麻省理工学院技术评论见解：利用人工智能扩展实时支持

Produced in association with Scaling livesupport with AI Key takeaways live agent spends hours each week manuallydocumenting routine interactions. Anothercombs through multiple knowledge bases tofind the right solution, scrambling to piece itA A third types out the same response they’ve writtendozens of times before. resolution for customers and improved jobsatisfaction for customer support agents. These repetitive tasks can be draining, leaving less timefor meaningful customer interactions—but generative AIis changing this reality. By automating routine workflows,AI augments the efforts of live agents, freeing them to Organizations are implementing advancedsafeguards to address ongoing concerns. The true value of generative AI emergeswhen human strengths—like judgment,creativity, and empathy—complement AI’s “Enterprises are trying to rush to figure out how toimplement or incorporate generative AI into theirbusiness to gain efficiencies,” says Will Fritcher,deputy chief client officer at TP. “But instead of Fritcher notes that TP is relying on many of thesecapabilities in its customer support solutions. For instance,AI-powered coaching marries AI-driven metrics withhuman expertise to provide feedback on 100% of Doing this requires solving two intertwined challenges:empowering live agents by automating routine tasksand ensuring AI outputs remain accurate, reliable, andprecise. And the key to both these goals? Striking the •Call summaries:By automatically documentingcustomer interactions, AI saves live agents valuable time A key role in customer support Generative AI’s potential impact on customer supportis twofold: Customers stand to benefit from faster,more consistent service for simple requests, whilealso receiving undivided human attention for complex,emotionally charged situations. For employees, “Instead of viewing AI as away to reduce expenses,companies should reallybe looking at it through •Automated routine inquiries:AI systems handlestraightforward customer requests, like resetting •Real-time assistance:During interactions, AI pulls upcontextually relevant resources, suggests responses, Will Fritcher, Deputy Chief Client Officer, TP Ongoing issues with accuracy, preci-sion, and reliability in generative AI Understanding generative AI’soutput quality As organizations strive to integrate generative AI intoevery facet of their operations, they face an onslaughtof challenges pertaining to accuracy, precision, and reli- Accuracy, precision, and reliability are relatedconcerns for generative AI applications, but Accuracy “There’s not yet a gold standard for addressing theseconcerns; most organizations are still trying to figure itout,” says Abhishek Sengupta, practice director on theIT services team focusing on data, analytics, and AIat global research firm Everest Group. “But there are Factual correctness and truthfulness PrecisionRelevance and specificity of outputs Reliability Overall consistency and dependability of a system, Source: Compiled by MIT Technology Review Insights, 2025 After all, a highly advanced AI system is useless if it directscustomers to the wrong troubleshooting guideor rebooks a flight to an incorrect destination. As generative AI becomes more mainstream andenterprise-level adoption expands, these experimentswill only become higher stakes. Addressing persistentissues associated with accuracy, precision, and reliability explains Arnal Dayaratna, research vice president ofsoftware development at market intelligence firmIDC. “From a purely foundation-model perspective, Beyond hallucinations, AI faces several other fundamentalchallenges. Models often misinterpret ambiguous inputs,leading to responses that are technically correct butirrelevant. They struggle with cultural nuances and Perhaps the most notorious issue related to AI accuracyis the phenomenon of hallucinations: instances whereAI generates content that is factually incorrect or evenentirely fabricated. “Hallucinations are indigenous to Although prepared for technology, enterprises lag in risk and governancewhen deploying generative AI The percentage of organizations that think they are highly prepared for generative AI across the following areas: In customer support operations, these limitations havereal-world implications. An AI system might misinterpreta customer’s frustration about a billing error as atechnical issue or fail to recognize when a routine intelligence, RLHF helps bridge the gap between rawmachine outputs and context-aware responses. •Guardrails for AI-generated content:These are hardboundaries for the types of content generative AI ispermitted to produce. They’re typically designed to The business consequences of these missteps can besignificant—when a customer is upset about a serviceinterruption during a major family event, for instance,no number of AI-generated responses can replace a live How TP is integratinggenerative AI for accurateand reliable customer Ove