Why most AI customer service pilots never scale

Containment rates of 70–85% for well-defined use cases, cost-to-serve reductions of 20–40%. These aren’t projections. They’re examples of the results successful organisations are achieving right now with AI-enabled customer service. Individual organisations are already using AI to manage tens of millions of interactions annually. The technology’s potential to deliver enterprise value is no longer in question. So why are so many organisations still stuck at pilot stage, watching while others move into production and capture value first? 

What many get wrong - and how to do it right 

The most common mistake we see is organisations treating the shift from pilot to production as a technology procurement exercise, rather than a business transformation programme. This is borne out by our recent research, From Chaos to Confidence, which includes interviews with dozens of organisations using AI to transform customer service. 

With many customer and operations leaders under pressure to implement AI, the temptation as you plan to scale is to focus on the technology. But that instinct is what causes programmes to stall for months, even years. In fact, a range of other factors determines whether AI delivers the results you want – not least whether your organisation has the operational, organisational and governance capabilities in place to run AI agents in production. 

This represents a significant shift from 18 months ago, when organisations were rightly asking ‘can this work?’, to today, where mature organisations are asking ‘can we operate, maintain and improve this?’. 

Here are six questions to ask before you start to scale AI in service:

1. Do you have a clear strategy? 

Organisations that achieve the strongest results treat deploying an AI agent as a business transformation that happens to involve technology, rather than a technology project looking for a business case. That might sound like a subtle difference, but it has a big impact on how your programme is structured, who owns it and how decisions get made when things don’t go to plan. 

To be ready, you need alignment across the organisation on the outcomes you are trying to drive with AI and agreement on how you’ll measure success against these outcomes. You also need a prioritised set of use cases and sponsorship by someone senior enough to make resourcing decisions and keep momentum behind them. It’s common to see AI activity springing up across different business units at the same time, each with its own vendor conversations and pilot projects. But that’s also where many programmes get stuck and never make it to production scale. 

In practice: a UK energy company started with strategy when its Group CIO set a 2030 technology vision first, and only then did a rapid assessment surface ~£10m of unrealised voice AI value against a prioritised set of use cases. Treating AI as a business transformation aligned to a clear ambition (rather than a standalone pilot) is what gave the programme the sponsorship and momentum to move toward production.

2. Have you defined the customer experience you're designing for? 

It’s easy to get lost in the technology without thinking about what you want the technology to enable – a better experience for your customers. A key question is ‘what does each segment value?’. A high-net-worth customer disputing a transaction wants reassurance, which may mean faster access to a human even when AI could resolve the issue technically. A customer resetting a password just wants speed, and any human involvement is friction. Designing for this means mapping journeys by intent and emotional stakes, not just query type, and deciding upfront where AI should own the interaction, where it should support a human and where it shouldn't feature at all. 

Also critical is being explicit about how customers move between the AI agent and a human, what the transition feels like and which segments need different treatment. Seamless handoff is harder than it appears. A customer who explains their situation to an AI agent and then repeats it to a human agent has a worse experience than one who reaches a human directly. Handoff design is a product decision that needs to be made before the platform is configured. 

In practice: A global airline now automating over 15 million interactions a year began with a forensic analysis of its contact drivers, mapping journeys by frequency, complexity and emotional stakes before deciding where AI should own, support or stay out of the interaction. Designing around intent first is precisely what enabled it to automate high-volume, well-defined tasks while routing sensitive cases to humans.

3. Do you have the capability to build, operate and improve an AI agent? 

Full value from AI is rarely achieved immediately from go-live – agents need to be refined and optimised over time. Organisations getting the best results from AI in service treat go-live as the start of a continuous improvement cycle, rather than the finish line.  

Your business structure and capabilities determine whether you can build, operate and continuously improve an AI agent after go-live. This is about ensuring not only that the right roles exist, but that those roles are set up to work effectively together. Conversation designers disconnected from contact centre operations, engineers stretched across several AI initiatives or product owners with the mandate but not the budget – these misalignments are as damaging as capability gaps. 

Skills are another key factor. If you’re deploying through a specialist platform or outcome-based provider, you don’t need a large, in-house AI engineering team. But you do need a level of AI literacy among the people who own and run the agent. That means people who can assess conversation quality, spot recurring failure patterns, brief improvements accurately and hold vendor partners to account. The good news is that these skills can be learned. Organisations that invest in them early typically recover the cost through faster iteration and fewer issues after launch. 

