical, and as MoneyPlus discovered, “it wasn’t an easy one because we did look at lots of different technologies in particular, but a few different implementation partners as well.” 

Technical capability matters, but it’s table stakes. What really separates adequate vendors from genuine partners is whether they understand your industry and regulatory context. MoneyPlus needed “a proven track record that the partners that we were going to work with understood our industry and our regulatory context.” 

Beyond credentials, “a willingness to work collaboratively, especially with the implementation partner” was essential. The goal was finding “a group of people that were going to feel like an extension of our team.” AI implementations aren’t transactional, they’re ongoing relationships requiring deep collaboration. 

James adds important context about the discovery process. Before entering into an AI project, understand “what are you trying to achieve?” Sometimes the question is “does that need to be achieved through AI or through technology full stop?” The answer might be simpler than AI. It might be a planning exercise where you can improve efficiency with existing resources. 

Start with a clear list of evaluation criteria weighted by importance. Look beyond product demonstrations to customer references and real-world implementations. Assess cultural fit and collaborative approach, not just technical capability. And ensure potential partners understand your regulatory landscape and can demonstrate relevant experience in similar contexts. 

Getting agent buy-in is fundamental to success. Andrew is clear that “the key to successful deployment is transparency. Make sure the key stakeholders within any particular project are engaged. They know what’s going on. They know why it’s going on.” You need feedback loops to measure participation and ensure the AI is having the desired impact. 

The fear of job loss is real when people hear ‘AI.’ Address it directly and honestly. Explain that AI isn’t about replacement, it’s about augmentation. Show agents how AI will make their jobs easier, whether that’s auto-generating summaries, providing real-time prompts, or flagging compliance risks before they become issues. 

Involving agents early makes a significant difference. Bring them into pilot groups, ask for their feedback, and let them help shape how tools are configured and deployed. When agents feel they have a say in the process, they’re far more likely to embrace the technology rather than resist it. 

Planning requires thinking in stages, not milestones. Andrew emphasises that “deployment success stages of AI, you need to build in considered success stages. It’s not just flip the switch and measure.” 

Phased rollouts consistently outperform big-bang deployments. Start with a pilot, choose a small group, test, learn, refine, then scale. This gives you the chance to iron out issues in a controlled environment before they affect the wider organisation. It also helps with change management, giving agents time to adapt and leadership time to see results before committing to full-scale deployment.” 

Integration complexity almost always exceeds expectations. AI tools need to connect with CRM systems, telephony platforms, workforce management software, and knowledge bases. Each integration introduces complexity with APIs to build, data formats to standardise, and security protocols to follow. 

Governance matters from day one. Before deploying AI, establish clear structures determining who makes decisions, who approves changes, and who’s accountable for outcomes. Without governance, projects drift and stakeholders end up pulling in different directions. 

Don’t treat go-live as the finish line. Post-implementation support is where you really start to see value through monitoring usage, gathering feedback, refining AI models, and continuously optimising performance. 

Managing expectations is one of the most overlooked yet critical parts of any AI implementation. AI is not a silver bullet for systemic operational issues. If your contact centre has poor processes or fragmented systems, AI won’t solve those problems. Fix your foundations first. 

When it comes to performance on the job, James explains that AI isn't going to be perfect, but if you take your best performing agent – AI will operate at 80% of their performance, but 100% of the time. 

Andrew then articulates a key concern about LLMs operating as black boxes. “If you ask an LLM to give you sentiment, it’ll give you sentiment. Ask it how it’s calculated it. It’ll give you a different view as to how it got to that conclusion today as it did tomorrow.” This inconsistency is problematic in regulated environments where “you need consistency, consumable output which is clearly defined.” 

That’s why human oversight remains critical. “You’ve got to keep a human in the loop” and “keep them skilled.” AI should support decisions, not make them autonomously. In cases where consistency is essential, you might need rule-based systems alongside AI to ensure predictable outcomes. 

Iteration is inherent to AI success. Implementations are never one-and-done. Models need retraining, prompts need adjusting, and feedback loops need acting upon. Set the expectation early that AI is a journey. The first version won’t be perfect. Neither will the second. But with each iteration, it improves. 

For organisations in regulated environments, risk management is fundamental. The MoneyPlus perspective is clear: “Regulation around AI is rapidly evolving, but perhaps not as fast as the technology is. Being mindful of that and continually adapting our approach is essential". 

