· AI
815 CX Leaders Say Fully Autonomous AI Never Wins. That Is Not the Case Against Automating.
The survey everyone will quote to go slow on AI agents is actually a map of exactly where to go fast.
A survey of 815 enterprise customer experience decision makers found that fully autonomous AI did not come out on top in a single function measured, not technical support, not billing, not complaint handling. Ryan Strategic Advisory's 2026 CX Technology and Global Services Survey, commissioned by TELUS Digital and fielded across 12 countries and 19 industries, found human agents working alongside AI were the most-cited approach in every category respondents could pick, ahead of human-only service, basic automation, and agents running unsupervised. Stop at that headline and 2026 looks like the year AI agents stalled out in customer service. Look at what is actually running inside those same functions and the story reads more like a map of exactly where agents are ready to take over, and where they are not yet.
My read is that the survey is right about functions and silent about tasks, and that gap is where most companies get this decision wrong. Technical support, billing, and complaint management are not single jobs. Each is a bundle of smaller tasks, some repetitive and low stakes, some carrying real financial or reputational weight, and the survey asked about the bundle, not the pieces inside it.
Here is the honest case for caution, because it deserves a fair hearing. Technical support, retention, and billing are exactly the functions where a bad autonomous call costs something real: a wrongly waived fee, a retention offer sent to the wrong account, a diagnosis that misses a safety issue. Databricks' framework for allocating work between agents and humans maps this directly: tasks with high ambiguity and high consequence should stay human-led, and the functions in the TELUS-commissioned survey sit heavy on both axes. A company that handed billing disputes or retention negotiations to an unsupervised agent this year was taking a real risk, and the respondents who picked the safer, assisted model were not being timid. They were pricing the downside correctly. At the function level, human-plus-AI is the right call, and nothing below should read as license to flip that switch wholesale.
But a function is not the unit that actually gets automated, and the usage data shows it. Comm100's 2026 Live Chat Benchmark Report, built from more than 220 million live chat interactions across 18 industries, found AI handled 75.3 percent of chats, the highest share the report has recorded. Only 44.8 percent of the chats AI handled were resolved without a human stepping in. Multiply those two figures and roughly a third of all chats in the dataset were started and finished by AI alone, a rough estimate that a separate analysis at Kommunicate arrived at independently using the same published figures. The other two out of every three AI-handled chats did not fail. They did what the TELUS-commissioned survey's preferred model describes: an agent took the first pass, then a person finished it. That is task-level automation already happening inside a function the industry still officially files under "needs a human."
So the two reports are not actually in tension. They describe the same split from two different altitudes. At the function level, almost nobody is choosing full autonomy, and the survey is right to say so. At the task level, inside those same functions, AI is already taking the first pass on three out of four conversations, and in roughly a third of them that first pass is the whole job. The real management question is not "can we trust an agent with technical support." It is "which pieces of technical support are a password reset and which ones are a warranty dispute," because the first kind clears Databricks' low-ambiguity, low-consequence bar for full automation and the second does not, inside the same ticket queue.
That is the audit a business owner should actually run. Pull a month of tickets or calls from one function, sort them by whether the correct answer is obvious and the cost of a wrong one is small, and hand only that pile to an agent with no human required. Keep the rest assisted, the way the TELUS-commissioned respondents already do, until the error rate earns more autonomy. Skip the sort and judge the whole function by its riskiest ticket, and a company ends up making both mistakes at once: trusting an agent with a refund decision it should not have, and sitting a person on password resets a machine could clear in seconds. Either one costs real money, just in different directions.
That ticket-by-ticket sort is most of the design work behind MojoAgents: not a verdict on whether a department is ready for AI, but a map of which tasks inside it already are. If you are staring at a function trying to guess which side of that line each piece belongs on, that conversation is worth having before you buy another seat license.
Sources
References used in this article. Links also appear alongside the relevant claims.
Let's make it happen.
You don't need a polished brief. A couple of lines about where you want your company to go is plenty, and we'll come back with what we'd tackle first.
We scope the work around your goals and timeline before quoting anything, so you know exactly what you're getting.