A third area where many technically sound deployments struggle to gain traction is culture and change. Redefining colleagues’ roles, being transparent about what AI will and won’t change, and helping teams move from handling interactions to supervising AI systems all require sustained investment. Organisations succeeding with AI treat front-line teams as co-creators of the AI, not just end users. When agents understand why the AI makes the decisions it does, contribute to improving it and see it as something that supports rather than replaces their role, the results are consistently better.  

In practice: A global sports and entertainment organisation running AI agents across 14 major events found the agents effective, but only because they have allocated dedicated capacity to updating content and refining responses. It's a reminder that sustaining quality depends on people with the AI literacy to keep improving the agent after the project team moves on. 

4. Is your data and integration estate ready? 

Clean, accessible data and codified institutional, process and policy knowledge are among the strongest predictors of whether an AI deployment succeeds in the first 12 months. 

On data, the standard you need to reach before starting is a unified customer view accessible through stable, documented APIs, with clean identity and authentication architecture that works for machine-to-machine interactions. Our research shows integration typically takes three to five times longer than building the agent itself. A unified Customer Action Gateway – a stable integration layer that abstracts back-end into a single clean interface – speeds this up considerably. 

Your knowledge base needs attention too. AI cannot act on knowledge that exists only in people’s heads. If escalation relies on an experienced agent’s judgement, that judgement needs to be made explicit and documented before any AI agent can be trained on it. Content written for human agents often includes informal language and implicit escalation cues that don’t translate for AI. Every knowledge article should clearly define the situation, the action required and the expected outcome, without relying on assumed context.  

In practice: One global e-commerce organisation reported that integration was straightforward because its order data was already well-groomed, centralised and accessible through clean APIs. Its Head of Product noted that without that foundation, implementation would have been significantly harder, illustrating why data readiness and API maturity are greater determinants of first-year success than technology 

5. Do you have the governance and risk clarity? 

It’s essential to resolve issues around internal governance before you start deployment, not work them out along the way. Clear ownership, clear accountability and a cross-functional forum with the authority to make and enforce decisions about what the agent can do, how it is monitored and when escalation is required are the fundamentals for success. Governance gaps expand as deployments scale. So, treat governance as an ongoing operational discipline (not a one-off procurement check) to ensure you spot any accumulated risks before they become serious problems. 

Regulatory compliance requires the same serious attention as internal governance. The obligations that matter most for UK organisations are UK GDPR, FCA obligations and the EU AI Act, which applies to UK organisations serving EU customers. Customer-facing AI agents in sectors such as financial services, insurance and healthcare are likely to fall into high-risk categories, bringing additional requirements around conformity assessment and human oversight before deployment. 

In practice: A global wealth-management organisation saw its proof-of-concept stall for months because data-residency questions surfaced after vendor selection rather than before. Resolving compliance and data-residency clarity upfront is essential. 

6. Have you chosen partners whose model fits your capability? 

The engagement model you agree with your vendor often matters as much as the technology. Even when you choose a vendor who takes full delivery responsibility, you need to be able to define what good looks like, provide clean data and integration and make decisions quickly. If you choose a vendor who hands over a platform and steps back, you need strong internal engineering and conversation design capability at a minimum.  

Post-go-live ownership matters more than the vendor selection decision itself since AI improves with ongoing tuning as customer language shifts, new products launch and edge cases accumulate, and gets worse without it. The question isn't just who builds the platform, but who owns its performance in month six, when the initial project team has moved on and the system needs active attention to keep pace with the business. 

In practice: A European financial services organisation deliberately chose a highly configurable platform because it matched its vision to own and extend the agent over time. This enabled them to add AI capabilities incrementally and retain the freedom to swap in different LLM, speech and text-to-speech components as the market moved.  

Why the pilots that scale start with honesty, not vendors 

None of these six questions are about technology. They're about organisational readiness — and that's precisely why so many pilots stall. It's far easier to shortlist vendors than to confront whether your governance model has teeth, your knowledge base makes sense to a machine or your teams have been brought along rather than left to find out from a press release. The organisations getting real value from AI in service are the ones willing to do the harder work of getting their own house in order first. In the end, success is determined less by the pilot itself and more by the capability you build around it.