Adoption of AI in customer service has reached nearly 70%, up from 54% in 2024. Of the organisations that have deployed it, 2% score in the top tier of a new maturity index. Seventy percent sit in the lowest one, labelled “Experimental.”
Those numbers come from the third annual State of AI in CX report, published on 13 August by Forethought AI Agents by Zendesk and based on responses from more than 600 CX leaders and practitioners. Read as a headline, it says AI in customer service does not work. Read in detail, it says something narrower and more useful: the way large companies deploy it does not work.
What is actually failing
Forethought scored programmes on five dimensions — measured outcomes, operating maturity, training sophistication, production trust and control, and ROI discipline. Antoine Nasr, the company’s head of AI, told CX Dive that “taking shortcuts with AI cannot repair a fragmented customer journey, inconsistent knowledge, poor handoffs or a lack of ownership.”
That list is worth reading twice. It is a list of organisational problems, not technical ones. It is corroborated by Sinch’s survey of 2,527 senior decision makers across ten countries, which found 74% of enterprises had rolled back a live AI customer communications agent after deploying it — rising to 81% at companies with the most mature governance frameworks.
A fragmented journey across six systems, knowledge that contradicts itself between departments, and nobody who owns the outcome: these are the specific conditions of a company with a thousand employees. If you run a restaurant, a clinic, a small firm or a shop, you do not have them. Your knowledge lives with two or three people, your handoff is one phone, and the person who owns the outcome is you. The thing that kills most of these projects is the thing your size protects you from — which is why the same technology behaves differently across smaller operations than the survey headline suggests.
Two numbers in the report worth acting on
The first is the gap between AI that answers and AI that does. Among programmes where the AI can take action, 74% reported improving resolution rates, against 43% where it only assists. Satisfaction followed the same split: 53% versus 33%. An assistant that says “we may have availability Thursday” has moved nothing. One that checks the calendar and books it has.
The second is the phone. Seventy-one percent of organisations use voice as a support channel, but only 20% have live AI on it. Voice is the least automated channel and the one where a missed contact is gone permanently — nobody leaves a second voicemail.
Where to start, in order
Spend one week counting: calls that rang out, voicemails left after closing, web forms and DMs answered the next day. That is your baseline, and for most operators it is larger than they expect.
Then write one page of answers to the five questions you field every day — hours, prices, availability, what is included, how to change a booking. The “inconsistent knowledge” that sinks enterprise deployments takes an afternoon to fix at your scale.
Decide what must reach a human: complaints, refunds, anything touching a specific booking. Then insist on the action-taking version the data favours, on the phone as well as chat — answering, booking and following up across every channel at once, not a widget that collects questions for you to answer later.
Recount in 30 days. If the number of missed contacts has not moved, the setup is wrong. If it has, you have the thing only 2% of the organisations in that survey could show: value you can actually measure.
The TalkRev blog
One short email per new post
No ads. Unsubscribe anytime.
