Adoption of artificial intelligence in customer experience has climbed to nearly 70% in 2026, yet only 2% of organizations using the technology qualify as a Center of Excellence, according to the third annual State of AI in CX Report from Forethought AI Agents by Zendesk.
Adoption rose from 54% in 2024 to 57% in 2025 before reaching 69.7% this year. Forethought scored respondents against a 100-point Center of Excellence Index that weighs measured outcomes, operating maturity, training sophistication, production control and ROI discipline. Seventy percent landed in the lowest “Experimental” tier, 10% in Foundational, 12% in Scaling, 7% in Advanced and 2% at the top.
The report tied results to what AI does rather than whether it is present. Among programs with action-taking AI, 74% reported improving resolution rates, versus 43% at the assistive stage. Organizations using AI reported improving resolution rates at 56%, compared with 19% of non-users, a 36-point gap the report called the widest in its trend analysis. AI users also reported improving cost per resolution at more than twice the rate of non-users, 42% versus 19%.
Operating model tracked with performance. Fifty-four percent of respondents said they use a purpose-built CX AI operating model, about one-quarter use a help-desk add-on, and 22% built in-house systems. Purpose-built programs reported improving resolution 16 points more often than do-it-yourself deployments and posted a mean Index score of 59.4, against 50.3 for embedded and 46.7 for DIY.
The report flagged a channel gap. Voice support is offered by 71% of respondents but backed by live AI at only 20%. Email and chat showed the closest alignment between support and AI coverage.
Production issues remained common. Hallucinations and incorrect answers were the most cited, at 74%, followed by poor escalation or handoff at 44% and workflow or integration failures at 40%. Just 3% reported no issues.
On autonomy, 46% of respondents said they were only somewhat confident that AI could resolve issues end to end without a human, while 25% were very confident and 10% extremely confident. Forethought said all AI-impact data is self-reported and should be read as directional rather than causal.
