Are CX Teams measuring the right things?

The reason our metrics need rethinking is because the although the definition of customer is no different, the there is now a new agent representing the customer that has an influence on the customer's experience: agents run by Artificial Intelligence (AI). AI representing a customer poses a different set of challenges for businesses - AI agents are now engaging businesses from multiple fronts and Gartner projects that 90% of B2B purchases will be handled by AI agents within three years, channelling more than US$15 trillion through automated exchanges. On the organisational front, Gartner predicts agentic AI will autonomously resolve 80% of common customer service issues by 2029, and that service teams will need to support both human customers and a growing population of machine customers.

What this means is that the Customer Journey map that you and your team have developed now has a different flavour to it: the customer is now a mix of human as well as AI. Since an agent intermediates the interaction, the human only experiences the outcome, not the journey. Which means that since the human doesn't experience the journey, perception metrics collected at the interaction layer lose their signal. And when perception metrics lose signal, dashboards built on them start describing a customer who was never actually there.

In other words, "the customer" is split into two audiences: the human who owns the outcome, and the agent that experiences your infrastructure. In today's world, most organisations (even we still advocate this) measure the human experience and not the agent.

What still holds in this AI world

There are still some metrics that still hold true with the advent of AI. Customer Effort Score (CES) is arguably the metric best suited to the agentic era. Ada's 2026 survey of 2,000 consumers found that accuracy and problem-solving ability ranked above empathy, and that a faster resolution beat a warmer but slower one by a wide margin. Effort removal is the entire value proposition of agentic AI, so a metric that measures effort is measuring the thing being sold.

Customer Satisfaction Score (CSAT) also remains valid, however with a condition that it is split it by handling mode (e.g. CSAT for humans, and CSAT for agents). Again by Ada, their research found that 55% of organisations measure AI and human interactions together, which makes it impossible to see where AI succeeds, where it leans on humans, and where it fails outright. You'll start to see pathways where AI resolves better than humans, and that could lead to reducing strain on your front line and up-skilling them to other parts of the business.

Churn, retention and customer lifetime value remain the ground truth. They are outcome metrics, not perception metrics, so agent intermediation cannot distort them. If anything they matter more, because as perception metrics degrade, outcomes become the only signal you can fully trust. However since they're outcome metrics, you still need correlation and causation tests.

NPS: is time to demote it?

Here is the firmer position. NPS should move from headline metric to lagging relationship indicator, and CX leaders should be explicit about the demotion.

The likelihood-to-recommend question assumes the respondent experienced the brand. In an agentic world, that assumption breaks. If a customer's AI agent researched the options, compared providers, negotiated terms and completed the purchase, what exactly is the human recommending? They experienced an outcome delivered through an intermediary. Their answer tells you the outcome landed, which retention already tells you, months earlier and without survey fatigue.

Even NPS's creator has conceded ground here. Fred Reichheld introduced Earned Growth Rate in 2021 as an accounting-based complement precisely because the survey score was being gamed. That is worth taking seriously: when the metric's inventor builds an alternative from financial data rather than sentiment data, the direction of travel is clear. Keep NPS as a quarterly relationship pulse if your board expects it. Stop treating it as the number that runs the operation.

Is first contact resolution still valid?

FCR deserves its own examination, because it looks like it should survive and mostly doesn't, at least not in its current form. The principle behind FCR is sound and permanent: resolve the customer's intent the first time they raise it. The unit of measurement, the "contact", is what breaks.

Three problems emerge:

1) AI responds instantly and is always the first contact, so FCR inherits the same gaming problem as deflection: an FAQ link served in two seconds counts as first-contact resolved whether or not the customer's problem was solved.

2) When the customer's own agent handles the interaction, failed attempts become invisible. The agent retries, reroutes and escalates silently in the background. From the human's perspective it was one contact, even if their agent needed nine attempts against your systems. The effort hasn't been removed, it has been hidden, and hidden effort still shows up later as churn.

3) Multi-agent journeys blur what a contact even is when a single intent spans your voice AI, your chat agent and a human handoff.

The evolution is to measure first intent resolution: was the customer's underlying intent resolved the first time it was raised, verified by repeat-contact rate over the following seven to fourteen days rather than by contact logs. Repeat contact is the honesty test. It cannot be gamed by a fast answer, and it catches the silent agent retries that traditional FCR never sees.

The new layer: measuring the machine customer

So we ascertained that some of the old guard metrics can still remain, but now we need a machine-experience layer alongside the human one. Five metrics to start with:

Autonomous resolution rate: End-to-end resolution without human involvement, verified by repeat contact, not containment.

Escalation quality: Not how few handoffs, but whether handoffs carry context and land well. Intercom's 2026 research shows hybrid escalation flows close the CSAT gap with human agents to 0.05 points. Escalation is a feature to measure, not a failure to minimise.

Agent accessibility: Can a third-party AI agent authenticate, navigate and complete tasks against your channels? Parloa tested the Global 2000 and found only 1% of enterprises ready for agent-to-agent interaction. Authentication assumes a human caller, IVR menus require button presses, and the APIs don't exist. This is accessibility testing for machine customers, and almost nobody passes.

Agent share of interactions. What proportion of your inbound traffic is non-human, tracked over time. Whilst this is tricky, there are certain behavioural anomalies that may indicate agents might be on your website (e.g. frequent skips in standard navigational pages, massive content scraping, etc.).

The uncomfortable truth

Adobe's 2026 AI and Digital Trends report found only 31% of organisations have a measurement framework for agentic AI at all. The dashboard gap is the norm, not the exception, which means the CX leaders who build the two-audience view now are not catching up. They are ahead.

The metrics that survive the agentic era share one property: they measure what the customer achieved, not what your systems did. Now we are in an era where understanding how agents engage with your website and your products / services (especially if you are in the software space) is just as important to improve, without losing focus on the human experience as well.