By Gregg McKittrick, Head of Strategic Partnerships, Symbos
Most customer experience AI is sold on a single number: the share of contacts the system can handle without a person. Containment. It sits in the vendor deck and the board update, and it is a large part of why so much of this investment is failing.
The evidence is hard to ignore now. MIT’s State of AI in Business 2025 found 95 per cent of organisations investing in generative AI were seeing zero return. Gartner expects a large share of agentic AI projects to be cancelled within two years, in a market it describes as full of vendors rebadging ordinary chatbots as agents. In Qualtrics’ 2026 consumer research, AI customer service was the worst-performing AI application people use.
We run contact centres, collections operations and back offices for more than 250 clients, so we see what containment hides. A contained contact is not always a resolved one. It can be a customer who gave up, a query that returns tomorrow, or a collections conversation that quietly disengages and turns into arrears. The dashboard logs a win. The operation wears the cost a month later.
AI in CX does work, and it works when it is built on the right measure and done in the right order. Most of the failures trace back to those two things. Here is what that looks like in an operation.
Automating a contact that should not exist
Automation makes a contact cheaper to handle. It does nothing about whether the contact should have happened. A large share of contact volume is failure demand, meaning contacts created by an earlier failure somewhere else in the business. A bill the customer could not understand. A promise that was not kept. A digital form that broke halfway through. Automate that volume and you have made your own failures cheaper to deliver, while the customer still has the problem and you still carry the cost, now spread across a bot and the escalation that follows.
So the first pass in our implementations is not a technical one. We map where contacts originate and why the same ones keep returning, then separate the volume worth automating from the volume worth removing at source. Removing a contact is worth more than deflecting it, and it almost never shows up on a containment dashboard.
A home care provider we work with is a good example. Before any model went near their operation, the work was unifying a customer view that had lived in separate platforms for years. That surfaced how much of their inbound volume was the same customers chasing the same unresolved issues across different channels. The value came from fixing those root causes and giving what remained a reliable foundation to run on. A bot layered over the original fragmented view would only have made the chasing cheaper.
A wrong answer at scale is a compliance problem
Between the strategy deck and go-live sits the work almost nobody scopes properly. An AI answer is only ever as current as the last time someone updated the knowledge base behind it. In a service context, a stale answer is a poor experience. In a collections or regulated context, it is a confidently wrong answer delivered at scale, around the clock, that can put you in breach before anyone notices.
Keeping that knowledge accurate is a funded, ongoing role, and most business cases leave it out. The model gets built, the launch happens, and answer accuracy starts decaying within weeks because nobody owns it as a daily job. The operations that get this right price knowledge accuracy in as a permanent operating cost from the first business case.
Automation leaves your hardest contacts for people
When the model takes the simple, high-volume contacts, what remains for people is the complex, multi-issue and emotional work. The residual contact is harder than the average contact used to be. Human handle time goes up, and the skill profile you need goes up with it.
That changes the workforce model. A plan built on the same agents doing the same work in smaller numbers will miss its own case. The honest version is fewer people, more capable and better paid, doing harder work, and it has to be costed that way. Gartner now expects half the organisations planning significant customer service cuts to abandon those plans, and predicts that firms cutting staff because of AI will be rehiring for similar roles by 2027. The ones caught out costed people as a line to cut. The capability was the thing they needed to keep.
The handover from model to person is the single most important moment in the design. The person needs the full context of what just happened, including its emotional temperature, so the customer is not made to start again. If they have to repeat the whole thing, the automation has already cost you the interaction, whatever the containment rate says.
Containment is the wrong number
Three numbers predict whether CX AI pays back: resolution rate, cost to serve per resolved contact, and the repeat-contact rate across thirty, sixty and ninety days. Containment sits in none of them.
The second-order effects matter as much. CSAT scored only on the customers who completed a self-service journey is survivorship bias. It says nothing about the ones who abandoned it and rang in angry, and those are the contacts that carry the real cost.
When we modernised support across channels for a major telco, we held the work to those measures. Handling time fell and first contact resolution rose, and the operation was left with a base of resolved contacts for automation to build on, because the operation underneath was resolving them properly.
An initiative that cannot state the resolution and cost-to-serve numbers it exists to move, and name the person accountable for them, has not been scoped yet. Calling it a plan does not make it one.
Four questions worth putting to any vendor
- Which of our contacts are failure demand we should remove at source, and which are real volume worth automating?
- Who owns the accuracy of the knowledge this system answers from as a funded daily role, and what does answer decay look like at ninety days?
- What does the customer and the agent inherit in the first ten seconds of an escalation?
- Will success be measured on containment, or on resolution and cost to serve, and who is accountable for that number?
None of these is a technology question. Every one of them decides whether the technology pays back.
The 95 per cent are not behind on technology
The organisations seeing zero return are not short of models. They are measuring deflection, automating demand they should have removed, and treating containment as a result. The ones seeing returns worked in a particular order. They found out which contacts should not exist. They made the knowledge trustworthy and kept it that way. They designed and funded people for what was left. And they held the whole thing to resolution and cost to serve. None of it is exotic. It is the work, and it is where the returns sit.
That is the work we do, and we are happy to get into the specifics of your operation.