In May 2025, Klarna's CEO publicly admitted what many in the industry had suspected for months. After replacing 700 human agents with AI, the company reversed course and began rehiring human agents. Customer satisfaction scores had dropped. Edge cases had overwhelmed the system. The AI that had looked polished in early demonstrations turned out to struggle with the full unpredictability of real customer conversations.
What happened at Klarna is not a story about AI failing. It is a story about implementation failing.
It is also not an isolated case. Analysis of contact centre AI deployments consistently finds that 95% of AI pilots in call centres are quietly failing. Not loudly, with press coverage and public post-mortems. Quietly. Pilots stall. Projects never reach production. The return on investment that was signed off in the budget meeting never materialises.
Gartner put a number on the broader trend earlier this year: over 40% of agentic AI projects will be cancelled by the end of 2027. The reasons are consistent with what we see across AI contact centre implementation in the field every week: escalating costs, unclear business value, and inadequate risk controls.
So why does this keep happening? And what does it actually look like when AI contact centre implementation works?
The pressure trap nobody is talking about
Before we get to the technical failures, there is a strategic one that sits upstream of all of them.
Gartner's research found that 91% of customer service leaders feel pressured into implementing AI. Not inspired by a compelling business case. Pressured. By their board. By their competitors. By the fear of being the person who did not move fast enough on AI contact centre implementation.
That pressure produces a recognisable set of bad decisions. It produces AI implementations where nobody agrees on what success looks like before the project starts. It produces budgets approved on projected ROI that is never tracked after go-live. It produces contact centres where AI is deployed before the customer journeys have been redesigned around it, and before the agents using it have been trained on how it changes their work.
The technology is not the problem. The rush is.
When the goal is to be seen doing AI rather than to solve a specific operational problem with AI, the project is already in trouble before a line of configuration is written. Every decision that follows gets made in service of the optics rather than the outcome. And when the outcome eventually arrives, it is rarely what anyone expected.
The question worth asking before any AI procurement starts is not 'which platform should we buy?' It is: 'what specific problem are we trying to solve, and is AI actually the right tool for it?' Getting that question answered honestly, with a defined metric and a realistic timeline, is the single most important step in a successful implementation.
The four places where implementations fall apart
Across the implementations we see in the market, failure tends to cluster around four specific points.
The first is vendor over-promise. Too many AI contact centre vendors sell on demonstrations rather than deployments. A demonstration shows a clean, controlled interaction with a cooperative caller and a well-defined problem. A real contact centre handles accents, frustrations, background noise, complex account histories, regulatory edge cases, and emotionally charged conversations that no AI model was trained to navigate cleanly. The gap between what the demonstration showed and what the deployment delivers is where most AI contact centre projects fail.
The second is integration failure. An AI layer that sits outside your existing systems is a liability, not an asset. If agent assist prompts cannot pull from your CRM in real time, if automated post-call summaries cannot be written back to your ticketing system, if the AI cannot access customer history at the point of need, you have created process complexity rather than removed it. Integration is not a post-go-live problem to be solved after deployment. It is a pre-purchase conversation that should happen before anything is signed.
The third is the escalation gap. This is the failure mode that cost Klarna most dearly. When the AI reaches the edge of its capability, what happens next? If there is no designed handoff to a human agent, customers get stuck. They repeat themselves. Their frustration compounds. The AI has failed and the human who could resolve the issue is nowhere in the contact flow. Building smart escalation paths, where the AI recognises its limits and routes to the right person with full context intact, is not optional. It is the thing that determines whether AI-assisted contact feels seamless or maddening.
The fourth is expectation misalignment at every level of the organisation. The board wants cost reduction within six months. The operations team expects agent productivity to double. The IT team is concerned about security and compliance. The agents are worried about their roles. When these expectations are never surfaced and aligned before the project begins, each stakeholder measures a different outcome and finds it wanting. The project that was greenlit to solve everyone's problem ends up solving nobody's.
The data problem sitting underneath everything
There is one more failure pattern worth naming, because it is the most common and the least discussed: poor data quality.
Analysis of failed enterprise AI projects consistently finds that 85% cite inadequate data as a root cause. The AI model is not the problem. The data feeding it is. Contact centres that attempt to deploy AI on top of disconnected CRM systems, inconsistent interaction records, and poorly structured knowledge bases are asking the AI to work with materials that would confound a human agent.
This is not an AI problem. It is a data architecture problem, and the right response is not to buy a more sophisticated AI model. It is to invest in connecting, cleaning, and structuring the data that the AI needs before you ask it to do anything useful with it.
The organisations that succeed at AI contact centre implementation treat data readiness as a prerequisite, not a parallel workstream. They audit what data the AI will need access to. They identify the gaps between what currently exists and what will be required. They fix those gaps before go-live. This is rarely glamorous work. It does not appear in product demonstrations or vendor case studies. But it is the work that determines whether the implementation performs in production.
What the contact centres getting this right have in common
The contact centres where AI contact centre implementation is working share a few consistent patterns.
They start with one clear, measurable problem. Not 'AI in the contact centre' as a strategic initiative, but a specific outcome: reduce post-call handling time on billing enquiries by 15%. Improve first-call resolution on a defined set of technical queries. Reduce average handling time in one queue by two minutes. One problem. One metric. One pilot. Prove it works. Then expand.
They build human escalation into the design from day one, not as an afterthought. The design question is not 'how do we automate this interaction?' The question is: at what point does this interaction need a human agent, and what information does that human need when they take over? The escalation architecture shapes everything else.
They invest in integration before AI capability. Getting the AI connected to the right data sources takes longer than vendors suggest and matters more than the AI feature set itself. A modestly capable AI with clean, connected data will outperform a sophisticated AI operating in isolation every time.
And they choose partners who are honest about what their platform can and cannot do today, who have enough contact centre AI implementations behind them to know where the problems arise, and who have the services depth to deliver the implementation themselves rather than hand it off.
A different approach to AI in the contact centre
NeonNow was built around the conviction that AI in a contact centre is most valuable when it makes human agents more effective, not when it tries to replace them. Our AI contact centre platform reflects that directly.
NeonNow IQ, the AI intelligence layer built into the NeonNow platform, surfaces agent assist prompts during live calls before agents have to search for the answer. It detects sentiment shifts in real time so supervisors can intervene before a conversation escalates. It produces automated post-call summaries that eliminate the after-call admin that consumes significant time in most contact centre operations. It routes customers based on detected intent so they arrive at the right person with context, not in a generic queue with none.
This is AI designed to amplify the human on the other end of the call. It runs on Amazon Connect and AWS infrastructure, which matters for organisations that need their contact centre platform to meet enterprise standards for security, compliance, and availability. It is supported by a team with consulting and delivery experience to make the implementation perform the way the business case said it would.
The AWS Contact Center blog regularly covers what enterprise AI contact centre deployments look like when they are well-architected. The pattern is consistent with what we see in the field: structured pilots, clear metrics, integration-first design, and human escalation paths built before anything goes live.
If you want to see what AI contact centre implementation looks like when it works, not in a demonstration but in a production contact centre, the conversation starts below.
Ready to see AI that works in practice, not just in theory?
If your organisation is evaluating AI contact centre implementation, or if a current project is not delivering what was promised, we would like to help. Book a demo to see NeonNow IQ working in a live contact centre environment, or talk to our team about where your AI contact centre implementation should begin.
.jpg)
.svg)
