LectureRechercheRadarCadre d'investissement
Connexion / Inscription
← Retour à la lecture
ActualitésSujet · AI智能体问责制Matériel4 min de lectureRédaction DisconfirmAI

Insurance Tech COO: The Hardest Problem with AI Agents Isn't Technology, It's Accountability

Rick Folgmann, Chief Operating Officer of Get Covered, draws on his operational experience running 16 AI agents across 3 million rental units to argue that the core bottleneck in deploying AI agents is organizational accountability rather than model capability. He outlines three specific management practices: starting with tasks governed by existing rules, setting autonomy levels based on workflow risk, and assigning a named human owner responsible for each agent.

2026-10-27 19:00 GMT+8

A landlord in California has 21 days after a resident moves out to return the security deposit or send an itemized list of deductions. Miss it in bad faith, and the penalty can reach twice the deposit. That deadline doesn’t care whether the letter was supposed to come from a person or a piece of software. Either way, someone will answer for it.

Rick Folgmann is the chief operating officer at Get Covered, an insurance technology company. His company now runs 16 AI agents across more than 3 million rental units. Recalling the experience of building these agents, he said that technical questions turned out to be the easier part; the truly difficult question for every agent is who it reports to.

He believes most companies are about to face the same issue, yet few are prepared. Deloitte’s 2026 State of AI survey found that 23% of enterprises currently use agentic AI at least moderately, and 74% expect to reach this level within two years. Meanwhile, Gartner predicts that more than 40% of agentic AI projects will be canceled by the end of 2027.

Folgmann’s conclusion is that these figures indicate agents should operate like employees in regulated roles—with written rules, defined scope, a manager, and regular reviews. He offers three practices distilled from his own operations.

First, start with tasks that already have rules. The best early candidates for agents are tasks where rules already exist outside anyone’s head: a statute, a contract clause, or insurance requirements in a vendor agreement. He cited the example of verifying vendor certificates of insurance—property teams across the industry spend roughly three hours per property every week chasing these documents via email. The work is tedious but completely specified. Agents have objective standards to validate output, and when the team audits results, there is no need to debate judgment calls. His rule of thumb is: if a new employee cannot make the decision according to your description on their first day, neither can an agent.

Connexion / Inscription
LectureRechercheRadarCadre d'investissement

Second, set autonomy by workflow. He noted that tech executives often ask whether an AI system is autonomous, as if it were a fixed attribute of the product. In reality, the same agent should operate at different levels of freedom depending on what it touches. At Get Covered, a reminder to a vendor whose certificate expires next month can be sent automatically; a response to a resident requesting an emotional support animal never will. He advises leadership to list every workflow an agent will touch and ask the same question for each: If this action is wrong, how long until we discover it, and what will the cost be by then? Quick detection and low cost earn the agent more room.

Third, assign a named human owner to each agent. Folgmann learned early in his career in the financial industry that every automated credit decision must eventually be explained to a customer or regulator.

The model’s output has never been an acceptable answer. This principle still holds, except agents now act more frequently and across more systems than any underwriting model did. Therefore, at Get Covered, every agent has a person ultimately responsible, just as a team or a budget has an owner. That person reviews its error logs, decides when to expand or restrict its autonomy, and takes responsibility for it.

He also points to a quieter problem: when multiple teams launch agents independently, no one sees where those agents contradict one another.

Folgmann acknowledges that models will continue to improve, but a company’s clarity on who decides what will not improve on its own. He believes Gartner’s cancellation prediction reflects exactly this point: most failed agent projects are not because the models are inadequate, but because no one defined value or controlled risk—both of which are management issues.

His advice is to treat each agent like a new employee in a regulated role. Write down its rules, limit its scope, and assign a manager to check its work and discuss its performance. Companies that do this will be able to give agents more responsibility over time because they have evidence to justify doing so. Companies that skip this step will spend 2027 explaining why their software did something nobody approved.

It is worth noting that Folgmann’s framework rests on a specific premise: the organization possesses the discipline to execute these management actions. The Gartner prediction he cites—that over 40% of agent projects will be canceled—implies precisely that many companies lack this discipline. His advice has direct operational value for teams that already possess basic operational governance capabilities, but for organizations unwilling to even write rule documentation, a correct diagnosis does not necessarily equate to executability.

Sources

  • news.crunchbase.comCrunchbase News↗
Sujet

AI智能体问责制

随着AI智能体在企业中大规模部署,核心瓶颈已从模型能力转向组织问责:当软件自主执行任务时,谁对结果负责?保险科技公司Get Covered首席运营官Rick Folgmann指出,在300万套出租房中运行16个智能体的经验表明,必须为每个智…

6 étapes1 piècesMis à jour 2026-10-28 08:06

Comment on en est arrivé là

  1. 2025-06-17Gabison与Xian发表研究,从委托-代理理论视角分析LLM智能体系统的固有及新兴责任问题,指出现有法律框架难以有效应对AI系统中的多重委托关系。arxiv.org ↗
  2. 2026-05-06Treude发布关于软件工程智能体问责制的愿景论文,对比主流AI编码助手的条款,发现提供商倾向于将正确性、安全性和合规责任转移给用户。arxiv.org ↗
  3. 2026-08-13南加州大学等机构提出“可审计智能体”框架,强调没有可审计性就没有问责制,并定义行动可恢复性等五个维度,指出当前开源项目普遍存在安全前提缺失。arxiv.org ↗

Ce qui reste non résolu

  • Gartner预测超过40%的智能体项目将在2027年底前取消,主要因价值未定义或风险失控,具体哪些行业或类型的智能体最易失败尚待观察。
  • IBM等推出的身份控制工具能否被广泛采纳,以及其实际降低违规成本的效果如何,需后续企业部署数据验证。
  • 学术提出的“可审计性”标准(如证据完整性)是否会转化为新的行业合规要求,目前尚无定论。
Lire le sujet complet →
Continuer la recherche →