Salesforce reports customer service AI agent adoption rising from 39% in 2025 to 66% in 2026, while roughly one in five consumers who have used AI support say they saw no benefit. The gap is explained by what teams automated rather than by whether they automated.
Key takeaways
- Adoption rose faster than the discipline needed to make it work.
- Automating judgement work produces the experiences that generate the complaints.
- Documentation quality, not model quality, sets the ceiling on outcomes.
- Escalation design is what customers actually judge the deployment on.
Two numbers from 2026 sit oddly together. Salesforce reports customer service organisations using AI agents rising from 39% to 66% in a year. Meanwhile roughly one in five consumers who have used AI for customer service report seeing no benefit, a failure rate several times higher than consumer AI use generally.
Adoption is not the problem. What got automated is.
The split that explains it
Support volume divides cleanly into two kinds of work, and they are not equally automatable.
Documented, repetitive, low-risk. Access and login questions, order and delivery status, plan and feature explanations, billing policy, setup how-tos. Automation handles these well, often better than a queue does, because the customer gets a correct answer instantly at three in the morning.
Judgement, exception, undocumented. Complaints, cancellations, commercial exceptions, anything involving money in dispute, anything your documentation does not cover. Automation handles these badly, and the failure is visible and memorable.
Teams reporting good outcomes automated the first category. Teams generating the complaints automated across both, usually because nobody categorised the queue before deploying.
The refund problem specifically
The consumer frustration reported in 2026 clustered noticeably around refunds and complaint handling.
That is not an accident. A refund conversation is a judgement call about money, frequently from someone already annoyed. It is close to the worst possible automation target, and it is a tempting one because the volume is high and the scripts look simple.
The fix is a routing rule rather than a better model. Complaints, cancellations, refund disputes, and any signal of financial difficulty should reach a person by explicit rule, not by a confidence threshold that occasionally does not fire.
Three things that separate the outcomes
Content quality. A grounded system reproduces your documentation faithfully, including its gaps and contradictions. Teams that audited their content before deploying got better results from the same technology. This is the largest single variable and the least discussed.
An honest refusal path. A system that always answers is not grounded, it is confident. The ability to say it cannot confirm something and route to a person lowers your resolution rate and is what keeps invented policies out of customer conversations.
Escalation design. Customers judge the whole deployment on what happens when it cannot help. A handoff carrying the full conversation means nobody repeats themselves. A handoff carrying a summary line means they do, and that is the experience people describe as being passed around.
The pressure that causes the mistake
Gartner found 91% of service leaders feeling pressure to implement AI in 2026. That pressure is the mechanism behind the gap.
A team under pressure to show adoption deploys broadly and quickly. A team optimising for outcomes deploys narrowly, measures, and expands. The second produces better numbers and worse-looking progress reports in the first quarter, which is exactly why the first happens.
The counter is to agree the measurement before the deployment: cost per resolution as the metric, reopen rate and satisfaction as guardrails, and deflection explicitly excluded because it counts customers who gave up as successes.
What good looks like
Run the categorisation first. A thousand tickets, grouped by topic, each marked documented, documentable, or judgement. Automate the first, write content for the second, route the third to people by rule.
Then measure honestly for a month. If cost per resolution falls while reopens and satisfaction hold, expand. If reopens rise, you are automating failure at scale and the fix is almost always content rather than vendor.
That sequence is slower to report and considerably less likely to put you in the one-in-five statistic. We wrote up the queue analysis and what AI actually resolves in more detail.
Sources: Salesforce adoption figures and Gartner survey data as reported in 2026 industry coverage; consumer sentiment reporting from CNBC.
Frequently asked questions
How many companies use AI in customer service?
Why do customers report no benefit from AI support?
Does AI customer support actually improve satisfaction?
What is the most common AI support mistake?
Resolve, don't deflect.
See Fidiora resolve a ticket, capture a lead, and keep the bill predictable.