Thirty-nine percent of U.S. contact centers have stopped, paused, or rolled back an artificial intelligence solution after deploying it.
That is not a figure anyone has an incentive to publish. Suppliers plainly don't. Buyers rarely volunteer it either, since "we spent eight months on something we then switched off" is not a sentence that features in many conference keynotes. It came out of ContactBabel's own survey work for "The Inner Circle Guide to Voice AI," where the respondents had no particular reason to flatter anybody.
Against this, 74 percent of the same respondents expect their contact center AI budgets to rise over the next two years. Only 4 percent percent expect it to fall.
Both of those are true, and reconciling them tells you more about the state of this market than either does alone. This is not a market losing its nerve; it's a market finding out, project by project, what actually works, and the failures are at least as informative as the case studies.
Of those who stopped something, 64 percent stopped at a limited pilot and 21 percent at proof of concept. That is evaluation doing its job, and nobody should lose sleep over it. The concerning slice is the 15 percent who pulled a solution out of full production because by that point the integration work is done, the agents have been trained and the customers have already met it.
The leading reason, by a wide margin, is accuracy or reliability not being good enough, at 68 percent. Integration proving too difficult follows at 34 percent, customers reacting badly at 32 percent, and cost being higher than expected at 30 percent. (Multiple reasons for stopping were allowed, and few respondents had only one reason for doing so).
I would note that accuracy is not always a property of the product bought. A voice AI deployment is a stack: the audio, then the transcription, then the understanding, then whatever the system is permitted to do in your business systems. Those layers frequently come from different suppliers, are likely bought at different times by different people, and the weakest one sets the ceiling for everything above it. A voicebot with excellent intent detection will still choke if what it hears is a customer with spotty cell service calling from a busy grocery store parking lot.
The audio is getting harder. Seventy-five percent of customers now call on cell phones, up from 48 percent in 2017. Half of U.S. agents work remotely or in a hybrid environment, which means the operation no longer controls the acoustic environment at either end of the call. More than half of consumers tell us they regularly have to repeat themselves or struggle to hear the agent. When companies conclude that voice AI does not work, the fault quite often sits a layer or two below the thing that got the blame.
It's worth mentioning again the 32 percent figure for customers reacting badly: a meaningful number of deployments got in front of customers, at volume, before anyone found out just how bad they were.
What buyers say they want, and what they get pitched
We asked what matters most when deciding where to source AI, with respondents choosing up to three factors.
Ease of integration with existing systems led at 69 percent, followed by data security and compliance at 66 percent, total cost at 63 percent and depth of capability at 43 percent. Those four account for very nearly the whole distribution.
Speed of deployment mattered most to only 20 percent, and the simplicity of having fewer suppliers to manage matters most to 7 percent.
"Deploy in days" is a common claim made by vendors in this category, but few buyers actually treat that as a priority. Buyers really want to know how the product will behave inside a 20-year-old core system and what happens to the call recordings. Both are questions where evidence is hardest to produce and where a demo tells you little.
We should note what buyers rank as the least significant barrier to further adoption over the next two to three years. Vendor immaturity, at 13 percent, comes last. Even after a 39 percent rollback rate, buyers still do not think the products are the problem. They think their own data, their own integration, and their own customers are the problem, which is a considerably more useful diagnosis and a much less comfortable one.
On customers, they are right. Customer resistance is the leading barrier to further AI implementations at 57 percent, well ahead of regulatory concerns at 37 percent. Our consumer research says the same thing from the other side of the call: 34 percent of U.S. consumers say they would never prefer AI under any circumstances, 63 percent are worried about being unable to reach a human, and consumers rate AI worse than a human on five of six service dimensions. The exception is speed, where they rate it clearly better. Businesses and their customers are describing the same constraint and agreeing about it.
I would encourage you to fix the bottom of the stack before buying at the top. Transcription accuracy is consumed by every layer above it, so improving it is the cheapest way to improve several things at once. It also lets you measure before committing to anything.
Then, write down what success is, in a number, before you sign. Between 3 percent and 16 percent of respondents set no measurable expectations at all in their original business cases, depending on the metric being considered. Between 26 percent and 46 percent cannot say whether their AI performs better or worse than their people on a given metric. If you never defined the target, you can't tell a project that is failing from one that is bedding in, and you will end up rolling back or moving forward on vibes.
Treat the gap between pilot and production as a real risk rather than a formality: the 15 percent who stopped in production did not fail the pilot.
A final point: target the AI investment at contacts your customers are content or even keen to have automated. Quick answers, queue avoidance, and out-of-hours automation works with the grain there.
But if deployed against complex or emotionally loaded contacts, it is working against the large majority of consumers who already think AI is worse than human agents at understanding the issue, and no amount of prompt engineering fixes that.
Steve Morrell is managing director and principal analyst of ContactBabel, which was founded in 2001 to provide research and analysis to the U.S. and U.K. contact center industries.