By the numbers
-
14%
Productivity gain from generative AI assistance across 5,179 support agents — but 34% for novices and near 0% for experts.
Brynjolfsson, Li & Raymond, QJE 2025 [1] -
−8%
Drop in customer rating (5-point scale) on chats an agentic AI system handled, despite chats running 16.8% shorter.
Alibaba/Taobao field experiment, 2026 [2] -
87%
Of customers say an option to reach a human agent is essential when a company uses generative AI in service.
Gartner survey of 3,566 customers, 2026 [3] -
24%
Of service and support leaders can demonstrate positive financial returns across their AI use cases.
Gartner survey of 1,303 senior leaders, 2026 [4] -
−79.7%
Less Sales when customers know they are talking to a bot. Customers became curt and disengaged, perceiving the disclosed bot as less knowledgeable and less empathetic. Late disclosure and prior AI experience both mitigated the effect.
Wang, Y., Zhu, C., Feng, T., Lu, L. X., & Jia, B. (2026).
What do people want when they call?
- Answers that are correct
- To be heard
- Low effort sessions
What people don't want when they call
- channel switching
- repeat contacts
- generic service
- unnecessary warnings
- difficult escalations
Bottom Line
Dixon, Toman & DeLisi — customer effort
The Classic Answer
- Give information problems to AI
- Emotional problems to People
The emotion trap
- What remains is harder If automation absorbs transactional volume, what remains for human agents is, by construction, harder.
- Support work is stressful The stress is concentrated in difficult contacts, and difficult contacts are exactly what remains after automation.
- Workers disengage Workers supervising the AI showed measurably lower engagement in emotional escalations — fewer messages, less proactivity.
A support model that routes a filtered stream of the hardest, angriest contacts to a shrinking team, and measures that team on handle time, is not a customer-first model. It is an attrition engine with a good dashboard.
The AI with the Agent Answer
On manufactured warmth
“What I see more often is brands talking about humanity in their AI strategy while systematically removing the moments where humanity could actually show up… The brands doing this well aren’t trying to script warmth into their AI. They’re using AI to remove the friction that exhausts their people, so that when a human moment is needed, their agents actually have the emotional capacity to show up for it. You can’t manufacture genuine care, but you can absolutely design the conditions that make it possible — or impossible.”
This aligns cleanly with the strongest academic finding. Brynjolfsson and colleagues did not find that AI made customers happier directly; they found it made agents better and more likely to stay, and customer sentiment improved as a consequence.[1] The causal path in the best-evidenced result runs through the workforce, not around it.
The unglamorous stack still decides the outcome
The guidance names the central challenge as ensuring customers get consistent, accurate, timely information, and warns that standing up a repository is only one part of knowledge management — the harder part is the process ensuring knowledge is captured, curated, and kept current.[13]
Gartner finds self-service success is a top priority but organizations face knowledge backlogs and inconsistent content review, with 58% of leaders planning to upskill agents into knowledge management specialists to curate AI-generated content.[5] A retrieval-augmented assistant over a stale knowledge base produces confident, current-sounding wrong answers at scale.
Intelligent routing — rules, time, geography, skills, and IVR-collected data — exists to get the contact to the right resource the first time.[13] Every subsequent generation of technology has been an improvement in the inputs to that decision, not a replacement for it.
Agentic AI is best understood the same way: a dramatically better classifier and executor sitting on the same routing problem. Which is why practitioners report the failure mode as amplification — as one puts it, if the routing logic was clunky before, “it’s confidently clunky now, at scale.”[10]
Full research review
The complete document, Customer-First Contact Centers