AI doesn't define its environment. It inherits it.

The real problem with AI in CX is not the AI

There's a pattern playing out across most enterprise AI in CX programmes right now. The technology arrives, the dashboards look promising, the early demos impress the executive team, and somewhere between month four and month nine the same complaints start showing up: customers repeating themselves, agents dealing with rework, the system sounding confident in places it shouldn't, and the savings that were promised somehow not arriving in the way they were modelled.

Every time this happens, the instinct is to blame the AI. Wrong model. Wrong vendor. Wrong prompt. Wrong configuration.

The instinct is almost always wrong.

AI doesn't get to define the conditions under which it operates. It inherits them.

Contact centres are not neutral environments. They are pressure systems built up over decades, shaped by operating models, KPIs, technology layers, and incentive structures that long predate the arrival of generative AI. When you deploy AI into that environment, it doesn't reset the conditions. It inherits every one of them. Organisational structure. Operating model. Incentives. Technology constraints. Human workflows. Cultural defaults. All of it.

If the environment is fragmented, misaligned, or built around the wrong metrics, AI will simply automate those problems at scale. The flaws don't disappear. They just get faster, more consistent, and more confident-sounding.

What "the environment" actually means

When people say AI is failing in customer experience, they usually mean one of two things: the bot doesn't work, or the project hasn't delivered the business case. Both are symptoms. The underlying cause is almost always something deeper in the operating environment.

Most AI in CX inherits at least five things from the system around it.

1. Fragmented customer journeys

Customers move across teams, channels, and systems that were never designed to work together. Identity gets re-asked. Context gets dropped. Each component is optimised in isolation while the overall experience feels exhausting. AI dropped into this becomes another station on the assembly line rather than a connective layer across it.

2. Misaligned metrics

What gets measured drives behaviour. When the dashboard rewards containment, deflection, and cost per contact, AI will be optimised to end conversations rather than resolve them. Containment goes up. Repeat contact goes up underneath it. The dashboard looks healthy while the experience quietly erodes.

3. Poor service design

Bad processes don't become good processes when wrapped in a model. They become faster bad processes. The places where customers were already frustrated are now frustrated with greater fluency.

4. Weak escalation paths

When AI cannot resolve, the handoff matters more than the model. Most environments are not ready for it. Context gets dropped at the moment customers most need it preserved. The agent inherits a customer who has already explained their problem twice and is now being asked to explain it a third time.

5. Vendor-led architecture

Roadmaps shaped by what vendors sell, not by what the operating model needs. Capabilities arrive faster than the organisation can absorb them. New tools layer on top of old tools without anyone owning the seams between them.

None of these are AI problems. They are environment problems. AI just makes them louder.

The Automation Trap

Here's where it gets interesting. The way contact centres have been measured for decades didn't just shape how contact centres operate. It shaped how the industry introduced AI into customer experience.

Contact centres have historically been measured as cost centres. Average handle time. Cost per contact. FTE reduction. Containment. These metrics weren't chosen because they were good measures of customer experience. They were chosen because contact centres were treated as a cost to manage, and these were the numbers easiest to put on a board slide.

So when AI arrived, the existing scoreboard was the only scoreboard available.

We didn't position AI as a way to deepen relationships, resolve underlying issues, or surface latent demand. We positioned it as a cost-reduction tool. Not because that's where the value sat. Because that's how the contact centre had always been measured.

AI gets optimised for deflection, not because deflection is the right outcome, but because deflection is what the dashboard was already counting.

This is the Automation Trap. Combine industrialised CX, fragmented systems, and interaction-centric metrics, and the predictable result is AI deployed as a containment engine. Customers don't experience that as a service improvement. They experience it as harder access. Slower resolution. Less accountability. The brand that prided itself on customer relationships starts to feel like a company trying to avoid speaking to its own customers.

AI doesn't just automate. It amplifies.

This is the part most executive teams underestimate. AI is not a neutral overlay on existing operations. It is an amplifier. Whatever the system already does well, AI makes it do faster and more consistently. Whatever the system already does badly, AI also makes faster and more consistent.