The AI landscape changes faster than regulators can respond, creating ambiguity about how existing frameworks apply. The tension between operations and compliance is common. “Operations might be say ‘we’re ready to go’. But then you’ve got your risk and compliance department that’s saying, no way. I’m not there yet.” 

Building compliance in from the start is far easier than retrofitting it later. That means considering data privacy, security, audit trails, and explainability at the design stage. For contact centres, think about how customer data is being used, where it’s stored, and who has access to it. 

Ensuring AI decisions can be explained and defended matters for both compliance and trust. Having processes to identify and mitigate bias is essential, particularly when AI is involved in vulnerability identification or service delivery decisions affecting customer outcomes. 

Measuring AI success requires new metrics alongside traditional ones. Andrew explains that “AI does introduce new data points and data sources for you to look at.” With auto-summary, “what I want to be able to do is review how many of the agents are reviewing it and if they’re making changes. That introduces a whole new set of metrics.” 

Usage metrics are critical because “if they’re [agents] not using it, I can’t measure it’s being used. Then the project failed right at the get-go.” If adoption is low, understand why and address it quickly through additional training, usability improvements, or system refinements. 

Track agent amendments when AI generates content. Lots of changes suggest the AI isn’t hitting the mark. Decreasing amendments over time suggest it’s learning and improving. If agents stop reviewing altogether, that could be a red flag suggesting they’re blindly trusting AI they shouldn’t. 

Quality metrics differ from efficiency metrics. Efficiency metrics like average handling time tell you how fast things happen. Quality metrics tell you how well they’re happening. Both matter, but they serve different purposes in understanding AI’s true impact. 

Define baseline metrics before AI deployment so you can measure change. Establish both leading indicators like usage rates and amendment patterns alongside lagging indicators like efficiency gains and customer satisfaction. Review metrics regularly and use insights to refine AI systems. And set realistic ROI timelines that account for adoption curves, with early wins typically visible within weeks or months but larger gains taking longer to materialise. 

Despite the hype, most organisations are at the beginning of their AI journey. James is blunt “We’re nowhere near. We’re still at the beginning of understanding the AI journey. The LLMs have only been around for a couple of years, really, and we’re only really scratching the surface of capability.” 

New use cases emerge through experimentation and happy accidents. “The AI might do something in a new way that you’ve never even considered, that could be a use case.” As organisations experiment, they discover capabilities they didn’t know existed. 

The MoneyPlus perspective adds necessary caution. “Even like on a personal level, there’s gotta be a little bit of caution exercised. I’ve had some questionable results from AI.” This reinforces the importance of validation and not blindly trusting AI outputs. 

Don’t feel pressure to be at the cutting edge immediately. Focus on solid foundations and incremental progress. Start with well-understood use cases that deliver clear value. Learn from each deployment before expanding. Share learnings across your organisation to build collective expertise. And stay informed about AI developments in your sector whilst maintaining realistic expectations about maturity timelines. 

Customer readiness is one of the most overlooked aspects of AI implementation. “I think one really interesting point that’s come out is, is the customer ready for it? That’s a big thing that isn’t really considered a lot of the time.” 

Context changes everything when it comes to customer comfort with AI. Andrew explores meeting customers “where they are, where they’re feeling more vulnerable at night.” An insurer considering agentic bots discovered that “most of our customers who potentially get stuck on this journey, do it when we’re closed.” Rather than limiting support to contact centre hours, AI enables help exactly when customers need it. 

The Japanese care bot example is instructive. When care bots were deployed with “a nice voice” to support elderly people, “the response was phenomenal.” The lesson is clear: “We have to be careful not to prejudge customers’ receptiveness to a solution if it’s actually providing a really useful outcome.” 

Christina mentioned that the key is that “we have to include them [customers] in the design and we have to have them in the feedback loop constantly.” Don’t make assumptions about what customers will or won’t accept. Test AI implementations with real customers across different scenarios and demographics. 

Design AI interactions with clear escalation pathways to human agents. Be transparent about when customers are interacting with AI. Gather customer feedback systematically during pilots and after deployment. Consider context when deploying AI, recognising that acceptance varies by situation, time of day, and customer demographics. And always provide choice, allowing customers to opt for human support when they prefer it.