By the end of 2026, an overwhelming 88% of companies say they will have artificial‑intelligence (AI) agents handling at least part of their customer conversations, according to a new Sinch survey of 2,527 senior decision‑makers in ten countries. This strong adoption signal reflects the growing confidence that businesses have in AI to improve efficiency while still preserving the human touch for high‑stakes interactions.
Human agents remain essential
Gartner research backs the trend, noting that nearly 80% of customer‑service organizations plan to reshape human‑agent roles as routine tasks become automated. Complex or sensitive issues will continue to be routed to a person, ensuring that customers receive accurate answers when it matters most.
When AI should hand over
The key to a successful deployment is a system that can recognize when it has reached its limits. An AI agent must detect questions it cannot answer correctly, spot signs of customer frustration, and transfer the conversation to a live representative. Failure to do so can lead to three outcomes: a correct answer, a confidently wrong answer, or no answer at all.
Recent high‑profile mishaps illustrate the stakes. In 2024, DPD’s AI assistant failed to locate a missing parcel, could not connect the caller to a human, and even responded with a poem. A British Columbia tribunal also found Air Canada liable for negligent misrepresentation after its website chatbot gave a grieving customer incorrect information about a bereavement discount.
Rollbacks and risk management
In the Sinch survey, 22% of organizations that had live AI agents already rolled one back because of hallucinations or brand‑risk concerns. When asked about the most significant business impact of an AI‑driven failure, 34% cited reputational damage and loss of customer trust—damage that can be difficult to reverse.
Companies are responding by tightening guardrails. More than three‑quarters (74%) of firms running AI agents have pulled an agent back at least once, and among those that describe their AI safeguards as fully mature, the rollback rate rises to 81%. This suggests that mature governance often includes proactive monitoring and timely handovers.
Designing effective handovers
Effective handover design relies on two main inputs: sentiment detection and intent recognition. Sentiment analysis gauges the emotional tone of each message in real time, while intent recognition identifies what the customer is trying to accomplish. Together they help the AI decide whether a human should intervene before the customer even asks.
When frustration is detected, the system can flag the case as urgent, prioritize it in the support queue, and route it to the appropriate team—whether billing, logistics, or technical support. The human agent receives the full conversation history, reducing friction and avoiding repeated explanations.
Balancing scope and usefulness
An AI agent’s error rate can be reduced but never eliminated. Companies must decide before launch which questions the bot will answer and which will be escalated. Narrowing the scope too much limits the agent’s usefulness; leaving it too broad increases the risk of confident misinformation.
Consumer confidence also varies by task. Sinch’s pre‑Black‑Friday survey of 2,501 consumers across eight countries showed high overall trust in AI agents, but confidence dropped for tasks involving money or account security. Rather than excluding those tasks entirely, firms should make the handover process seamless, allowing customers to reach a human quickly when needed.
Measuring performance responsibly
Many organizations track containment and deflection rates—how many conversations the AI resolves without human help. While high rates can look good on paper, they may mask underlying failures if customers are forced to repeat issues with a live agent later. A balanced metric set that rewards successful handovers as well as accurate AI resolutions provides a clearer picture of true customer experience.
Overall, the data suggest that AI‑driven customer support is moving from experimental pilots to mainstream deployment, with companies learning to blend automation and human expertise for better service outcomes.
Original reporting: KTVZ (Central Oregon) — read the source article.