Three patterns repeat in almost every enterprise programme I've worked across at Cisco and through advisory engagements:

The question is not how many humans can AI replace? The question is how well does the system perform as a whole?

Why this matters now more than it used to

Failure used to be obvious. A bot that didn't understand. A workflow that broke in front of the customer. Everyone could see the problem. The fix was visible too.

That's changed. Modern AI in CX sounds correct, looks correct, and behaves with confidence even when it's wrong. It generates fluent answers that contain errors. It hands off to agents with summaries that subtly misrepresent the conversation. It closes interactions confidently while the customer's underlying issue remains unresolved.

Fluent failure is harder to detect than broken failure. By the time it shows up in your lagging indicators, churn, complaint volume, regulator interest, the damage is already compounded.

This is why the cost of getting AI wrong is going up, not down. The technology is more capable, which means the failures look more credible, which means they go undetected for longer, which means the damage runs deeper before anyone notices.

Readiness is not a project phase. It is the project.

Most AI in CX initiatives treat readiness as a step. Something you do at the start, before the real work begins. Data readiness. Integration readiness. Stakeholder readiness. Tick the boxes, then build the bot.

That framing is wrong. Readiness is not the precondition for the work. It is the work.

An AI system that's still maturing inside an organisation that has not done the underlying work on journeys, metrics, governance, escalation, and ownership is being asked to compensate for things it cannot see. The model has no view of why the customer's data is split across three systems. It has no view of why the KPI rewards interaction closure over resolution. It has no view of why the escalation path delivers an agent a context-less customer who has already explained themselves twice.

None of that is fixable inside the model. All of it shapes whether the model succeeds or fails.

The readiness work is not the thing you do before the AI project. It is the AI project.

What actually changes the trajectory

None of this is an argument against deploying AI in customer experience. The capability is real. The opportunity is real. The risk of doing nothing is real. The argument is about where the leverage actually sits.

Four moves consistently change the trajectory of an enterprise AI in CX programme.

Fix the environment before scaling the model

Persistent customer context across channels. Identity that travels with the customer rather than being re-asked at every step. A coherent platform that gives the AI a single signal to work from. These are not exciting roadmap items. They are the foundation that determines whether anything downstream actually works.

Change what the dashboard counts

If your primary AI metric is containment, deflection, or cost per contact, you are measuring the thing that is creating your problem. Introduce a journey-level metric alongside, repeat contact rate, end-to-end customer effort, resolution at the issue level rather than the interaction level. Put it on the same dashboard, with the same cadence, with the same audience. Within two quarters you'll see whether the containment number is real or whether it's redistributed effort in disguise.

Treat the escalation path as part of the AI design, not a fallback

The moment AI hands off to a human is the moment trust is either reinforced or broken. Context must travel with the customer. The agent must inherit understanding, not just a transcript. Authentication should not break at the seam. Reversibility should be designed in. Every escalation is a UX event, not a system event.

Name an owner for the live system

Most AI systems drift because nobody owns them after launch. Appoint a person, or a team, with authority and budget to observe the system continuously and modify it based on what it actually does, not what the design said it would do. Launch is not the end of design. It's the point at which design stops being theoretical.

What this means for executives making AI decisions right now

If you are commissioning, sponsoring, or being asked to fund an AI in CX programme in the next six months, the single most useful question to ask is not "what model are we using?" or "what's the business case?" Those questions matter, but they assume the model is the variable.

The more useful question is this: what environment is this AI inheriting, and is that environment actually ready for what we're about to deploy?

If the answer is no, the project is not a model project. It's a readiness project that happens to involve a model. Naming it accurately changes how you scope it, fund it, sequence it, and measure it.

That reframing is uncomfortable, because it makes the work bigger and slower than the vendor pitch suggested. It is also the only framing that consistently produces AI in CX programmes that don't quietly disappoint twelve months in.

AI is not failing customer experience. It is doing what it was always going to do. It is inheriting the environment it was deployed into, and showing the organisation, at scale, exactly what that environment was built to produce.

The question is whether the organisation likes what it sees, and whether it's willing to do the work that genuinely changes the trajectory.