Why AI will fail without a single version of truth

If there's one area where AI was expected to revolutionise business and bring efficiency gains it was customer service. We've seen massive investments in AI chatbots and the expectation remains that Agentic AI agents will rival their human counterparts in years to come, with Gartner predicting that 80 percent of customer service issues will be resolved by autonomous agentic AI by 2029. But right now, AI is tanking customer satisfaction scores.

According to the UK Institute for Customer Service (UKICS), customer satisfaction has hit its lowest point in a decade, with almost a quarter of those polled reporting a bad experience. In fact, the CX Trends 2025 report found 55 percent of customers report feeling frustrated due to being subjected to too many questions by chatbots and voicebots with only 34 percent made a purchase after speaking to a bot, revealing that customers are having to make far more effort when dealing with the technology.

Organisations are now waking up to the fact that AI may not be delivering as expected, with half doing a U-turn on plans to cull their workforce of live agents over the course of the next year. The consensus is that AI must be integrated with and not just supplant human agents because, while AI can respond faster and enable the business to scale that response, it's not able to resolve issues satisfactorily.

There are some fundamental issues as to why this is happening and crucially why it will continue to happen, even with the introduction of agentic AI, unless we change how the technology is utilised. 

Firm foundations

Firstly, the data made available to the AI is far from complete. Customer Experience (CX) platforms typically summarise disconnected sessions rather than interrelate the entire customer journey, so some interactions inevitably get missed. Failed contacts, multiple customer calls, partial resolutions, escalations, failed deflections and all the friction between systems and channels can be omitted even though these valuable insights have a direct correlation to the propensity of the customer to churn. 

Individual channels are also typically siloed or supported by separate platforms, with email, IVR, bots and voice all dealt with separately and that prevents the unification of data, making it impossible to gain a holistic view of the customer journey and experience. And it's a problem that becomes even more evident in large enterprises that have multiple business units or regional operations which can see customers soon become frustrated because there is no process in place to track their journey across the enterprise.

Without a data-first integrated CX stack, the AI cannot both draw upon and contextualise the CX to date. AI Retrieval Augmented Generation (RAG) initiatives need to be able to identify and pull-down live customer journey data while supplementing this with other validated sources such as third-party apps and web journeys from platforms like Adobe, Braze, and MS Dynamics. If the AI is only able to be trained upon and respond to partial data sets, it's inevitable that its output and relevance will be poor.

Even when that data is available, it needs to be suitably sequenced and structured in a well-defined data architecture to avoid the AI having to deal with conflicting information. Often businesses will have more than one version of a product, for example, which can cause confusion, and it's vital that the AI can reference one single validated version of truth to avoid erroneous responses or confabulations. 

A capable knowledge management system can often solve these issues. It ensures that the AI can see all the historic and contextual information about the company's products and services that it needs to support a customer. The Agentic AI can then leverage that information in real-time in conjunction with the full customer journey detail to determine the likely intent and sentiment of the customer, allowing it to respond appropriately and in way that is more likely to lead to a successful resolution.

Earned trust

Another issue is one of trust. If the customer or the human agent receives an incorrect response to a question from the AI, that trust is eroded and the user is less disposed towards using these channels again. There is already evidence of this, with the CX Trends 2025 report revealing that 47 percent of customers struggle to get accurate answers from AI chatbots which can damage brand perception and customer loyalty. It's for this reason that AI initiatives need to be precision focused rather than simply rolled out under the expectation they can be used to address all customer queries.

To achieve this, AI guardrails must be put in place that aren't just shaped by the technical limitations of the technology but are determined by customer need. The AI can then learn over time from every interaction and because its learning from clean, live data that gives a comprehensive view of the customer, it can deliver consistent trustworthy responses. It then becomes possible to explore the potential use cases for AI rather than expecting it simply to replicate a human agent.

There is also no reason why the AI can't also adopt a persona that more accurately reflects the image of the brand, for instance, by adopting a different tone of voice. It can be designed to capture information from the customer at the get-go, so that it determines intent more readily, enabling faster and more accurate resolution. And, perhaps most importantly of all, it's customer handling can be analysed to determine how accurately the bot has gauged intent and responded. Should this start to decline, a human can be brought back into the loop to re-educate the bot.

It's this ability to equip and guide AI that will determine if these initiatives succeed or fail. These bots can be used to increase the speed of response and deal with the most common issues to reduce cost to serve but they should also be able to recognise when they need to handover to a human. In this way, the business can harness the best of both.